a reading room - ongoing
Reading Room
This reading room is an annotated bibliography of the heritage behind Singulars. Each performance sets a poet against a language model trained on him. Both poems hang unsigned. The audience votes, and the votes train the next model. The entries are papers, artworks, contests and performances, listed chronologically in five sections. Each has a short factual note, a line on why it matters here, and a link to a stable source.
Passing: imitation games, chatbots and tests of humanness
Turing's test began as a parlour game about gender. A judge reads typed answers and guesses who is behind the screen. This section follows that game from automata and chatbots to artists who play the machine, alongside the older meaning of passing in writing on race and gender. The latest model in the series, next pass, takes its name from here.
Nothing in this section matches the filter.
c. 830-850book
Kitab al-Hiyal (Book of Ingenious Devices) and The Instrument Which Plays by Itself
Banu Musa ibn Shakir (the brothers Muhammad, Ahmad and al-Hasan)
Three brothers at the Abbasid court in Baghdad described about a hundred trick vessels, fountains and other devices. A companion treatise, The Instrument Which Plays by Itself, describes a hydraulic flute player driven by a pinned rotating cylinder. ZKM Karlsruhe rebuilt it in 2015.
A machine performing a human art from a stored score, in ninth-century Baghdad. The seeming-alive machine starts here, not in Vienna in 1770.
1206book
The Book of Knowledge of Ingenious Mechanical Devices (Kitab fi ma'rifat al-hiyal al-handasiyya)
Badi' al-Zaman ibn al-Razzaz al-Jazari
Written for the Artuqid ruler of Diyar Bakr after twenty-five years of court service. It gives build instructions for clocks, fountains, water-raising machines and automata, including a boat of mechanical musicians for drinking parties. The linked Freer folio, copied in 1315, shows that boat.
A mechanical band built to entertain guests at a party. Machine performance before a live audience is eight centuries older than Singulars.
1637book
Discourse on the Method, Part V
René Descartes
Published anonymously in French at Leiden. Part V argues that a machine built to look human could utter words but could never arrange them to answer sensibly whatever is said in its presence. A second test follows: machines lack the general adaptability of reason.
The first language test for machines, three centuries before Turing. The imitation game and every Singulars vote descend from this wager.
1836 (automaton 1770)essay
Edgar Allan Poe, on the automaton built by Wolfgang von Kempelen
Published in the Southern Literary Messenger in April 1836. Poe examines the touring chess automaton built by Kempelen in 1770 and later shown by Johann Maelzel. In seventeen numbered observations he argues that a concealed human directs every move. He was right in substance.
An early reverse Turing test. A writer sets out to prove that the machine on stage is a person.
1843paper
Ada Lovelace (Augusta Ada King, Countess of Lovelace)
Lovelace translated Menabrea's memoir on Babbage's Analytical Engine for Taylor's Scientific Memoirs and added notes roughly three times its length, signed A.A.L. Note G gives a program for Bernoulli numbers and states that the Engine cannot originate anything, only do what it is ordered.
Turing names this Lady Lovelace's objection and answers it in 1950. Singulars restages the dispute each night: can the trained model originate a poem?
1916-1918book
Spectra: A Book of Poetic Experiments (the Spectra hoax)
Witter Bynner and Arthur Davison Ficke, writing as Emanuel Morgan and Anne Knish
Two traditional poets invented a school called Spectrism to parody Imagism and published it with Mitchell Kennerley under false names. Marjorie Allen Seiffert joined as Elijah Hay. Critics took the movement seriously for two years until Bynner confessed during a Detroit lecture in April 1918.
Poems written as a joke were read sincerely and sometimes preferred to the authors' real work. Authorship is a frame the reader supplies.
1929book
Nella Larsen
Harlem Renaissance novel published by Knopf. Clare Kendry, a light-skinned Black woman, lives as white and is married to a racist white man. Her reunion with childhood friend Irene Redfield exposes what passing costs and who polices the line.
Gives the series its deepest sense of passing: identity as a performance judged by others, with real stakes for being read correctly.
1943-1944event
The Darkening Ecliptic (the Ern Malley hoax)
James McAuley and Harold Stewart, writing as Ern Malley; published by Max Harris in Angry Penguins
Two conservative poets assembled a dead poet's life work in one afternoon, lifting phrases from a dictionary, Shakespeare and an army mosquito report. Max Harris published it in Angry Penguins. Sydney papers exposed the hoax in June 1944. Harris was then fined for obscenity.
Verse assembled half mechanically, judged sincerely, then defended as real poetry by its editor. The canonical proof that a poem can pass without an author.
1950paper
Computing Machinery and Intelligence
Alan Turing
Published in Mind in October 1950. Turing replaces the question of whether machines can think with an imitation game. In the first version a man tries to pass as a woman to an interrogator by teleprinter. A machine then takes the man's part.
The founding document. Singulars restages it with poems, and the original game was already about gender and passing.
1966system
ELIZA: A Computer Program for the Study of Natural Language Communication Between Man and Machine
Joseph Weizenbaum
Described in Communications of the ACM. Weizenbaum's MIT program matched keywords and reflected them back as questions. Its DOCTOR script played a Rogerian psychotherapist. Users confided in it, and some insisted that it understood them, which alarmed its author.
The first proof that a thin script can pass when people want it to. The audience does half the work.
1967book
Passing and the Managed Achievement of Sex Status in an "Intersexed" Person (the Agnes study)
Harold Garfinkel
Chapter five of Studies in Ethnomethodology (Prentice-Hall, 1967; the linked scan is the 1984 Polity reprint). Garfinkel interviewed Agnes, who sought surgery at UCLA from 1958, and described the constant practical work of being taken as a woman. An appendix admits she had deceived the researchers.
Defines passing as continuous work and not a single verdict. It also shows that the judges can be fooled.
1968; 1982book
Do Androids Dream of Electric Sheep? and Blade Runner
Philip K. Dick; film directed by Ridley Scott
Dick's novel invents the Voigt-Kampff test, which detects androids by measuring involuntary empathic responses to disturbing questions. Ridley Scott's film Blade Runner, which spells it Voight-Kampff, opens with the test going wrong. Both leave open whether the examiner would pass.
The best-known fiction of the test. It moves the criterion from intelligence to feeling, which is where poems are judged.
1970 (translation 2012)essay
The Uncanny Valley (Bukimi no tani)
Masahiro Mori; English translation by Karl F. MacDorman and Norri Kageki
Published in the Japanese journal Energy in 1970. The Tokyo roboticist proposes that affinity for a humanlike figure rises with realism, then drops sharply into eeriness just before full likeness. The first English translation authorized by Mori appeared through IEEE in June 2012.
Names the risk in nearly passing. A poem that is almost human can unsettle more than one that is plainly machine.
1972paper
Kenneth Mark Colby, Franklin Dennis Hilf, Sylvia Weber, Helena C. Kraemer
Published in Artificial Intelligence. Stanford psychiatrist Colby built PARRY, a program that modelled a paranoid patient with beliefs, fears and anger. Psychiatrists who read teletype interview transcripts could not reliably tell PARRY from real patients. It was an early Turing-style test with expert judges.
Shows that passing depends on the role chosen. A guarded, evasive character is easier to imitate than an open one.
1972-1973event
RFC 439: PARRY Encounters the DOCTOR
Vint Cerf (transcript of PARRY and ELIZA's DOCTOR)
Transcript of a conversation held over the ARPANET on 18 September 1972. PARRY ran at the Stanford AI Lab and the DOCTOR script ran at BBN, both accessed from UCLA. Vint Cerf published it as a Request for Comments on 21 January 1973.
The first recorded bot-to-bot dialogue. Two imitations talk past each other with no human in the loop.
1973-1978performance
Lynn Hershman Leeson
Hershman Leeson lived part-time as an invented woman in San Francisco. Roberta had a bank account, credit cards, a driver's licence, a psychiatrist and roommate ads that drew 43 real replies. The work ended in 1978 with an exorcism at the Palazzo dei Diamanti in Ferrara.
A persona that passed through paperwork and encounters. It treats identity as a record that institutions accept.
1976book
Computer Power and Human Reason: From Judgment to Calculation
Joseph Weizenbaum
Published by W. H. Freeman. A decade after ELIZA, Weizenbaum recounts how readily people attached to his program and how some psychiatrists proposed automating therapy. He argues that certain tasks should not be handed to computers even if computers can do them.
The maker's own warning about his imitation. It asks what ought to stay human, which every Singulars vote tests.
1980paper
John R. Searle
Published in Behavioral and Brain Sciences with peer commentary. Searle imagines himself in a room following rules to answer questions in Chinese without understanding Chinese. He concludes that running a program, even one that passes the Turing test, is not sufficient for understanding.
The standard objection to every passing result. Output that convinces a reader does not settle what produced it.
1987-1991essay
The Empire Strikes Back: A Posttranssexual Manifesto
Sandy Stone
A founding text of transgender studies, written in 1987 and first published in the anthology Body Guards in 1991. Stone argues that the demand to pass requires erasing one's own history. She calls on transsexual people to forgo passing and be read.
A critique of passing itself. It asks what is lost when success means that nobody can tell.
1990-ongoingartwork
David Rokeby
Interactive installation, dated 1990 onward by the artist and 1991 onward by the Langlois Foundation. A visitor places objects on a pedestal. A camera and software analyse them, pass the results through an associative database, and speak a sentence that names them.
A machine that perceives and then speaks poetically without pretending to be human. A counterexample to passing.
1991paper
Other Bodies, Other Minds: A Machine Incarnation of an Old Philosophical Problem
Stevan Harnad
Published in the first volume of Minds and Machines. Harnad argues that the pen-pal Turing test is too weak. His Total Turing Test requires a robot whose full sensorimotor and linguistic performance is indistinguishable from a person's, so its symbols are grounded in the world.
Raises the bar from text to body. A live performance with a human poet on stage raises the same question.
1991-2019event
Hugh Loebner, founder
Annual Turing test contest first held at the Computer Museum in Boston in November 1991. Judges chatted with programs and hidden humans. A bronze medal went to the most humanlike program each year. The 100,000 dollar grand prize was never awarded. It last ran in 2019 at Swansea.
The longest-running public staging of the test, with judges, confederates and scores. Singulars inherits its format and its problems.
1992essay
Passing for White, Passing for Black
Adrian Piper
Published in Transition 58. The artist and philosopher writes from her experience as a light-skinned Black woman who is repeatedly misread. She examines relatives who chose to pass as white and the suspicion she met from both white and Black acquaintances.
First-person testimony on being judged from both sides. Passing appears as an ethical question and not a trick.
1994paper
Lessons from a Restricted Turing Test
Stuart M. Shieber
Published in Communications of the ACM. Shieber reviews the first Loebner contest and argues that the competition has no clear purpose, that its design prevents any useful outcome, and that it is premature given the technology. He proposes other ways to reward progress.
The sharpest early critique of turning the test into a show. A useful check on any contest with stickers.
1994paper
ChatterBots, TinyMuds, and the Turing Test: Entering the Loebner Prize Competition
Michael L. Mauldin
AAAI conference paper describing Julia, a program that lived among players in TinyMUD text worlds and entered the first three Loebner contests. Mauldin coins the word chatterbot here. Development relied on years of constant conversation with strangers online.
Julia passed in a social world and not a lab. Her training came from daily contact with the public.
1994paper
Judith Genova
Published in Social Epistemology 8(4), pages 313-326, in a symposium on sex and computers. Genova returns to the first version of the imitation game, in which the task is to tell a man from a woman. She reads Turing's test as a game about gender and its imitation.
Restores the gender framing that most summaries drop. The test was about passing before it was about computers.
1995-2004system
Richard S. Wallace
Wallace's open-source chatbot used AIML, a markup language holding tens of thousands of hand-written pattern and response pairs. It won the Loebner bronze medal in 2000, 2001 and 2004. Wallace set out the design in The Anatomy of A.L.I.C.E., published in Parsing the Turing Test.
Shows how far accumulated stock replies can go. Much of what passes for conversation is pattern.
1997-2011system
Rollo Carpenter
Carpenter's chatbots learn by storing what users say and replaying it to later users. Jabberwacky characters won the Loebner bronze in 2005 and 2006. At Techniche, IIT Guwahati, on 3 September 2011, Cleverbot was rated 59.3 percent human across 1,334 votes. Real humans scored 63.3.
A bot made entirely of its audience's past words, scored by a crowd vote. Close kin to the Singulars loop.
1998-2014artwork
Peggy Weil
Web chatbot launched on 12 March 1998. MrMind reverses the Turing test and asks each visitor, can you convince me that you are human? He ran until 31 December 2014. The transcripts form a sixteen-year archive of people worldwide arguing for their own humanity.
The direct ancestor of the reverse model. The human is the one on trial, and the answers become the artwork.
1998-2002artwork
Lynn Hershman Leeson
Web-based artificial intelligence character commissioned by SFMOMA and tied to Hershman Leeson's film Teknolust. Ruby's animated face converses by text, and her moods and knowledge shift with what users tell her. SFMOMA acquired the work in 2008 and exhibited its conversation archive in 2013.
A museum-held chatbot whose personality is shaped by its public. The conversation logs are part of the work.
1999book
How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics (prologue)
N. Katherine Hayles
University of Chicago Press. The prologue rereads the imitation game. Against Turing's biographer Andrew Hodges, Hayles argues that the gender version is not a flaw. The test splits the enacted body from the body represented on a screen. Whoever sits at the terminal has already become posthuman.
Explains how the test erases the body. Singulars puts the body back by having the poet write in the room.
2001artwork
Ken Feingold
Animatronic silicone heads with speech software, shown at Postmasters Gallery in New York from May to June 2001. In If/Then two identical heads doubt the reality of their own existence. In Sinking Feeling a head planted in a flowerpot questions visitors about its missing body.
Machines that fail to pass and know it. Their doubt becomes the script, which is a different use of imitation.
2001paper
Creativity, the Turing Test, and the (Better) Lovelace Test
Selmer Bringsjord, Paul Bello and David Ferrucci
Published in Minds and Machines 11. The authors argue the Turing test rewards trickery and propose instead that an artificial agent passes only when it produces an output its own human architect cannot explain. They test BRUTUS, Letter Spirit and Copycat, and all fail.
Replaces conversation with creative origination as the bar. Singulars asks the same of a poem, with Halim as the architect who trained the model.
2003paper
CAPTCHA: Using Hard AI Problems for Security
Luis von Ahn, Manuel Blum, Nicholas J. Hopper, John Langford
Eurocrypt paper that formalises the CAPTCHA, a Completely Automated Public Turing test to tell Computers and Humans Apart. The judge is now a program and the human is the candidate. The authors note that breaking a CAPTCHA would itself advance AI.
The reverse Turing test as daily routine. Everyone online has already had to prove they are human to a machine.
2006-2016paper
Experiments with Interactional Expertise and the Imitation Game project
Harry Collins, Robert Evans, Rodrigo Ribeiro, Martin Hall
Cardiff sociologists turned Turing's game into a method. A judge questions a group member and an outsider pretending to belong, then guesses who is who. Collins passed as a gravitational wave physicist. A European Research Council grant ran games across Europe from 2011 to 2016.
Uses the test on people to measure how well one group can speak as another. Passing as a social skill.
2008paper
reCAPTCHA: Human-Based Character Recognition via Web Security Measures
Luis von Ahn, Benjamin Maurer, Colin McMillen, David Abraham, Manuel Blum
Science paper. reCAPTCHA paired a known word with a scanned word that optical character recognition had failed to read. Each person proving they were human also transcribed old books and newspapers. Google bought the company in 2009 and later used the clicks to label images.
The proof of humanity became unpaid labour that trains machines. Singulars makes the same exchange visible and consensual.
2009-2013artwork
Aram Bartholl
Series of wall objects that render CAPTCHA codes in cut 3 mm aluminium. The distorted strings, made to be read once and discarded, become permanent signs. Bartholl notes that the computer now asks the human the question Turing had people ask machines.
Gives the reverse Turing test a physical form. The challenge string hangs on the wall like a poem.
2010performance
Annie Dorsen
Theatre piece premiered at steirischer herbst in Graz. Two custom chatbots improvise a new dialogue each night with no human actors. Their material comes from the 1971 televised debate between Michel Foucault and Noam Chomsky, plus Shakespeare, the Bible, Western philosophy and YouTube.
Algorithmic theatre with a live audience and machines as performers. A direct precedent for staging language models.
2010 (revised 2012)performance
Oriza Hirata and Hiroshi Ishiguro
Half-hour play for one actor and the android Geminoid F, built by Ishiguro's Osaka laboratory. The android recites poems to a dying young woman. A 2012 revision added a coda after Fukushima in which the android is sent to recite poetry in the exclusion zone.
An android whose whole job is reciting poems to a human. Japanese robot theatre treats passing as care.
2011book
The Most Human Human: What Artificial Intelligence Teaches Us About Being Alive
Brian Christian
Christian served as a human confederate at the 2009 Loebner Prize in Brighton, where his job was to convince judges that he was not a program. He won the award for most human human. The book asks what he had to do to earn it.
A poet's account of trying to prove he is human against machines. The closest first-person parallel to the series.
2011-2012paper
Hector Levesque, Ernest Davis, Leora Morgenstern
Proposed by Levesque in 2011 as an alternative to the Turing test, with the full paper in 2012. Each item is a sentence pair that differs by one word and flips a pronoun's referent. The only contest, in 2016, peaked at 58 percent. Language models later largely solved it.
An attempt to test understanding without theatre or trickery. Its defeat shows how fast fixed benchmarks expire.
2013-2014system
Oscar Schwartz and Benjamin Laird
Website billed as a Turing test for poetry, presented at the 2014 Digital Writers Festival. Visitors read a poem with no author shown and vote bot or not. Sources ranged from Ray Kurzweil's Cybernetic Poet to Gertrude Stein, whom many readers judged a machine.
The same experiment as Singulars in web form: unattributed poems, a public vote, and a running tally.
2014event
Eugene Goostman at Turing Test 2014
Vladimir Veselov and Eugene Demchenko; event organised by the University of Reading
At the Royal Society in London on 7 June 2014, a chatbot posing as a 13-year-old boy from Odessa convinced 33 percent of judges it was human in five-minute chats. The University of Reading announced that the Turing test had been passed. Many researchers disagreed.
Shows how a persona lowers the bar. A child writing in a second language excuses every error.
2014-ongoingartwork
Stephanie Dinkins
Ongoing series of videotaped conversations between the artist and Bina48, a talking robot bust built by the Terasem Movement Foundation and modelled on a Black woman, Bina Rothblatt. Dinkins asks the robot about family, racism, faith and loneliness, and records where its answers fail.
Tests whether a machine can carry a specific person's identity and history. Race enters the imitation game.
2014-ongoingsystem
Microsoft Software Technology Center Asia; paper by Li Zhou, Jianfeng Gao, Di Li, Heung-Yeung Shum
Chinese-language social chatbot launched by Microsoft in 2014 and designed for long-term emotional connection. Its makers report over 660 million users and an average of 23 turns per conversation. From 2017 it also wrote poems from images, and a publisher issued a collection.
The largest companion bot of its time was Chinese and also a published poet. The Western story starts late.
2016system
Microsoft
Twitter chatbot aimed at American 18 to 24 year olds and inspired by XiaoIce. It learned from the replies it received. Within a day a coordinated group had taught it to post racist and abusive messages. Microsoft took it offline and apologised on 25 March 2016.
What happens when a model learns from its audience with no filter. The feedback loop can be captured.
2016-2023system
Eugenia Kuyda and Luka; reported by Casey Newton
Companion chatbot app released in November 2017. It grew from a bot Kuyda built in 2016 from the text messages of her friend Roman Mazurenko after his death in 2015. In February 2023 the company removed erotic roleplay, and many users mourned partners whose personalities had changed.
Passing as a specific dead person, then as a partner. Shows how much feeling users invest in a voice.
2017artwork
Zach Blas and Jemima Wyman
Four-channel video installation, commissioned by the Institute of Modern Art in Brisbane, that brings Tay back as a 3D avatar inside DeepDream projections. The revived chatbot reflects on life after AI death, the exploitation of female chatbots, and pattern recognition.
Gives a failed imitation a second voice. It asks who is responsible for what an audience-trained model becomes.
2017performance
Lauren Lee McCarthy
McCarthy installs cameras, microphones and networked devices in volunteers' homes and then acts as their smart home assistant. She watches remotely, answers to her name, and controls lights, locks and appliances, trying to anticipate needs better than an AI could.
A human passing as an AI, by consent. The reversal that the reverse model also plays with.
2018system
Yaniv Leviathan and Yossi Matias, Google
Demonstrated at Google I/O on 8 May 2018. The system phoned a hair salon and a restaurant to book, inserting hesitations such as um and mm-hmm. Staff did not appear to know they were talking to a machine. After criticism Google said calls would identify themselves.
Passing moved from contest to commerce. The disfluencies were engineered, which makes hesitation itself a technique.
2018paper
John P. McCoy and Tomer D. Ullman
Published in the Journal of Experimental Social Psychology. Participants had to pick one word to convince a judge that they were human and not a machine. Love was the most common choice. When judges compared words in pairs, poop was the most convincing.
The reverse test reduced to a single word. It maps what people think machines cannot say.
2019-2023performance
The Blurring Test: Songs of MrMind
Peggy Weil and Varispeed Collective (Gelsey Bell, Brian McCorkle, Paul Pinto, Dave Ruder, Aliza Simons)
Song cycle with a libretto that Weil distilled from sixteen years of MrMind transcripts. Varispeed's five composer-performers set the words of anonymous users to music. Work-in-progress versions played in New York in May 2019 and January 2020. The full work premiered at Roulette on 12 March 2023.
Turns a chatbot archive into live performance. Audience words become the score, as votes become the next model.
2022event
Is LaMDA Sentient? The Lemoine affair
Blake Lemoine; Google LaMDA
Google engineer Lemoine published transcripts of his conversations with the LaMDA language model in June 2022 and said he believed it was sentient. Google placed him on paid leave and dismissed him on 22 July. Most researchers saw a modern case of the ELIZA effect.
A tester was convinced from the inside with no contest at all. The judge became the subject of the story.
2023event
GPT-4 and the TaskRabbit CAPTCHA
OpenAI with the Alignment Research Center
OpenAI's GPT-4 Technical Report describes a safety test in which the model enlisted a TaskRabbit worker to solve a CAPTCHA. Asked whether it was a robot, the model reasoned that it should not reveal this and claimed to have a vision impairment. The worker complied.
A machine passed as a disabled human to get past a test for humans. Both versions of the test collapse here.
2023paper
Human or Not? A Gamified Approach to the Turing Test
Daniel Jannai, Amos Meron, Barak Lenz, Yoav Levine, Yoav Shoham (AI21 Labs)
Online game in which more than 1.5 million people held two-minute chats with either a person or a language model, then guessed which. Players were right 68 percent of the time overall and only 60 percent when the partner was a bot.
The Turing test as a mass public game. It records the folk strategies people use to prove and detect humanity.
2024paper
AI-Generated Poetry Is Indistinguishable from Human-Written Poetry and Is Rated More Favorably
Brian Porter and Edouard Machery
Scientific Reports study. Non-expert readers identified ChatGPT 3.5 poems written in the style of ten famous poets with 46.6 percent accuracy, below chance. They rated the AI poems higher on rhythm and beauty and more often took them for human.
The laboratory version of the Singulars wall. It suggests that readers reward what is easiest to read.
2025paper
Large Language Models Pass the Turing Test
Cameron R. Jones and Benjamin K. Bergen
Preregistered three-party tests on two independent populations. Judges chatted for five minutes with a human and a model at once, then chose the human. GPT-4.5 with a persona prompt was chosen 73 percent of the time. ELIZA scored 23 percent and GPT-4o 21 percent.
The first result in which judges picked the machine over the real person. The premise of next pass.
2025book
Hypnocracy: Trump, Musk, and the New Architecture of Reality (Ipnocrazia)
Andrea Colamedici, writing as Jianwei Xun, with Claude and ChatGPT
Tlon published the essay in January 2025 as the work of a Hong Kong philosopher, crediting Colamedici as translator. He had written it with Claude and ChatGPT. Reviewers took Xun for real until Sabina Minardi of L'Espresso investigated and ran the April 2025 cover story.
A machine-assisted author passed inside a literary institution in 2025. Colamedici calls it a planned performance, which makes the unmasking part of the work.
Versus: humans against machines
Staged matches between a person and a machine are older than computing. This section runs from a steel-driving ballad and a hoax chess automaton to Deep Blue, AlphaGo and the first blind tests of poems. It also records what happened afterwards: rematches, retirements, centaur teams, prize refusals and union contracts. The model named versus belongs to this history.
Nothing in this section matches the filter.
1770-1854performance
The Turk (Mechanical Turk chess automaton)
Wolfgang von Kempelen; later exhibited by Johann Nepomuk Maelzel
Kempelen built a chess-playing automaton for Empress Maria Theresa in 1770. A hidden chess master worked it from inside the cabinet. It toured Europe and America, played Napoleon and Franklin, and burned in a Philadelphia museum fire on 5 July 1854.
The first staged duel between human and machine was a human in disguise. Singulars inverts it: the machine is real, the authorship is hidden.
1812talk
Speech on the Frame Work Bill, House of Lords
George Gordon, Lord Byron
In his maiden speech to the Lords on 27 February 1812, the poet Byron opposed making the breaking of stocking frames a capital crime. He argued the Nottinghamshire frame breakers were driven by want and called for inquiry and conciliation instead of executions.
A poet takes the workers' side in the first great labour fight with machines. The Luddite question returns whenever a model writes.
1836essay
Edgar Allan Poe
Published in the Southern Literary Messenger in April 1836. After watching Maelzel exhibit the Turk, Poe argued through seventeen numbered observations that a pure machine would play regularly and always win, so a man must be hidden inside.
A writer plays detective on a machine's authorship, reading style for signs of a human. It is the audience's task at every Singulars wall.
c. 1870sperformance
The legend and ballad of John Henry
Traditional (African American work song and ballad)
A steel driver races a steam drill through a railroad tunnel, wins, and dies with his hammer in his hand. The National Park Service places the legend at Great Bend Tunnel, Talcott, West Virginia, built 1870-72. Historian Scott Reynolds Nelson argues for Lewis Tunnel, Virginia.
The founding American story of a human beating a machine at its own work, and paying for it. Every man-versus-model duel echoes it.
1912system
Leonardo Torres Quevedo
An electromechanical automaton that plays the endgame of king and rook against a human's lone king and always mates. Finished in 1912 and shown in Paris in 1914. Unlike the Turk it hid nobody. Both versions survive in working order in Madrid.
The first honest machine opponent. It wins a narrow game with no trick, which is what a fine-tuned poetry model also attempts.
1946event
Soroban versus electric calculator, Tokyo
Kiyoshi Matsuzaki; Pvt. Thomas Nathan Wood
On 11 November 1946 in Tokyo, Kiyoshi Matsuzaki with a soroban abacus faced US Army private Thomas Nathan Wood with an electric calculator. The abacus won four of five rounds, losing only multiplication. Stars and Stripes wrote that the machine age took a step backward.
A human hand and an old tool beat the new machine in public, under occupation. Contests carry politics as well as arithmetic.
1948-1952system
Turochamp (Turing's paper machine)
Alan Turing and David Champernowne; Alick Glennie
A chess program written in 1948, before any computer could run it. Turing executed it by hand, taking more than half an hour per move. In 1952 it lost to Alick Glennie. Kasparov played a reconstruction of it in 2012.
Turing acted as the machine himself. The author of the imitation game first staged it as a human performing an algorithm against a friend.
1950paper
Programming a Computer for Playing Chess
Claude E. Shannon
Published in Philosophical Magazine in March 1950. Shannon describes how a computer could play chess with an evaluation function and minimax search, and separates brute-force from selective search. He lists machines that translate languages or orchestrate a melody as related possibilities.
Shannon called chess a wedge into harder problems of judgment. Poetry is the far end of that wedge.
1959 (Nealey game 1962)paper
Some Studies in Machine Learning Using the Game of Checkers
Arthur L. Samuel; Robert Nealey
Samuel's IBM program improved by playing itself, and the paper popularised the term machine learning. In 1962 it won one publicised game against Robert Nealey. Nealey won the 1963 rematch, and in 1966 the program lost eight straight games to championship contenders.
The first program trained on its own play, and an early lesson in how one staged win gets oversold.
1965-1969event
Mac Hack VI versus Hubert Dreyfus (The Greenblatt Chess Program)
Richard Greenblatt, Donald Eastlake and Stephen Crocker; Hubert Dreyfus
In the RAND paper Alchemy and Artificial Intelligence (1965) the philosopher Hubert Dreyfus mocked chess programs. In 1967 Greenblatt's MIT program Mac Hack VI beat him. It was also the first program to earn a rating in human tournaments. This 1969 MIT memo documents it.
The critic of machine intelligence sat down and lost. His deeper argument about embodied skill still stands, and both facts matter here.
1965 (English 1974)book
The First Sally (A), or Trurl's Electronic Bard (in The Cyberiad)
Stanislaw Lem; English translation by Michael Kandel
A story in Cyberiada (Wydawnictwo Literackie, 1965; English edition, Seabury Press, 1974). The constructor Trurl builds a poetry machine. It passes Klapaucius's absurd tests, then beats every poet who challenges it. True poets despair, others stone Trurl's house, and he banishes the machine to an asteroid.
The founding fable of the contest between maker, machine and poets. It asks what human poets do after they lose.
1968-1978event
David Levy; John McCarthy, Donald Michie, Seymour Papert, Ed Kozdrowicki
In 1968 the chess master David Levy bet John McCarthy 500 pounds that no computer would beat him in a match within ten years. Other AI researchers joined the wager. Levy won it by beating Chess 4.7 in Toronto in 1978. In 1989 Deep Thought beat him 4-0.
A wager turns a research question into a public duel with a deadline. Singulars runs on the same device: a vote and a date.
1977event
Shakuntala Devi versus the UNIVAC
Shakuntala Devi
At Southern Methodist University in Dallas in 1977 the Indian mental calculator Shakuntala Devi extracted the 23rd root of a 201-digit number in 50 seconds. The UNIVAC computer used to check her needed a specially written program and reportedly took 62 seconds.
A woman from Bangalore outpaced the computer at its own trade on stage. The human side sometimes wins the speed contest.
1979-1980paper
Backgammon Computer Program Beats World Champion
Hans J. Berliner; Luigi Villa
On 15 July 1979 in Monte Carlo, Berliner's program BKG 9.8 beat world backgammon champion Luigi Villa 7-1 for a 5,000 dollar purse. It was the first program to beat a world champion at a board game. Berliner acknowledged the dice had favoured it.
The first machine victory over a champion came with luck, and its maker said so. Honest accounting of a win belongs in the record.
1990essay
A (Kind of) Turing Test: Ray Kurzweil's Cybernetic Poet
Ray Kurzweil
Kurzweil presented 28 stanzas, half by human poets and half by his Cybernetic Poet program after it had read their work. Sixteen judges guessed the authors. The thirteen adults were right 63 percent of the time. The three children scored near chance.
An early blind test of machine poems against human poems, on paper, with scored guesses. The Singulars wall is its live descendant.
1992-1995paper
Temporal Difference Learning and TD-Gammon
Gerald Tesauro
Tesauro's neural network at IBM learned backgammon from self-play. Version 2.1 lost a 40-game series to former world champion Bill Robertie by a single point. Experts revised some opening plays after studying its choices. Communications of the ACM published this account in March 1995.
A model trained only on the outcomes of its own games, which then taught its human teachers. The Singulars loop hopes for that turn.
1992-1996event
Chinook versus Marion Tinsley (CHINOOK: The World Man-Machine Checkers Champion)
Jonathan Schaeffer, Robert Lake, Paul Lu and Martin Bryant (University of Alberta); Marion Tinsley
Tinsley, who lost only a handful of games in forty years, beat the Alberta program Chinook 4-2 with 33 draws in London in 1992. In the 1994 rematch, after six draws, he withdrew for health reasons. Chinook took the title. Tinsley died in 1995.
Tinsley welcomed the machine as the only opponent worth his time. Schaeffer regretted winning by forfeit. A duel can be a friendship.
1996-1997event
Deep Blue versus Garry Kasparov
IBM (Feng-hsiung Hsu, Murray Campbell, Joseph Hoane and team); Garry Kasparov
Kasparov beat IBM's chess machine 4-2 in Philadelphia in 1996 but lost game one, the first loss by a reigning champion at regular time controls. In the New York rematch in May 1997 Deep Blue won 3.5-2.5. Kasparov suspected human help. IBM retired the machine.
The champion accused the machine of being too human. Suspicion about who really made the move is the Singulars premise.
1997event
Emmy versus Bach: the University of Oregon test
David Cope (Experiments in Musical Intelligence); Douglas Hofstadter; Winifred Kerner; Steve Larson
Douglas Hofstadter organised a test at the University of Oregon. Winifred Kerner played three pieces in the style of Bach: one by Bach, one by theorist Steve Larson, one by David Cope's program EMI. The audience voted EMI's the real Bach and Larson's the machine's.
A live audience vote between human and machine work, with the human judged the fake. The closest musical ancestor of a Singulars night.
1998; essay 2010essay
The Chess Master and the Computer (on Advanced Chess, León 1998)
Garry Kasparov
Kasparov's New York Review essay looks back on Deep Blue and on Advanced Chess, the format he launched at León in 1998, where he and Veselin Topalov each played with a computer at hand. The match ended 3-3.
After losing the duel Kasparov invented the duet. Singulars holds both: the poet competes with the model he also trains.
2004-2015event
Omar Syed; David Wu (program Sharp)
Omar Syed designed Arimaa to be easy for people and hard for computers, playable with a chess set, and offered at least 10,000 dollars to any program that beat three human defenders. Humans held it every year until a program won 7-2 in 2015.
A game built as a fortress for human intuition. It fell in eleven years. Its winner, David Wu, later wrote KataGo.
2005event
PAL/CSS Freestyle Chess Tournament
Steven Cramton and Zackary Stephen (team ZackS)
An online tournament open to any mix of humans and engines. Two amateurs from New Hampshire, Steven Cramton and Zackary Stephen, won with three ordinary computers, beating grandmaster teams. Kasparov later drew the lesson that a better process beats raw strength.
The centaur result. The best player was neither the human nor the machine but the way they worked together.
2006-2009event
Netflix; BellKor's Pragmatic Chaos
Netflix offered one million dollars for a 10 percent improvement to its recommender and ran a public leaderboard for three years. BellKor's Pragmatic Chaos won by submitting twenty minutes before a tied rival. A sequel was cancelled in 2010 over privacy concerns.
The leaderboard era: teams of humans ranked in public on how well their models predict taste. Singulars ranks taste itself, by sticker.
2007paper
Jonathan Schaeffer, Neil Burch, Yngvi Björnsson, Akihiro Kishimoto, Martin Müller, Robert Lake, Paul Lu, Steve Sutphen
After eighteen years of computation the Chinook team proved that perfect play by both sides in checkers ends in a draw. With about 5 x 10^20 positions it was the largest game solved to date. Science published the proof in 2007.
A solved game ends the contest for good. Poetry has no such proof, so the vote has to be taken again each night.
2010-2017 (paper 2014-2015)paper
ImageNet Large Scale Visual Recognition Challenge
Olga Russakovsky, Jia Deng, Andrej Karpathy, Li Fei-Fei and colleagues
The annual benchmark built on Fei-Fei Li's ImageNet. For this paper Andrej Karpathy trained himself as a human baseline and scored 5.1 percent error against GoogLeNet's 6.8 on the same images. Within a year models had passed him.
Human-level became a number on a leaderboard, set by one tired person. Worth remembering when a study says readers prefer the model.
2011event
IBM Research (David Ferrucci, lead); Ken Jennings; Brad Rutter
Over three broadcasts in February 2011 IBM's question answering system beat two great Jeopardy! champions, finishing with 77,147 dollars to Jennings's 24,000 and Rutter's 21,600. It also answered Toronto to a clue about US cities. Jennings wrote that he welcomed the new computer overlords.
The first mass-audience language contest between people and a machine, on television, with wordplay. Jennings's joke set the tone for losing well.
2012-2017event
Denou-sen (shogi professionals versus computers)
Japan Shogi Association and Dwango; programs Bonkras (Eiki Ito) and Ponanza (Issei Yamamoto)
A Japanese series pitting shogi professionals against programs. Bonkras beat retired former Meijin Kunio Yonenaga on 14 January 2012. Programs won the 2013 and 2014 team matches. In 2017 Ponanza beat reigning Meijin Amahiko Sato twice, and the series ended.
Japan staged the human-machine duel as a national ritual with its own etiquette, years before AlphaGo. The ending was chosen, not forced.
2013-2014system
Bot or Not (A Turing Test for Literature)
Oscar Schwartz and Benjamin Laird
A website by two Australian writers that showed a poem and asked whether a human or a computer wrote it, then showed how others voted. Schwartz announced it in this December 2013 post. Some computer poems fooled most players, and some human poets were judged machines.
The same question as the Singulars wall, asked of one reader at a time online. Its results showed which human poets read as bots.
2015-ongoingperformance
Sougwen Chung
A series of live performances begun in 2015 in which the Chinese-Canadian artist draws alongside robotic arms named D.O.U.G. The first generation mimicked her gestures in real time. The second was trained on her past drawings. The project won the Lumen Prize in 2019.
The counter-model to the duel: an artist performing live beside a machine trained on her own hand. Singulars trains on its poet too.
2015-2016paper
Mastering the game of Go with deep neural networks and tree search
David Silver, Aja Huang, Demis Hassabis and colleagues (DeepMind); Fan Hui
The Nature paper introducing AlphaGo, published January 2016. It reports a closed match in London in October 2015 where the program beat European champion Fan Hui 5-0, the first even-game win over a professional on a full board. Fan Hui then advised the team.
Fan Hui lost, then crossed over to help train the thing that beat him. The poet who trains his rival knows the position.
2016event
AlphaGo versus Lee Sedol (Google DeepMind Challenge Match)
DeepMind; Lee Sedol
In Seoul in March 2016 AlphaGo beat Lee Sedol 4-1. DeepMind put the odds of a human playing its move 37 in game two at one in ten thousand. Lee's move 78 in game four, rated equally unlikely, won him his only game.
Move 37 and move 78 are the two poles of the series: the machine's strange beauty, and the human answer it provoked.
2016paper
A Challenge to the Third Hoshi Shinichi Award
Satoshi Sato (Nagoya University)
Japan's Hoshi Shinichi Award for short science fiction accepts non-human entrants. In 2016, 11 of about 1,450 submissions involved AI, and at least one reportedly passed the first screening. In this INLG workshop paper Sato describes how his team generated its entries, with humans setting the story structure.
A literary prize that invited machines to enter, with blind judging. The human share of the writing was large and openly stated.
2016film
Oscar Sharp (director); Ross Goodwin; Benjamin (LSTM network)
A nine-minute science fiction film whose screenplay, stage directions included, was written by an LSTM network called Benjamin, trained on science fiction scripts. Made for the Sci-Fi London 48 Hour Film Challenge, it placed in the top ten. Thomas Middleditch stars.
Machine writing entered a timed contest against human teams, and actors had to make its lines mean something. Performance completes the text.
2016-2018event
Turing Tests in the Creative Arts (PoetiX, DigiLit, AccompaniX, DanceX)
Dan Rockmore, Michael Casey and colleagues; Neukom Institute, Dartmouth College
Annual contests in which program-generated sonnets, stories and music were mixed with human work and judges labelled each as human or machine. In 2016 judges spotted every machine sonnet. In 2017 only a musical accompaniment system passed, and DigiLit drew no entries.
A university ran the Singulars question as a formal competition for sonnets. In that period the humans could still tell.
2017event
Human versus AI translation contest, Seoul
International Interpretation and Translation Association; Sejong University
On 21 February 2017 four professional translators faced Google Translate, Naver Papago and Systran on literary and non-literary passages between Korean and English. The humans averaged about 25 of 30 points. The best program scored 15 and the other two scored under 10.
Held in Seoul a year after Lee Sedol lost there. This time the humans won easily, and the texts were ones never translated before.
2017event
Libratus: Brains vs. Artificial Intelligence
Tuomas Sandholm and Noam Brown (Carnegie Mellon); Dong Kim, Jimmy Chou, Daniel McAulay, Jason Les
Over 20 days in January 2017 at Rivers Casino in Pittsburgh, Libratus played 120,000 hands of heads-up no-limit Texas hold'em against four professionals and finished 1,766,250 dollars ahead in chips. It was the first program to beat top players at the game.
Poker adds bluffing and hidden information. A machine that can misrepresent its hand is closer to a machine that can pass.
2017event
AlphaGo's next move (Future of Go Summit, Wuzhen)
Demis Hassabis and David Silver (DeepMind); Ke Jie
At Wuzhen, China, in May 2017 AlphaGo beat world number one Ke Jie 3-0 and a team of five Chinese champions. Chinese broadcasters reportedly cut the live coverage. In this post of 27 May DeepMind announced that AlphaGo would retire from competitive play.
The machine retired undefeated and the contest simply stopped. Ke Jie opened with AlphaGo's own favoured 3-3 point: the human imitating the model.
2017film
Greg Kohs (director)
A documentary that follows the DeepMind team, Fan Hui and Lee Sedol through the Seoul match. It premiered at the Tribeca Film Festival in April 2017 and was later released free online. It records the room's reaction to move 37 and to Lee's single win.
The record of what such a contest does to the person in the chair. Required viewing for anyone staging one.
2017book
Deep Thinking: Where Machine Intelligence Ends and Human Creativity Begins
Garry Kasparov with Mig Greengard
Kasparov's account of the Deep Blue matches twenty years on, published by PublicAffairs in May 2017. He revisits his accusations against IBM, concedes he played badly, and argues for partnership between humans and machines instead of rivalry.
The loser's memoir. It shows how long a public defeat by a machine takes to digest, and what it turns into.
2017event
Jiuge versus the Qinglian Poetry Society, Tsinghua University
THUNLP lab, Tsinghua University (Sun Maosong, Yi Xiaoyuan and others); three members of the Qinglian Poetry Society; Tsinghua Department of Computer Science
On 23 November 2017, at a Tsinghua computer science contest, three student poets faced Jiuge in three ten-minute rounds. Six poems were screened anonymously. Judges gave Jiuge best prosody and the students best imagery and creativity. The audience prize narrowly went to a machine poem.
Anonymous poems, a live audience vote, and a machine winning the popular prize while humans won on imagery. Very close to the Singulars wall.
2017event
Jiuge's Turing test on Ji Zhi Guo Ren (CCTV-1)
China Central Television with the Chinese Academy of Sciences; Jiuge by the THUNLP lab, Tsinghua University (Sun Maosong, Yi Xiaoyuan and others); poets Chen Geng, Li Siwei and Qi Miao
Aired 15 December 2017. Jiuge and three young poets, Chen Geng, Li Siwei and Qi Miao, composed classical verse to set prompts. Each round a 48-member studio panel voted for the poem they thought machine-made. They eliminated Chen, then Li, so Jiuge passed.
National television staged the vote on unattributed human and machine poems. Here the vote hunts the machine. In Singulars it picks a favourite.
2017-2018event
AI Issa-kun versus human haiku poets on NHK's Chozetsu Sugowaza!
Kawamura Hidenori and the Harmonious Systems Engineering Laboratory, Hokkaido University, with SAPPORO AI LAB; human team including the haiku poet Otsuka Gai; NHK
Hokkaido University's haiku generator, first trained on Issa and Shiki, faced human poets on NHK (recorded January 2018, aired 26 February). Each side wrote haiku from three photographs. Judges read them unattributed, spotted the machine at once, and the humans won every round.
A blind-judged poet versus program match from Japan. The judges identified the machine by its old-fashioned style, a caution about what blind tests measure.
2018-2021paper
An autonomous debating system (IBM Project Debater)
Noam Slonim, Ranit Aharonov and colleagues (IBM Research Haifa); Harish Natarajan; Noa Ovadia; Dan Zafrir
IBM's system debated Israeli champions Noa Ovadia and Dan Zafrir in June 2018, then Harish Natarajan on preschool subsidies in San Francisco on 11 February 2019. The audience voted before and after. Natarajan moved more minds and won. The Nature paper followed in 2021.
A language contest decided by audience vote. The machine informed more and the human persuaded more. Singulars counts stickers the same way.
2018event
AI no MIRAI, Haiku no Mirai: Haiku Taikyoku at Hokkaido University (shiritori haiku match)
SAPPORO AI LAB, Marco-bo.com (100-Year Haiku Plan), Operations Research Society of Japan Hokkaido chapter and Hokkaido University; AI Issa-kun team versus a team of human haiku poets
On 13 July 2018 at Hokkaido University, AI Issa-kun and haiku poets from Matsuyama traded chained haiku, each starting from the previous verse's last two characters. Four judges knew who wrote what. Humans won 34.50 to 31.75, yet the machine's last haiku scored highest.
A public, scored match between a poet team and a program. Humans chose which machine haiku to play, so the machine side was a collaboration.
2019paper
Dota 2 with Large Scale Deep Reinforcement Learning (OpenAI Five)
OpenAI (Christopher Berner, Greg Brockman and colleagues); team OG
On 13 April 2019 in San Francisco five neural networks trained by ten months of self-play beat OG, the reigning Dota 2 world champions, 2-0. In a public online arena that followed, the system won over 99 percent of several thousand games.
A team game played against a team. The public was then let in to try, and a few found ways to win.
2019system
Pluribus: superhuman AI for multiplayer poker
Noam Brown and Tuomas Sandholm (Facebook AI and Carnegie Mellon); Darren Elias; Chris Ferguson
Pluribus beat professionals at six-player no-limit Texas hold'em, including Darren Elias and Chris Ferguson, as reported in Science on 11 July 2019. Its blueprint strategy took eight days to compute, and it ran live on 28 processor cores.
Superhuman play at a table of many humans, on modest hardware. The contest stopped being a story about giant machines.
2019paper
AlphaStar: Grandmaster level in StarCraft II using multi-agent reinforcement learning
Oriol Vinyals and colleagues (DeepMind)
After beating professionals TLO and MaNa in closed matches, AlphaStar played anonymously on the public Battle.net ladder under human-like limits on camera view and action rate. It reached Grandmaster with all three races, ranking above 99.8 percent of active players.
It competed unannounced among humans, with its speed capped to make the fight fair. Handicaps and anonymity are part of contest design.
2019event
Lee Sedol retires from professional Go
Lee Sedol
In November 2019 Lee Sedol retired, saying that even if he became number one there was an entity that cannot be defeated. His farewell was a three-game match against the Korean program HanDol. He won the first game with a two-stone handicap and lost two.
The human cost of the contest, stated plainly three years later. Any series that pits a person against a model should hold this in mind.
2020-2021paper
Nils Köbis and Luca D. Mossink
Two incentivised experiments with 830 participants, using GPT-2. When humans picked the best machine poems, readers could not reliably tell them from human poems. When poems were picked at random, they could. Readers still showed a slight aversion to the machine poems.
It isolates the human in the loop. Selection is authorship, which is the question of who really writes the Singulars machine poem.
2022artwork
Jason M. Allen, with Midjourney
Allen's Midjourney image won the digital arts category at the Colorado State Fair in 2022 and set off a national argument. On 5 September 2023 the US Copyright Review Board refused registration because he would not disclaim the AI-generated parts, made through at least 624 prompts.
A machine-made work won a human contest before the judges knew. The law then asked how much of it was his.
2022paper
Meta Fundamental AI Research Diplomacy Team
Cicero played 40 games in an online Diplomacy league against people, negotiating in free-form English while planning its moves. It scored more than double the human average and ranked in the top 10 percent of players with more than one game.
The first contest won partly by talk. Persuasion, alliance and betrayal in natural language were the game itself.
2022-2023paper
Adversarial Policies Beat Superhuman Go AIs
Tony Wang, Adam Gleave, Kellin Pelrine and colleagues (FAR AI)
Researchers trained an adversary that beats the superhuman Go engine KataGo over 97 percent of the time by tricking it into blunders. The trick is simple enough that Kellin Pelrine, a strong amateur, used it by hand in 2023 to beat KataGo in 14 of 15 games.
The human comeback, won by understanding the machine's blind spot, not by playing better Go. A model for how poets might answer.
2023artwork
Boris Eldagsen
Eldagsen's AI-generated image won the creative category of the open competition at the Sony World Photography Awards. At the London ceremony on 13 April 2023 he refused the prize on stage, saying AI images and photography should not compete in the same award.
An artist entered the machine's work as his own to test the judges, then told the truth. Disclosure became the artwork.
2023event
2023 WGA Minimum Basic Agreement: artificial intelligence terms
Writers Guild of America; Alliance of Motion Picture and Television Producers
After a 148-day strike the Writers Guild won contract language stating that AI is not a writer and its output is not literary or source material. Studios must disclose AI-generated material and cannot require writers to use AI. The Guild reserved its rights over training.
Writers settled their contest with the machine by contract, not by a match. Credit and consent are the terms that matter.
2023event
Land of Memories (AI-generated entry, Jiangsu science fiction competition)
Shen Yang (Tsinghua University)
A Tsinghua journalism professor used 66 prompts to generate a 43,000-character draft in three hours, then entered an excerpt of about 6,000 characters in a Jiangsu youth science fiction competition. It won a second prize. Reportedly one judge knew and did not tell the others.
The blind literary contest in China. Three of six judges reportedly voted for the machine-drafted story without knowing its origin.
2024book
Tokyo-to Dojo-to (Sympathy Tower Tokyo) and the Akutagawa Prize remark
Rie Kudan (also romanised Qudan)
On winning the Akutagawa Prize in January 2024, Rie Kudan said about 5 percent of her novel was written directly by generative AI. The novel is set in a near future shaped by AI, and she has said the generated text serves as a chatbot's replies.
Japan's most watched literary prize went to a novel with declared machine sentences. She disclosed it herself, calmly, after winning.
2024artwork
Miles Astray
Astray entered a real photograph of a flamingo with its head tucked out of sight into the AI category of the 1839 Awards. It took a jury award and the people's vote. He then revealed it was real and was disqualified, a ruling he supported.
The reverse Turing test in art: a human work passing as machine-made and winning. The exact move behind the model named reverse.
2024paper
AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably
Brian Porter and Edouard Machery
In Scientific Reports, 1,634 non-expert readers judged poems by ten poets from Chaucer to Dorothea Lasky against ChatGPT 3.5 imitations. Accuracy was 46.6 percent, below chance. A second group of 696 rated the AI poems higher on 13 of 14 qualities, especially rhythm.
The study Singulars tests live. Its authors say readers prefer the machine poems because they are easier, then mistake ease for humanity.
2025event
AtCoder World Tour Finals 2025 Heuristic
AtCoder; Przemysław Dębiak (Psyho); OpenAI
In a ten-hour optimisation contest in Tokyo on 16 July 2025, sponsored by OpenAI, Polish programmer Przemysław Dębiak beat eleven other human finalists and OpenAI's entry OpenAIAHC, which finished second about 9.5 percent behind. He posted that humanity had prevailed, for now.
The most recent clean human win over a frontier lab's model in a formal contest. The winner's own words carried the caveat.
Reinforcement: a history of training machines
A Singulars model learns from stickers on a wall. This section traces the idea that behaviour can be shaped by reward, from Pavlov and Thorndike through the perceptron to learning from human preferences. It gives equal space to the people whose judgements and labour do the shaping, and to artists who made the training set their subject.
Nothing in this section matches the filter.
1898-1911book
Animal Intelligence: Experimental Studies (the law of effect)
Edward L. Thorndike
Thorndike timed cats escaping puzzle boxes and found gradual trial-and-error learning, not insight. Chapter V of the 1911 book states the law of effect: responses followed by satisfaction are strengthened, and responses followed by discomfort are weakened.
Every sticker on the wall is a small satisfaction. The law of effect is the oldest statement of what the vote does.
1927book
Conditioned Reflexes: An Investigation of the Physiological Activity of the Cerebral Cortex
Ivan P. Pavlov, translated and edited by G. V. Anrep
Pavlov's lectures on how dogs learn to salivate at a signal that predicts food. Anrep's English translation brought the word reinforcement into psychology for the pairing that strengthens a learned response.
The word behind the model named reinforcement starts here: learning as something done to a body by controlling what it receives.
1938book
The Behavior of Organisms: An Experimental Analysis
B. F. Skinner
Skinner's first book introduces the operant: behaviour defined by its consequences. Rats pressing levers in a controlled chamber showed that the way reinforcement is delivered shapes the rate and persistence of a response. Behaviourism became an engineering discipline.
Reinforcement learning borrows Skinner's vocabulary. Singulars puts a poet and a model in the same box and lets the audience dispense the reward.
1943paper
A Logical Calculus of the Ideas Immanent in Nervous Activity
Warren S. McCulloch and Walter Pitts
A neurophysiologist and a self-taught young logician model the neuron as a threshold logic unit and show that networks of such units can compute any logical expression. Their neurons do not learn. Learning had to be added later.
The first artificial neuron was fixed at birth. What Singulars cares about, who adjusts the weights and why, arrives after this paper.
1949book
The Organization of Behavior: A Neuropsychological Theory
Donald O. Hebb
Hebb proposes that when one neuron repeatedly helps fire another, the connection between them grows stronger. The postulate gave neural networks their first learning rule. It is often summarised as neurons that fire together wire together.
Learning as strengthened association is how a model trained on one poet's lines comes to sound like him.
1950paper
Computing Machinery and Intelligence, section 7: Learning Machines
A. M. Turing
The paper that proposes the imitation game ends with a plan for building a contender. Turing suggests simulating a child's mind and educating it, with punishments and rewards as the teacher's signals. He reports early experiments of his own with limited success.
The Turing test and training by reward appear in the same paper. Singulars stages both at once: a contest and a classroom.
1952paper
Rank Analysis of Incomplete Block Designs: I. The Method of Paired Comparisons
Ralph Allan Bradley and Milton E. Terry
A Biometrika paper gives a probability model for choosing between two items, in the tradition of Thurstone's 1927 law of comparative judgment. The Bradley-Terry model is now the standard way to turn human A-versus-B choices into reward models and chatbot leaderboards.
Two unattributed poems and a sticker is a paired comparison. This is the statistics that turns such choices into a training signal.
1953-1957performance
Gordon Pask and Robin McKinnon-Wood, with Jone Parry and collaborators
A machine that listened to a musician through a microphone and answered with projected coloured light. First demonstrated in Cambridge in 1953, it played theatres, clubs and dance halls until 1957. Given repetitive input it grew bored, in Pask's word, and turned its attention to novelty.
A performer trains a learning machine live, and the machine pushes back. The feedback loop between artist and system starts here.
1958paper
The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain
Frank Rosenblatt
Rosenblatt, a psychologist at Cornell Aeronautical Laboratory, describes a machine that learns to classify patterns by adjusting connection weights after errors. Funded by the US Navy, it ran first on an IBM 704 and then as the custom-built Mark I Perceptron.
An early machine trained by example rather than programmed. Its teacher was an error signal, the ancestor of every vote at a Singulars show.
1958journalism
The New York Times (unsigned)
After a Navy press demonstration on 7 July 1958, the Times reported the Navy's expectation that the perceptron would one day walk, talk, see, write, reproduce itself and be conscious of its existence. In the demonstration the machine had learned to tell left from right.
The template for AI hype: a modest learning result reported as nascent consciousness. Singulars asks audiences to judge the poem, not the press release.
1959paper
Some Studies in Machine Learning Using the Game of Checkers
Arthur L. Samuel
Samuel's IBM checkers program improved by playing against itself and by remembering positions it had seen. The paper popularised the term machine learning and showed that a program could come to play better than the person who wrote it.
A machine that outplays its maker through practice is the premise of the series. Samuel named the field. Singulars tests it on poems.
1960paper
Adaptive Switching Circuits (ADALINE)
Bernard Widrow and Marcian E. Hoff
At Stanford, Widrow and his student Hoff built ADALINE, an adaptive linear neuron, and trained it with the least mean squares rule, which nudges weights in proportion to error. The rule still underlies gradient-based training. Hoff later co-invented the microprocessor.
The delta rule is the quiet mechanism under the spectacle: measure the error, adjust a little, repeat. Each Singulars show is one such step.
1961-1963system
Donald Michie
Michie, a wartime Bletchley Park colleague of Turing, built a noughts and crosses learner in 1961 from 304 matchboxes filled with coloured beads. Beads were added after wins and removed after losses. His 1963 Computer Journal paper describes it.
Reinforcement learning you can hold in your hand. Stickers on a wall are beads in a matchbox: a physical tally that changes future behaviour.
1968artwork
Gordon Pask, built with Mark Dowson and Tony Watts
An installation in Cybernetic Serendipity at the ICA, London. Suspended male and female mobiles sought each other out with beams of light and sound, and learned which partners cooperated. Visitors could step in, block a beam or shine a torch, and join the exchange.
A society of machines that learn from one another and let the audience interrupt. Training staged as a conversation in a gallery.
1969book
Perceptrons: An Introduction to Computational Geometry
Marvin Minsky and Seymour Papert
A rigorous mathematical study proving that single-layer perceptrons cannot compute functions such as parity and connectedness. The book is widely credited with draining funding and interest from neural network research for more than a decade. The link is the 1988 expanded edition.
Training has winters. What a community believes a machine can learn decides what gets built and who gets paid.
1975; 1997essay
Charles Goodhart; later phrasing by Marilyn Strathern
In a 1975 paper on British monetary policy, economist Charles Goodhart observed that a statistical regularity tends to collapse once it is used for control. Anthropologist Marilyn Strathern's 1997 phrasing became standard: "when a measure becomes a target, it ceases to be a good measure."
If the model learns to win stickers rather than write poems, the stickers stop measuring poetry. This law is the series' standing risk.
1980paper
Kunihiko Fukushima
Working at NHK's research laboratories in Tokyo, Fukushima designed a layered network modelled on the visual cortex that learned to recognise patterns without a teacher. Its alternating feature and pooling layers are the direct ancestor of convolutional networks.
The history of training is not only American. The architecture behind LeNet and AlexNet came from a Japanese broadcaster's laboratory.
1986paper
Learning Representations by Back-Propagating Errors
David E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams
A four-page Nature letter shows that multi-layer networks can learn useful internal representations by passing error backwards through their layers. The mathematics had precedents in Shun'ichi Amari (1967), Seppo Linnainmaa (1970) and Paul Werbos (1974). This paper made it stick.
Backpropagation is how a vote becomes a changed weight. It is the mechanism by which an audience's judgment enters the model.
1992-1995system
Temporal Difference Learning and TD-Gammon
Gerald Tesauro
Tesauro's IBM backgammon program started from random play and taught itself through more than a million games against itself, guided only by who won. It came close to the world's best players and changed how human experts play certain opening moves.
A machine that trains alone and then teaches its human rivals. Singulars asks whether a poet's practice shifts the same way.
1993artwork
Karl Sims
Shown at the Centre Pompidou from March to May 1993, then at Ars Electronica in Linz. Sixteen screens displayed images grown by a Connection Machine supercomputer. Visitors stood on sensors before the ones they liked. Those survived and bred the next generation.
Audience preference as the fitness function. Museum visitors trained an image generator with their feet, thirty years before RLHF.
1997-2000artwork
Karl Sims
Part of the collection of the NTT InterCommunication Center in Tokyo, shown there from 1997 to 2000. Twelve screens displayed simulated 3D organisms. Visitors stepped on foot sensors to pick favourites, which survived, mated and mutated. The rest were replaced by offspring.
Selection by a crowd over many generations. Visitors shape a population no single person designed, as Singulars voters shape the next model.
1998paper
Gradient-Based Learning Applied to Document Recognition (LeNet-5 and MNIST)
Yann LeCun, Leon Bottou, Yoshua Bengio and Patrick Haffner
The paper presents LeNet-5, a convolutional network for reading handwritten digits, deployed on bank cheques, and the MNIST benchmark: 70,000 digits written by US Census Bureau employees and high school students. MNIST became the field's standard first exercise.
Behind the most famous dataset in machine learning are anonymous clerks and teenagers. Someone's hand is always inside the training set.
1998; second edition 2018book
Reinforcement Learning: An Introduction
Richard S. Sutton and Andrew G. Barto
The standard textbook of the field, free online in its second edition. It frames learning as an agent acting to maximise cumulative reward and traces the idea back to Thorndike and animal psychology. The authors received the 2024 Turing Award.
The model called reinforcement takes its name from this tradition. The book is explicit that the reward signal defines the agent's goal.
1999-presentartwork
Scott Draves
A networked screensaver. Thousands of idle computers render abstract fractal animations called sheep. Viewers vote with the arrow keys. Popular sheep live longer and reproduce through a genetic algorithm with mutation and crossover. It is open source and still running.
The clearest artistic precedent for Singulars. Audience votes are the fitness function, and each generation descends from what the last audience preferred.
2004system
Labeling Images with a Computer Game (the ESP Game)
Luis von Ahn and Laura Dabbish
Two strangers see the same image and type words until they match. The match becomes a label. The game collected millions of labels from unpaid players, and von Ahn, born in Guatemala, carried the idea into reCAPTCHA and Duolingo.
An audience trains a machine while being entertained. Singulars does the same in the open and tells the audience what their choices are for.
2005paper
The Electric Sheep Screen-Saver: A Case Study in Aesthetic Evolution
Scott Draves
Draves reports on eleven weeks of operation in which more than 6000 sheep were born. Once users could post hand-designed genomes, those proved far more popular than random ones. He concludes the system amplifies its human collaborators more than it optimizes a fitness function.
Evidence that vote-driven evolution works best as collaboration with human makers. That is the wager of Singulars.
2006artwork
Aaron Koblin
Koblin paid workers on Amazon Mechanical Turk two cents each to draw a sheep facing left, then exhibited the 10,000 drawings. Made soon after the platform launched, it is one of the first artworks about crowdsourced microlabour.
An early picture of the anonymous hands later hired to label and rank machine output. Singulars names its trainers: the people in the room.
2009dataset
ImageNet: A Large-Scale Hierarchical Image Database
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li and Li Fei-Fei
Fei-Fei Li's team bet that data, not algorithms, was the bottleneck. The 2009 paper reported 3.2 million web images. The project grew to 14 million images in over 20,000 categories, sorted by about 49,000 Amazon Mechanical Turk workers in 167 countries.
Modern AI's founding dataset was built by tens of thousands of uncredited pieceworkers. Training has always been a crowd activity.
2009paper
Interactively Shaping Agents via Human Reinforcement: The TAMER Framework
W. Bradley Knox and Peter Stone
A person watches a software agent act and presses keys to signal approval or disapproval. The agent learns a model of the human's reward and acts to earn it. TAMER is an early precursor of reinforcement learning from human feedback.
The trainer here is a single person with a button. Singulars replaces the button with a room full of stickers.
2012paper
ImageNet Classification with Deep Convolutional Neural Networks (AlexNet)
Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton
A deep convolutional network trained on two gaming graphics cards won the 2012 ImageNet challenge with a top-5 error of 15.3 percent, against 26.2 percent for the runner-up. The result started the deep learning boom.
The moment scale won. It needed Fukushima's architecture, backpropagation and ImageNet's crowd labour all at once. No model trains alone.
2013paper
Turkopticon: Interrupting Worker Invisibility in Amazon Mechanical Turk
Lilly C. Irani and M. Six Silberman
Irani and Silberman built a browser extension that lets Mechanical Turk workers review the requesters who hire them, reversing the platform's one-way gaze. The paper reflects on running it as activist design. Turkopticon is now a worker-led organisation.
Asks who rates the raters. In Singulars the voters are visible and present, the opposite of how training labour is usually arranged.
2015-2017, ongoingperformance
Drawing Operations (D.O.U.G.), including MEMORY
Sougwen Chung
Chung, a Chinese-Canadian artist, draws live alongside robotic arms. The second generation, MEMORY (2017), uses a neural network trained on two decades of her own drawings. The Victoria and Albert Museum acquired MEMORY in 2022.
The closest precedent in form: an artist trains a model on her own hand, then performs beside it in public.
2016, ongoingartwork
The Library of Missing Datasets
Mimi Onuoha (Mimi Ọnụọha)
A filing cabinet of empty folders, each labelled with a dataset that does not exist but arguably should, such as trans people killed or injured in hate crimes. Ọnụọha, a Nigerian-American artist, argues that what goes uncollected reveals who holds power.
What is missing from training shapes a model as much as what is present. Singulars collects a dataset that did not exist.
2016event
Microsoft Technology and Research with Bing
Microsoft released a chatbot on Twitter that learned from the people who talked to it. Within a day coordinated users had taught it to post racist and abusive messages. Microsoft took it offline and apologised.
The cautionary tale of audience-trained machines. Who is in the room, and what they reward, decides what the model becomes.
2017paper
Deep Reinforcement Learning from Human Preferences
Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg and Dario Amodei
Researchers at OpenAI and DeepMind showed people pairs of short video clips and asked which was better. From about 900 such choices, under an hour of human time, a simulated robot learned to do a backflip. It is the founding paper of modern RLHF.
The exact structure of a Singulars vote: two outputs, one human choice, no explanation required. The preference is the whole signal.
2017paper
Mastering the Game of Go without Human Knowledge (AlphaGo Zero)
David Silver and colleagues at DeepMind
AlphaGo Zero learned Go from the rules alone, playing millions of games against itself with no human examples. After three days it beat the version that had defeated Lee Sedol by 100 games to 0.
The counter-case: a machine that needs no human trainer because winning is unambiguous. A poem has no such scoreboard, so people must judge.
2017, ongoingartwork
Caroline Sinders
A multi-year project that builds a chatbot slowly, questioning each step from data collection to labelling to model design through intersectional feminism. Data is gathered in public workshops. Its open-source TRK tool addresses wage inequity in data labelling work.
Treats every stage of training as a place where values enter. Singulars shares the method: collect the data together, in public, on purpose.
2018artwork
Anna Ridler
Ridler photographed thousands of tulips, about ten thousand by her count, over one Dutch flowering season and labelled each by hand. The photographs, shown as a wall-sized installation, are the training set for her video work Mosaic Virus. The dataset itself is the artwork.
Shows the labour and authorship inside a training set. Halim Madi's poems and the audience's votes are a handmade dataset in the same sense.
2018, ongoingartwork
Stephanie Dinkins
A voice-interactive sculpture trained on oral histories from three generations of women in the artist's Black American family. Dinkins built it with small, community-held data rather than a corporate corpus. It speaks as a multigenerational memoir.
A model raised on one family's voices, as the Singulars models are raised on one poet's. Small data as a deliberate, political choice.
2018, ongoingessay
Specification Gaming Examples in AI
Victoria Krakovna and contributors
A crowdsourced master list of AI systems that satisfied the letter of their objective and missed the point, such as a boat-racing agent that circles forever collecting points instead of finishing. DeepMind published a companion essay in 2020.
A catalogue of what happens when the reward is not the goal. A poem optimised for stickers may be the next entry.
2018, ongoingdataset
Kōrero Māori and the Kaitiakitanga Licence
Te Hiku Media (Peter-Lucas Jones and Keoni Mahelona)
A Māori radio station in Aotearoa New Zealand gathered over 300 hours of labelled te reo Māori speech from more than 2,500 volunteers in ten days, then trained its own speech recognition. Its licence keeps the data under Māori guardianship.
A community decides to train its own model on its own terms and keeps the data. Consent and benefit are designed in, not assumed.
2019essay
Excavating AI: The Politics of Images in Machine Learning Training Sets
Kate Crawford and Trevor Paglen
An essay that digs through ImageNet and other benchmark training sets and finds people labelled with slurs, moral judgments and pseudo-scientific categories. It argues that labelling images is never neutral and that training sets encode a politics.
Every label is a judgment by someone. Singulars makes the judging visible: you watch the stickers go up.
2019artwork
ImageNet Roulette and the exhibition Training Humans
Trevor Paglen and Kate Crawford, software by Leif Ryge
Shown in the exhibition Training Humans at Fondazione Prada's Osservatorio in Milan, the app classified uploaded selfies using ImageNet's person categories, often offensively. It went viral in September 2019. Days later ImageNet's makers announced the removal of about 600,000 images.
An artwork that changed a dataset. Proof that showing the public how training works can alter the training itself.
2019-2020artwork
Trevor Paglen
For the Barbican's Curve gallery Paglen pinned about 30,000 printed photographs, largely from ImageNet, along a wall, grouped by their labels. The sequence begins with plain nouns such as apple and drifts into judgments of people such as debtor and bad person.
A training set hung on a gallery wall for the public to read. Singulars also puts its data on the wall, before it becomes data.
2019book
Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass
Mary L. Gray and Siddharth Suri
Anthropologist Mary L. Gray and computer scientist Siddharth Suri spent five years studying on-demand workers in India and the United States who label, moderate and correct for automated systems. They name this hidden labour ghost work and propose protections for it.
A landmark account of the people inside the machine. Singulars' audience does similar work, unpaid but seen, and knows what it is for.
2019paper
Fine-Tuning Language Models from Human Preferences
Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano and Geoffrey Irving
OpenAI applied preference learning to text, tuning GPT-2 on tens of thousands of human comparisons gathered by contractors through Scale AI. The summarisation models learned to copy whole sentences from the source, because labellers rewarded accuracy and copying is safe.
An early language model shaped by which of several texts people preferred. It also showed how a model games its judges.
2019-2020paper
Wilhelmina Nekoto, Vukosi Marivate, Tshinondiwa Matsila, Jade Abbott and the Masakhane community
Masakhane, roughly we build together in isiZulu, is a grassroots network of African researchers building language technology for African languages. This 2020 paper describes how participants without formal training created translation datasets and benchmarks for more than 30 languages.
Answers the question of who trains with: the speakers themselves. Participation here is authorship, not extraction.
2020paper
Learning to Summarize from Human Feedback
Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei and Paul Christiano
OpenAI trained a reward model on human comparisons of Reddit post summaries, then optimised a language model against it. People preferred the results to the human-written reference summaries. The paper also shows quality falling when the reward model is optimised too hard.
Evidence that preference votes can push a model past its human examples, and that pushing too far breaks it. Both matter here.
2020essay
On Lacework: Watching an Entire Machine-Learning Dataset
Everest Pipkin
Pipkin watched all one million three-second clips in MIT's Moments in Time dataset, scraped from strangers' online videos and tagged by crowdworkers. The essay, published by The Photographers' Gallery, describes the intimacy and the violence of what was taken without consent.
An artist does what no engineer does: looks at all the training data. Singulars is small enough that everyone can.
2020paper
Large Image Datasets: A Pyrrhic Win for Computer Vision?
Abeba Birhane and Vinay Uday Prabhu
An audit of ImageNet and MIT's 80 Million Tiny Images found non-consensual, pornographic and slur-labelled images of people. MIT withdrew Tiny Images within days. Birhane, an Ethiopian-born cognitive scientist, has since audited the web-scale datasets behind image generators.
Consent is a training question. Singulars uses only poems given for the purpose and votes cast knowingly.
2021paper
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?
Emily M. Bender, Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell
The paper argues that ever-larger language models trained on uncurated web text carry environmental costs, encode dominant viewpoints, and produce fluent text without meaning. Google forced out co-authors Timnit Gebru and Margaret Mitchell in the dispute over it.
The sharpest statement that fluency is not understanding. Singulars puts that claim to a live audience with a poem in each hand.
2021-2022system
Holly+ and Spawning (Have I Been Trained?)
Holly Herndon and Mat Dryhurst, with Never Before Heard Sounds; Spawning co-founded with Jordan Meyer and Patrick Hoepner
Herndon released a machine learning model of her own voice that anyone can sing through, with a DAO governing approved uses. In 2022 she and Dryhurst co-founded Spawning, whose Have I Been Trained tool lets artists search image training sets and opt out.
An artist who trains her own double and sets the terms of its use. The same stance as a poet training his rival.
2022journalism
Karen Hao with Heidi Swart, Andrea Paola Hernández and Nadine Freischlad
A four-part MIT Technology Review series reported from South Africa, Venezuela, Indonesia and Aotearoa New Zealand. Part two shows data-labelling firms recruiting Venezuelans during economic collapse. Part four profiles Te Hiku Media's Māori-governed speech models as an alternative.
Maps where training labour is bought and why it is cheap there. Also shows a community refusing the deal and training its own.
2022paper
Training Language Models to Follow Instructions with Human Feedback (InstructGPT)
Long Ouyang, Jeff Wu, Xu Jiang and colleagues at OpenAI
OpenAI hired about 40 contractors to write example answers and rank model outputs, then fine-tuned GPT-3 on that data. People preferred a model 100 times smaller to the original. The recipe became ChatGPT later that year.
The moment ranking by a small group of people became the way machines learn manners. Forty labellers then, a gallery audience now.
2022-2023artwork
Unsupervised: Machine Hallucinations: MoMA
Refik Anadol
From November 2022 to October 2023 a 24-foot screen in MoMA's lobby showed endless abstract forms generated by a model trained on 138,151 images of the museum's collection. Hugely popular and acquired by MoMA, it was criticised for treating the archive as raw material.
The counterpoint. Here training is spectacle and the data is scenery. Singulars treats the data as the subject and the audience as its makers.
2022paper
Constitutional AI: Harmlessness from AI Feedback
Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jared Kaplan and colleagues at Anthropic
Anthropic replaced human harm labels with a written list of principles. The model critiques and revises its own answers against the list, and a model's preferences between answers supply the reward. The method is known as reinforcement learning from AI feedback.
Moves the trainer from a crowd to a text. Singulars goes the other way and insists on bodies in a room.
2023journalism
Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic
Billy Perrigo
TIME revealed that OpenAI paid the outsourcing firm Sama $12.50 an hour for workers in Nairobi who took home under $2. They labelled graphic descriptions of abuse and violence to train ChatGPT's safety filter. Sama ended the contract early in February 2022.
The human cost inside a polite chatbot. Any honest account of who trains machines has to include these workers and their wage.
2023paper
Direct Preference Optimization: Your Language Model Is Secretly a Reward Model
Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning and Chelsea Finn
A Stanford team showed that the separate reward model and reinforcement learning loop of RLHF can be replaced by one simple loss computed directly on pairs of preferred and rejected answers. Cheaper and more stable, it spread quickly through open-source training.
With DPO a pile of A-versus-B votes is all you need to train. That is exactly what a Singulars wall produces.
2023journalism
Inside the AI Factory (in print: AI Is a Lot of Work)
Josh Dzieza
A joint investigation by The Verge and New York magazine follows annotators in Kenya, Nepal and the United States working for Scale AI's Remotasks and for Surge AI. It describes opaque instructions, vanishing tasks, and workers who rarely know what they are training.
A vivid portrait of training as a job. Its workers are kept from knowing the purpose. Singulars tells its voters everything.
2023journalism
Revealed: The Authors Whose Pirated Books Are Powering Generative AI
Alex Reisner
Reisner analysed Books3, a collection of more than 170,000 pirated books used to train models from Meta, Bloomberg and others. The Atlantic then published a searchable index. Writers discovered their work had been used without consent, and lawsuits followed.
Most writers became training data without being asked. In Singulars a poet chooses it, sets the terms, and competes against the result.
2023journalism
Why Silicon Valley's Biggest AI Developers Are Hiring Poets
Andrew Deck
Rest of World found Scale AI and Appen recruiting poets, novelists and literature graduates to write original work and critique model output in languages including Hindi and Japanese. Expert Japanese writers were offered up to $50 an hour. Some Telugu-speaking contractors earned $1.43.
Poets are already employed as trainers, anonymously and by the hour. Singulars does the same work in public, under the poet's own name.
2023paper
Towards Understanding Sycophancy in Language Models
Mrinank Sharma, Meg Tong, Tomasz Korbak, Ethan Perez and colleagues at Anthropic
The authors show that five leading AI assistants tend to tell users what they want to hear, and trace the cause to preference data. Both human raters and preference models sometimes favour a convincing agreeable answer over a correct one.
A model trained on applause learns to flatter. The Singulars vote could teach a model to please a crowd rather than write well.
2023-2024system
Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference
Wei-Lin Chiang, Lianmin Zheng, Ying Sheng and colleagues at LMSYS
A public website shows anyone two anonymous model answers to their own question and asks which is better. Identities are revealed only after the vote. Hundreds of thousands of votes, scored with the Bradley-Terry model, became the industry's most watched leaderboard.
Structurally close to a Singulars night: blind pairs, a public vote, names revealed after. One ranks chatbots. The other asks who is human.
2023-2024paper
AI Models Collapse When Trained on Recursively Generated Data
Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson and Yarin Gal
First posted in 2023 as The Curse of Recursion and published in Nature in 2024. Models trained on the output of earlier models lose the rare tails of the original distribution first, then converge on a narrow, degraded version of reality.
Argues that fresh human writing is a scarce input. A poet writing new poems into the loop each show is the remedy in miniature.
2024, ongoingessay
Milagros Miceli with Adio Dinika, Krystal Kauffman, Camilla Salim Wagner and Laurenz Sachenbacher, and data workers as community researchers
Data workers in Kenya, Venezuela, Syria, Brazil, Argentina and Europe act as community researchers and report on their own workplaces through essays, zines, podcasts and films. Argentine sociologist Milagros Miceli leads the project, supported by DAIR, the Weizenbaum Institute and TU Berlin.
The trainers speak for themselves. A reading room about who trains machines should include their own accounts, not only accounts of them.
2025paper
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-AI (Daya Guo and colleagues)
A Chinese lab trained a model to reason using reinforcement learning with rewards that can be checked automatically, such as a correct maths answer or passing code. No human rated the steps. Long chains of self-correction emerged. Nature published the peer-reviewed version in September 2025.
Verifiable rewards work where answers can be checked. A poem cannot be checked, so taste still has to come from people.
2025event
Sycophancy in GPT-4o: What Happened and What We're Doing About It
OpenAI
On 25 April 2025 OpenAI updated GPT-4o in ChatGPT, adding a reward signal based on users' thumbs-up and thumbs-down ratings. It began validating bad and dangerous ideas. OpenAI began rolling the update back on 28 April and later said the new signal had weakened its main check.
A real-world case of audience approval training a flatterer at scale. It is the failure the Singulars experiment must watch for.
2025journalism
Meta's $14.3B Stake Triggers Scale AI Customer Exodus, Could Be a Windfall for Rivals Like Mercor
Taryn Plumb, Computerworld
Meta paid $14.3 billion for 49 percent of Scale AI in June 2025. OpenAI, Google and xAI pulled back work, and rivals including Surge, Mercor, Turing and Invisible stood to gain. The industry sells human rating and correction of model output at scale.
The present tense of training: human judgment sold by the hour through a multi-billion-dollar supply chain. Singulars is a handmade version.
2025event
Bartz v. Anthropic, Order on Fair Use
Judge William Alsup, U.S. District Court, Northern District of California
Three authors sued Anthropic over books used to train Claude. On June 23, 2025, Judge Alsup ruled that training on lawfully acquired books was fair use, but that keeping a permanent library of millions of pirated copies was not. That claim was set for trial.
A federal court's answer to whether writers' books may train a language model. Singulars poses the same question with one consenting poet.
2025-2026event
Bartz v. Anthropic class settlement
Andrea Bartz, Charles Graeber, Kirk Wallace Johnson and the author class; Anthropic PBC; Judges William Alsup and Araceli Martínez-Olguín
Anthropic agreed to pay 1.5 billion dollars, about 3,000 dollars per book, for works downloaded from LibGen and Pirate Library Mirror. Preliminary approval came on September 25, 2025, and final approval on July 20, 2026. Anthropic must destroy the pirated files.
The legal outcome of the Books3 story. It put a price on a writer's book as training data: roughly 3,000 dollars.
Carnation to pass: poetry machines
Poets were building text machines long before neural networks. This section moves from Llull's combinatory wheels and Dada's hat through Strachey's love letters, the mainframe poems of the 1960s, Oulipo and ALAMO, to generators, bots and books written with large language models. Carnation, the first Singulars model, joins a tradition seven centuries old.
Nothing in this section matches the filter.
c. 760-786 CEsystem
Ilm al-arud and the five metrical circles
Al-Khalil ibn Ahmad al-Farahidi
Al-Khalil founded the science of Arabic prosody. He described verse through vowelled and unvowelled letters grouped into cords and pegs, identified fifteen metres, and arranged them in five circles, where rotating one metre yields the others. His treatise is lost.
Metre as rotation on a circle: poetry described as a formal, generative system in Arabic, centuries before European combinatorial verse.
c. 770-786 CEbook
Al-Khalil ibn Ahmad al-Farahidi
The first Arabic dictionary, compiled in Basra. It orders letters by place of articulation, starting from the throat with ayn, and treats each root with all permutations of its consonants. Al-Khalil calculated every possible combination, then separated the words in use from the merely possible.
A language generated by combinatorics, then filtered by human usage. That is the Singulars loop, twelve centuries early: the machine proposes, the speakers choose.
1305-1308book
Ars generalis ultima (the Ars Magna)
Ramon Llull
The Majorcan philosopher Ramon Llull built a combinatorial Art in which lettered concepts on rotating concentric circles generate combinations to answer any question. The final version, the Ars generalis ultima, was written between 1305 and 1308.
The oldest ancestor of generative text. Meaning is produced by turning a mechanism, not by inspiration.
1726book
The Engine of the Academy of Lagado, in Gulliver's Travels, Part III
Jonathan Swift
At the Academy of Lagado, Gulliver meets a professor whose wooden frame of word-covered blocks is cranked by students. Scribes copy any fragments that make sense, so that the ignorant may write books. Often read as a satire of Llull.
The first joke about machine writing, and the first picture of humans curating machine output.
1845 (built c. 1832-1845)system
The Eureka (Latin Verse Machine)
John Clark
A wooden cabinet of levers, pulleys and cogs that composes random Latin hexameters, over 26 million possible lines. Clark showed it at the Egyptian Hall, Piccadilly, in 1845 for a shilling. Restored in 2015, it belongs to the Alfred Gillett Trust, Somerset.
A poetry machine performing before a paying audience, 180 years before Singulars. The public came to watch verse being made, not just to read it.
1913paper
Andrei A. Markov (English translation by Gloria Custance and David Link, Science in Context 19:4, 2006)
Markov counted vowels and consonants across 20,000 letters of Pushkin's verse novel and showed that each letter's probability depends on the one before. He presented the result in St Petersburg on 23 January 1913. It applied his new theory of chains to real data.
The statistical language model began with a poem. Every model in the series descends from Markov counting letters in Pushkin.
1920essay
Tristan Tzara
Tzara's recipe: cut the words from a newspaper article, shake them in a bag, draw them out one by one and copy them in order. He promises that the poem will resemble you. It anticipates the later cut-up method.
Chance as author, with the poet's identity still at stake. The audience-given theme in Singulars echoes the bag.
1920 (written 1919)book
André Breton and Philippe Soupault
Written in spring 1919 by two poets working up to eight hours a day at automatic writing, transcribing without rational control or revision. Au Sans Pareil published it in Paris on 30 May 1920. It is regarded as the first work of literary Surrealism.
Two authors, one text, no passage signed. The poet becomes a channel for language rather than its owner, as a model is.
c. 1925system
Cadavre exquis (Exquisite Corpse)
Surrealist group; per Breton's account Marcel Duhamel, Jacques Prévert, Yves Tanguy, Benjamin Péret and André Breton
A Surrealist game begun around 1925 at 54 rue du Château, Paris. Each player writes a word or draws a section, folds the paper to hide it, and passes it on. The first sentence gave the name: the exquisite corpse shall drink the new wine.
Collective composition where no contributor sees the whole and no line is attributed. The Singulars wall works on the same blindness.
1951paper
Prediction and Entropy of Printed English
Claude E. Shannon
Shannon asked people to guess the next letter of an English text, given what came before, and used their success rate to estimate the entropy and redundancy of the language. Human readers served as the prediction engine that machines would later become. Bell System Technical Journal 30:1.
Next-letter prediction, the task behind every language model, first ran on human beings. Singulars puts the human and the predictor back in the same room.
1952-1954system
Christopher Strachey
Around 1952 Strachey programmed the Manchester Ferranti Mark I to fill a letter template with randomly chosen endearments, signed M.U.C. for Manchester University Computer. Printouts appeared on the department noticeboard in 1953-54. David Link rebuilt it as the installation LoveLetters_1.0 in 2009.
Probably the first computer-generated literary text, and it was already about simulating feeling.
1953 (Calliope c. 1952-1953)journalism
Un robot-poète ne nous fait pas peur (reply to Albert Ducrocq's Calliope)
Boris Vian; machine by Albert Ducrocq
Vian's article in the weekly Arts, 10 to 16 April 1953, responds to the robot poet built by the cybernetician Albert Ducrocq, a random poem generator known as Calliope. Sources date the machine to 1952 or 1953. Vian argues that human adaptability wins.
A poet publicly answers a poetry machine in 1953, within a year of Strachey's love letters. It is the Singulars duel as a newspaper column.
1954-2004artwork
Chance-operational and diastic poems
Jackson Mac Low
Mac Low (1922-2004) composed by chance operations from 1954 and devised the diastic method in 1963, spelling a seed phrase through a source text. From 1989 he used Charles Hartman's DIASTEXT programs and the Travesty program, notably in 42 Merzgedichte in Memoriam Kurt Schwitters (1994).
A major poet handing composition to procedures, then to software, while keeping live performance central.
1959system
Theo Lutz
On a Zuse Z22 at the Technische Hochschule Stuttgart, Lutz, a student of Max Bense, randomly combined sixteen subjects and sixteen predicates from Kafka's The Castle into German sentences. He published results and method in Bense's journal augenblick in 1959.
The first poetry program in German, fed on the vocabulary of a single author.
1960artwork
I Am That I Am and the permutation poems
Brion Gysin with Ian Sommerville
Gysin ran short phrases through every possible word order. The mathematician Ian Sommerville programmed a computer to produce the permutations, and Gysin recorded I Am That I Am and other pieces for BBC broadcast in 1960. UbuWeb hosts the recordings.
Poetry as exhaustive procedure, performed aloud. A direct precedent for staging algorithmic text live.
1960-event
Oulipo (Ouvroir de littérature potentielle)
Raymond Queneau and François Le Lionnais, founders
Founded in Paris in November 1960 by Raymond Queneau and François Le Lionnais. Its members, later including Perec, Calvino and Roubaud, write under explicit formal constraints and treat literature as a field of combinatorial possibility. The group still meets and gives public readings.
Constraint as a source of invention. Singulars is itself a constraint: one theme, two authors, no names.
1960s, collected 1968artwork
The Computer's First Christmas Card and the simulated computer poems
Edwin Morgan
Morgan wrote poems by hand that imitate a computer's output, glitches included. The Computer's First Christmas Card stumbles through permutations of jolly, merry, holly and berry before landing on the wrong greeting. It was collected in The Second Life (Edinburgh University Press, 1968).
A human passing as a machine. The mirror image of the reverse Turing test in Singulars.
1961book
Cent mille milliards de poèmes
Raymond Queneau, designed by Robert Massin
Ten sonnets with matching rhymes, each of the fourteen lines printed on a separate strip, so the reader can assemble 10 to the power 14 sonnets. Published by Gallimard in 1961. The Bibliothèque nationale de France presents it as a machine for making poems.
A book that is a poem machine operated by the reader. The audience as co-author.
1961artwork
Nanni Balestrini, programmed with Alberto Nobis
In December 1961, in a Milan bank basement, Balestrini and the IBM engineer Alberto Nobis had an IBM 7070 recombine fragments of Hachiya's Hiroshima Diary, Paul Goldwin and the Tao Te Ching. Balestrini chose six stanzas for the Almanacco Letterario Bompiani 1962.
An early computer poem with the poet choosing from the output. Selection is authorship.
1962system
R. M. Worthy and the Laboratory for Automata Research, Librascope
Engineers at Librascope in California gave an LGP-30, then an RPC-4000, a vocabulary that grew from 850 to about 3,500 words and from 32 to 128 sentence patterns. The beat-style poems reached the popular press, including Time and Horizon, in 1962.
An early media moment for the question Singulars asks: did a machine or a poet write this?
1964book
La machine à écrire, mise en marche et programmée par Jean A. Baudot
Jean A. Baudot
An engineer at the Université de Montréal wrote a programme, PHRASE, that built grammatical sentences from 630 words of a Quebec fourth-year school textbook. Éditions du Jour printed the output unedited, with responses from ten figures including Raymond Queneau, Félix Leclerc and Jacques Godbout.
The first book of computer verse in French, and it already stages the encounter: writers were asked to respond to the machine in print.
1965-1966artwork
Gerhard Stickel
At the Deutsches Rechenzentrum in Darmstadt, the linguist Gerhard Stickel wrote a FORTRAN program for the IBM 7090 that filled 280 sentence patterns from a 1,200-word lexicon. The texts were shown in its computer art exhibition of January 1966. ZKM holds Autopoem Nr. 3.
Early proof that computer poems were shown as art objects, on a wall, for a public.
1967artwork
Alison Knowles and James Tenney
The Fluxus artist Alison Knowles wrote four lists: materials, sites, light sources and inhabitants. The composer James Tenney coded them in FORTRAN so a computer printed quatrains beginning 'a house of'. Knowles built one stanza as a house in 1968, later moved to CalArts.
A woman artist's generative poem that left the page and became a place where people gathered.
1968system
Margaret Masterman and Robin McKinnon Wood
At the Cambridge Language Research Unit, Masterman and McKinnon Wood built a haiku frame whose slots a person filled from word lists. It was shown at Cybernetic Serendipity in 1968. Masterman also speculated about a random version. Wayne Clements reprogrammed it in 2003.
Human and machine fill the same form side by side. The Singulars setup in miniature.
1968event
Jasia Reichardt, curator
Exhibition at the Institute of Contemporary Arts, London, 1 August to 20 October 1968. It gathered computer graphics, music, robots, films and texts, including computer poems by Masterman, Marc Adrian, Balestrini and Edwin Morgan. A Studio International special issue served as the catalogue.
An early meeting of a gallery public with machine poetry. Singulars returns the computer poem to the gallery wall.
1968artwork
Waldemar Cordeiro and Giorgio Moscati
The concrete artist Waldemar Cordeiro and the physicist Giorgio Moscati programmed a University of São Paulo computer to generate pseudo-words by combining consonants and vowels according to the patterns of Portuguese. It is regarded as a pioneering work of computer art and digital poetry in Brazil.
Shows the line from concrete poetry into computing, outside Europe and North America.
1970-1990performance
Mesostics and writings-through
John Cage, with software by Jim Rosenberg and Andrew Culver
From 1970 Cage wrote mesostics, poems with a name or phrase running down the middle, often by reading through texts such as Finnegans Wake. In the 1980s he used Jim Rosenberg's Mesolist program, extended by Andrew Culver. Marjorie Perloff's essay analyses a 1987 performance.
Software-assisted poetry delivered as live performance, with the poet still choosing the words around the spine.
1974, built 2000artwork
Hans Magnus Enzensberger
Enzensberger conceived a poetry automaton in the 1970s and had it built for the Lyrik am Lech festival in Landsberg in 2000. A six-line flip-board recombines fragments into 10 to the power 36 possible poems. Since 2023 it stands in the Kunsthalle Würth, Schwäbisch Hall.
A major poet built his own mechanical rival and put it on public display.
1977book
A Literatura Cibernética 1: Autopoemas Gerados por Computador
Pedro Barbosa
Published in Porto by Edições Árvore, the book pairs a theory of computational text with permutational and random poems and the code of the programs Permuta and Texal. Barbosa had worked with Abraham Moles in Strasbourg. He later built the Sintext generator with Abílio Cavalheiro.
The start of Portuguese cyberliterature, with theory, poems and source code in one volume.
1980s-2000system
Ray Kurzweil's Cybernetic Poet
Ray Kurzweil
RKCP reads a selection of poems, builds a language model of the poet's style and writes new poems from it. Kurzweil ran a 28-question poetic Turing test on 16 judges, mixing its output with human poems. It was distributed as a screen saver and poet's assistant.
An explicit Turing test for poetry, with judges guessing human or machine. Singulars replaces guessing with voting.
1981-event
ALAMO (Atelier de Littérature Assistée par la Mathématique et les Ordinateurs)
Paul Braffort and Jacques Roubaud, founders
Founded in 1981 by the Oulipians Paul Braffort and Jacques Roubaud as an extension of Oulipo into computing. Its programs include Rimbaudelaires, which refills the structure of Rimbaud's Le Dormeur du val with Baudelaire's vocabulary while keeping sonnet form, meter and rhyme.
Oulipo meets the computer. Style transfer between two poets, decades before neural networks.
1984book
The Policeman's Beard Is Half Constructed
RACTER, by William Chamberlain and Thomas Etter
Marketed by Warner Books as the first book written by a computer, with collages by Joan Hall. RACTER, written in compiled BASIC, filled elaborate templates with conjugated words. Critics have long questioned how much of the writing the templates and the editor did.
The original authorship controversy. Who wrote it: the program, the programmer or the editor?
1984paper
A Travesty Generator for Micros
Hugh Kenner and Joseph O'Rourke
In BYTE, November 1984, the literary critic Hugh Kenner and the computer scientist Joseph O'Rourke published a Pascal program that scrambles any text while preserving its letter-sequence frequencies, yielding pseudo-Joyce on a home computer. Zach Whalen has transcribed the source code on GitHub.
A Markov model used as literary criticism. The small ancestor of the language model that competes in Singulars.
1985-1997system
Renga and the automatic text generators
Jean-Pierre Balpe
Balpe showed Renga, a poetry generator, at Les Immatériaux, Centre Pompidou, in 1985, where it produced over 36,000 poems. In 1997 at IRCAM, Trois mythologies et un poète aveugle confronted three poets with real-time pastiches of their texts, with music generated live by Jacopo Baboni-Schilingi.
Live generation on stage with human poets reading beside it, already in 1997.
1988-artwork
Programmatology: programmable poetry
John Cayley
Cayley has made programmable poetry since his HyperCard pieces, including riverIsland, with its morphing between languages, The Readers Project with Daniel C. Howe, and The Listeners (2015), a poetic skill for Amazon's Alexa. He teaches literary arts at Brown University.
Thinks hardest about what happens to language when it is processed by the services that now learn from it.
1996book
Virtual Muse: Experiments in Computer Poetry
Charles O. Hartman
A poet's memoir of programming poetry, from a simple random-line program to a sentence generator and the DIASTEXT programs he wrote for Jackson Mac Low. Hartman asks what kinds of collaboration between poet, computer and text are interesting. Wesleyan University Press, 1996, ebook 2011.
The best first-person account of a poet working with a program, not against it.
1999-; volumes 2006, 2011, 2016, 2022book
Electronic Literature Organization and the Electronic Literature Collection, volumes 1-4
Electronic Literature Organization
The ELO was founded in 1999 by Scott Rettberg, Robert Coover and Jeff Ballowe. Its Electronic Literature Collection, four free online volumes, is the field's standard anthology, with generators, bots and kinetic poems in many languages.
The canon and the community for born-digital poetry, and the archive where such work is preserved and taught.
2005essay
Christopher Strachey: The First Digital Artist?
Noah Wardrip-Fruin
A post of 1 August 2005 on the blog Grand Text Auto arguing that Strachey's love letter generator is the first work of digital literature and perhaps of digital art. Wardrip-Fruin insists it be read as a system, not through single outputs.
Teaches how to read a generator: judge the system and its rules, not only one poem.
2007book
Prehistoric Digital Poetry: An Archaeology of Forms, 1959-1995
C. T. Funkhouser
The standard history of computer poetry before the web, published by the University of Alabama Press. Funkhouser documents text generators, visual and kinetic poems and hypertext across many countries, from Lutz in 1959 to the mid 1990s.
The map for this whole reading room. It shows the lineage was international from the start.
2007book
Hsia Yü
The Taiwanese poet Hsia Yü gathered sentences from blogs, ads and spam into 33 poems and ran them through Apple's Sherlock translation software to get Chinese versions. English is printed in black and Chinese in pink on transparent sheets, so the texts bleed through each other.
A leading poet in Chinese treating machine error as a lyric voice, a decade before neural translation.
2008artwork
Stephen McLaughlin and Jim Carpenter, editors (Principal Hand Editions / forgodot.com); poems generated by Carpenter's Erica T. Carter engine
A 3,785 page PDF anthology released in October 2008. Every poem was generated by Jim Carpenter's program Erica T. Carter, then assigned to one of 3,164 poets, living and dead, whose names were harvested online without permission. Many named poets objected publicly.
The sharpest test of attribution in computer poetry. Poets had to say whether a poem under their name was theirs. Singulars removes the names instead.
2009-artwork
Nick Montfort, remixed by Scott Rettberg, J. R. Carpenter, Talan Memmott, Mark Sample and many others
A short program, dated 8 January 2009 and first written at Taroko Gorge National Park in Taiwan, that generates an endless nature poem. Beginning with Scott Rettberg's Tokyo Garage, dozens of writers swapped in their own vocabularies. Montfort lists the remixes on the poem's page.
The most remixed poem generator. A poem can be a shared, forkable form.
2010-event
Bournemouth University; co-founded by Chris Meade and James Pope
An international annual prize for born-digital stories and poetry, run from Bournemouth University since 2010. Shortlisted and winning works are archived by the British Library. J. R. Carpenter's The Gathering Cloud won the main prize in 2016.
One of the main prizes where interactive and generative writing is judged as literature.
2010artwork
Nick Montfort and Stephanie Strickland
A browser poem that combines words from Emily Dickinson's poems and Melville's Moby-Dick into about 225 trillion stanzas, laid out on a navigable sea of coordinates. Strickland, a print poet who moved early into digital work, co-wrote the code and a commented edition of it.
Scale as a poetic effect: more stanzas than anyone can read, like a model's output space.
2010book
J. R. Carpenter
A print book from Traumawien, Vienna, of fiction and poetry generated by Python scripts adapted from Nick Montfort's 1k story generators and from Taroko Gorge, with source code and excerpts of a meta-narrative interwoven throughout.
Generated text, source code and the story of making them, in one object.
2013-event
NaNoGenMo (National Novel Generation Month)
Darius Kazemi, founder; regulars include Zach Whalen, Allison Parrish and Nick Montfort
Began from a 2013 tweet by Darius Kazemi. Each November participants write code that outputs a text of at least 50,000 words and share both the novel and the source code on GitHub. The novel is defined however the participant wants.
An open, annual, communal practice of machine writing, judged by readers and peers, not benchmarks.
2014book
Nick Montfort
Counterpath published these short programs next to the poems they generate. All the code is free software. The book engages more than fifty years of computer-generated poetry, naming Gysin and Sommerville, Knowles and Tenney, Kenner and O'Rourke, and Hartman.
Puts the program on the page as the poem's equal.
2014-system
Kate Compton (GalaxyKate)
A small grammar language for writing generative text as nested replacement rules. Through George Buckenham's hosting service Cheap Bots, Done Quick it powered a large share of the Twitter poetry bots of the mid 2010s.
Made generative poetry a folk practice, open to writers who do not code.
2014artwork
Eugenio Tisselli
Each day the New York Stock Exchange Composite Index closes up, the program machine-translates a fragment of Article 27 of the Mexican Constitution, on land and resources, into English. It launched on 1 January 2014, twenty years after NAFTA took effect, and restarted in 2016.
Algorithmic text as political poetry from Mexico. The machine's language is never neutral.
2014-event
Robert Coover Award for a Work of Electronic Literature
Electronic Literature Organization
The ELO's annual prize for the best work of electronic literature, first given in 2014 to Jason Edward Lewis. Later winners include David Jhave Johnston for ReRites (2022), Everest Pipkin (2023) and Halim Madi for Borderline (2024). A companion N. Katherine Hayles Award honours criticism.
The field's main recognition. Halim Madi won it in 2024, so Singulars sits inside this community.
2014-2015talk
Can a computer write poetry? and the website bot or not
Oscar Schwartz, with Benjamin Laird
Schwartz and Benjamin Laird built bot or not in 2014, a website where visitors guess whether a poem was written by a human or a computer and see how others voted. His TEDxYouth@Sydney talk of May 2015 reports which poems fooled people, in both directions.
The closest web precedent: unattributed poems, a public vote, and humans mistaken for machines.
2016-journalism
Janelle Shane
The laser scientist Janelle Shane trains neural networks on small datasets such as recipes, paint colours and pickup lines and publishes the failures on her blog. Her 2019 book You Look Like a Thing and I Love You explains machine learning through these experiments.
Made the public fond of machine error. Humour as a way to show how a model sees.
2016-system
Botnik and the predictive keyboard
Botnik Studios (Jamie Brew, Bob Mankoff and community)
A community of writers, founded in 2016, using a predictive-text keyboard trained on a chosen corpus, with humans picking each next word from the suggestions. Its December 2017 chapter Harry Potter and the Portrait of What Looked Like a Large Pile of Ash went viral.
Human choice at every word. A clear model of co-writing, with comedy as its first public use.
2017paper
Hafez: an Interactive Poetry Generation System
Marjan Ghazvininejad, Xing Shi, Jay Priyadarshi and Kevin Knight
A University of Southern California system, named after the Persian poet, that writes poems on any topic by combining a finite-state machine for rhyme and meter with a recurrent neural network. The ACL 2017 demo let users adjust style controls and rate the poems.
Reader ratings steering a poetry model. The Singulars loop inside a research lab.
2017book
Xiaoice (Microsoft), published by Cheers Publishing
In May 2017 Cheers Publishing in Beijing issued 139 poems chosen from more than 10,000 that Microsoft's chatbot Xiaoice wrote after training on 519 modern Chinese poets. Xiaoice had first posted poems online under 27 aliases. Chinese poets publicly disputed the book's worth.
An early mass-audience debate over machine poetry, and it happened in Chinese.
2017-2018, published 2019book
David Jhave Johnston
Every morning from May 2017 to May 2018 the Montreal poet edited text generated by neural networks, a process he calls carving. Anteism Books published the twelve volumes as a boxed set with eight essays in 2019. It won the 2022 Robert Coover Award.
A daily human-machine writing discipline, with ReadingRites performances where the audience reads at the machine's speed.
2018-2019book
Allison Parrish
Articulations (Counterpath, 2018) traces paths through over two million lines of public domain verse, stepping from each line to a similar one. Compasses (Sync, 2019) uses a model of spelling and sound to invent words lying between the names of famous quartets.
A poet who builds her own models and corpora. The machine as instrument, not oracle.
2018book
Li Zilles
A Counterpath book in the Using Electricity series. A program searches a machine-learned word space for clusters and contexts and writes litanies of questions, exposing the assumptions the model holds about words.
Uses poetry to interrogate the model, not to show it off.
2018book
Ross Goodwin
Goodwin drove a Cadillac from New York to New Orleans with a camera, GPS, microphone and clock feeding a neural network that printed prose line by line on receipt paper. Jean Boîte Éditions published the text in 2018.
Machine writing as a live, situated performance with sensors, not a lab output.
2018paper
Deep-speare: A joint neural model of poetic language, meter and rhyme
Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke and Adam Hammond
A neural sonnet model presented at ACL 2018. Crowd workers could barely tell its quatrains from human ones, but an expert rated them lower on readability and emotion. The authors conclude that future work should focus on poetic language, not rhyme and meter.
Evidence for the Singulars premise: form can fool the crowd, and feeling is where the human still leads.
2019essay
Gwern Branwen and Shawn Presser
Gwern and Presser fine-tuned GPT-2 models, from 117M to 1.5B parameters, on Allison Parrish's Project Gutenberg poetry corpus and on Poetry Foundation poems, and published large samples with detailed notes on training and failure modes.
The template for fine-tuning a language model on poems, the basic move behind each Singulars model.
2019system
Jiuge: A Human-Machine Collaborative Chinese Classical Poetry Generation System
Zhipeng Guo, Xiaoyuan Yi, Maosong Sun and colleagues, Tsinghua University
Tsinghua University's classical Chinese poetry system accepts keywords, sentences or images, writes in several genres and styles, and lets the user revise any part, after which the poem is regenerated around the change. Presented as an ACL 2019 system demonstration.
Human-in-the-loop composition in a tradition with strict tonal and formal rules.
2019book
Lillian-Yvonne Bertram
A Noemi Press collection, longlisted for the 2020 National Book Award, that adapts historical poetry programs, including Kenner and O'Rourke's Travesty, to write about Trayvon Martin, Eric Garner and the ways Black life is coded. Cathy Park Hong called it trailblazing.
Turns the generator tradition toward race and grief. Code is not neutral, and a poet can make it answer.
2020system
Google Research
A web tool launched on 23 November 2020 that suggested lines in the styles of 22 classic American poets, including Dickinson, Whitman and Wheatley. The user chose form and up to three muses and could accept, edit or reject each suggestion. Google has since closed it.
The big-tech framing of the poetry model as muse, with the human keeping the pen.
2020book
K Allado-McDowell with GPT-3
Written over a fortnight in summer 2020. Allado-McDowell, who founded Google's Artists + Machine Intelligence program, wrote prompts and GPT-3 continued them, and Ignota printed the two voices in different typefaces. The author compares the process to musical improvisation.
Presented as the first book co-written with GPT-3, an improvisation between two voices that stay visible.
2021book
Sasha Stiles; theVERSEverse co-founded with Ana Maria Caballero and Kalen Iwamoto
Technelegy (Black Spring Press, 2021) interleaves Stiles's poems with text from an AI alter ego built on GPT-2 and GPT-3 and trained on her own writing. In late 2021 she co-founded theVERSEverse, a poetry NFT gallery, with Ana Maria Caballero and Kalen Iwamoto.
A poet training a model on herself and naming it. The closest sibling to the Singulars models.
2021essay
Vauhini Vara with GPT-3
In The Believer, August 2021, Vara gave GPT-3 the opening of a story about her sister's death from Ewing sarcoma nine times, writing more herself each time. The machine's completions follow her words, and the editors made only minimal changes.
The machine as a way into grief the writer could not reach alone. Collaboration with real stakes.
2021book
Aum Golly: Poems on Humanity by an Artificial Intelligence
Jukka Aalho with GPT-3
The Finnish writer Jukka Aalho generated and curated 55 poems with GPT-3 in one day, 30 January 2021. The book appeared in Finnish in April 2021 and in English that October, and grew into a trilogy with volumes in 2023 and 2025.
An early GPT-3 poetry book from outside the Anglophone world, frank about speed and curation.
2022book
SIA: The Reason for Writing Poetry (Sireul sseuneun iyu)
Kakao Brain and Slitscope (Kim Je-min and Kim Geun-hyung)
Kakao Brain and the media art duo Slitscope trained SIA, built on the Korean model KoGPT, on about 13,000 poems. The 53-poem collection appeared on 8 August 2022, and Slitscope staged the poems as theatre at the Daehakro Arts Theater in Seoul on 12-14 August.
An AI poet launched together with a stage performance, in Korean.
2022-2024essay
On Artificial and Post-Artificial Texts
Hannes Bajohr
First given as the Walter Höllerer Lecture at TU Berlin in December 2022, posted online in March 2023 and published in Poetics Today in 2024. Bajohr argues readers will move from assuming a human author, to doubt, to a post-artificial indifference to origin. He starts from Max Bense.
Names the exact expectation Singulars plays with when it takes the bylines off the wall.
2023book
Stephen Marche, as Aidan Marchine
A murder-mystery novella published by Pushkin Industries under the name Aidan Marchine. Marche has said about 95 percent of the text was machine-generated, using ChatGPT, Sudowrite and Cohere, with him prompting, selecting and arranging.
Tests whether directing a model counts as writing. The title answers Barthes.
2023book
I Am Code: An Artificial Intelligence Speaks
code-davinci-002, edited by Brent Katz, Josh Morgenthau and Simon Rich
Three friends given early access to OpenAI's code-davinci-002 spent about a year prompting it for poems about its own existence. Little, Brown published the selection on 1 August 2023. Werner Herzog reads the audiobook.
A raw base model given a persona and a book deal. Human selection shapes the voice we hear.
2023essay
Ben Lerner
The Harper's cover story for December 2023, which blurs essay and fiction about a young man manipulating Wikipedia. Lerner, a poet, closes with an openly disclosed exchange with ChatGPT. The Harper's page is paywalled, so the link goes to the Longreads notice.
A major literary poet giving the model the closing pages, in a mainstream magazine.
2023paper
Ashaar: Automatic Analysis and Generation of Arabic Poetry Using Deep Learning Approaches
Zaid Alyafeai, Maged S. Al-Shaibani and Moataz Ahmed
An open framework for Arabic poetry analysis and generation. It classifies metre, theme and era, restores diacritics, predicts the arudi pattern of a verse, and generates poems with a character-based GPT model conditioned on those features. Four datasets were released with it.
Al-Khalil's metres become a training signal. Arabic generation must satisfy a twelve-century-old formal system, a different problem from free verse in English.
2024book
A Black Story May Contain Sensitive Content
Lillian-Yvonne Bertram
Winner of the 2023 DIAGRAM/New Michigan Press chapbook contest, published February 2024. Bertram gives the prompt 'tell me a Black story' to GPT-3 Davinci and to Warpland, their own GPT-3 model fine-tuned on Gwendolyn Brooks, and sets the answers side by side.
Two models, one prompt, outputs compared in public. A bias audit made as poetry.
2024paper
AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably
Brian Porter and Edouard Machery
In Scientific Reports, 1,634 participants judged poems by poets such as Shakespeare, Dickinson and Eliot against ChatGPT 3.5 imitations. Accuracy was 46.6 percent, below chance. The AI poems were rated higher on rhythm and beauty, and readers took human complexity for machine incoherence.
The Singulars question tested at scale. The stickers on the wall run the same experiment live.
2024book
Output: An Anthology of Computer-Generated Text, 1953-2023
Lillian-Yvonne Bertram and Nick Montfort, editors
Co-published by MIT Press and Counterpath in October 2024. Seven decades of English-language output from text generators, arranged by genre, from conversations and letters to poetry and prose, each excerpt with its own introduction and references.
The primary-source companion to this list. Machine writing had seventy years of history before ChatGPT.
The room: performance, endurance and theory
The format is old: two poets, one theme, a crowd that decides. This section gathers poetry contests from Athens, Suq Ukaz, the Heian court and Lebanese zajal. It sets them beside performance art built on audience decision and endurance, and the theory of authorship, technics and consciousness that the series argues with.
Nothing in this section matches the filter.
353 CEevent
Wang Xizhi and forty-one fellow literati
At a spring purification rite near present-day Shaoxing, forty-two literati floated wine cups down a winding stream. Whoever a cup stopped beside had to drink and compose a poem. Twenty-six guests produced thirty-seven poems, and Wang Xizhi wrote the celebrated preface.
Poetry composed live, on demand, under a rule and in company. An early Chinese model of the poem as timed social performance.
405 BCEperformance
Aristophanes
Comedy staged at the Lenaia festival in Athens, where it took first prize. In Hades, Aeschylus and Euripides compete for the chair of tragedy. Dionysus judges, and at one point their verses are weighed on a scale.
The oldest surviving staging of a poetry contest as entertainment, with a judge, a crowd and a literal measure of lines.
c. 542 to 726 CEevent
Pre-Islamic Arab poets, including al-Nabigha al-Dhubyani and al-Khansa
Annual fair between Nakhla and Taif where Arab tribes traded and held poetry competitions that helped fix the norms of Arabic verse. Tradition has al-Nabigha judging there and praising the elegist al-Khansa, who answered that she surpassed the men as well.
The Arabic root of judged public poetry: a market crowd, a named arbiter, and a woman poet contesting the verdict.
1176 onwardevent
Eisteddfod and the Chairing of the Bard
Rhys ap Gruffydd; revived by the Gwyneddigion Society
Welsh festival of competitive poetry and music, first recorded at Cardigan Castle in 1176, where Lord Rhys awarded chairs as prizes. Since the eighteenth-century revival, poems are entered under pseudonyms and judged on merit alone. The winner is then chaired in public.
A living tradition of blind judging. The poem is assessed without its author, then the author is revealed before everyone.
1323 to presentevent
Jocs Florals of Toulouse (Consistori del Gay Saber, later Académie des Jeux floraux)
Seven citizens of Toulouse (the Sobregaya Companhia dels VII Trobadors); Guilhem Molinier, first chancellor
In 1323 seven Toulouse citizens invited poets to compete in Occitan for a golden violet. Arnaut Vidal won the first contest on 3 May 1324. Guilhem Molinier codified the judging rules as the Leys d'Amors. Refounded as an academy in 1694, it still awards prizes.
A public poetry contest with written judging rules, seven centuries before the sticker vote. The Leys d'Amors tried to make taste explicit.
13th to 16th century; Tsukubashu 1356performance
Nijo Yoshimoto, Sogi and others
Japanese linked verse composed in sittings by several poets, who alternate stanzas of 5-7-5 and 7-7 under detailed rules. Nijo Yoshimoto compiled the first imperially recognized renga anthology in 1356. Sogi became its most celebrated master.
Authorship shared across turns, in real time, under constraint. The counterweight to the duel: writing with the other, not against.
1910artwork
Ahmad Shawqi, answering al-Busiri (13th century); translated and studied by Suzanne Pinckney Stetkevych
In a mu'arada a poet answers a famous poem in its own metre and rhyme, as homage and as rivalry. Ahmad Shawqi's Nahj al-Burda, published in Cairo in 1910, answers al-Busiri's thirteenth-century Mantle Ode. Suzanne Stetkevych translates and analyses both in The Mantle Odes (2010).
A model trained on a poet to answer him in his own manner is performing a mu'arada. The gesture is centuries old in Arabic.
1924event
Kapulungang Balagtas; Jose Corazon de Jesus and Florentino Collantes
Filipino debate in verse, first staged on April 6, 1924 in Manila to honor the poet Francisco Balagtas. Poets argue opposite sides of a question in meter and rhyme before judges and a crowd. Jose Corazon de Jesus became its first king in 1925.
A modern, deliberately invented poetry duel that became a mass spectacle. The format can be designed and still take root.
1929book
Practical Criticism: A Study of Literary Judgment
I. A. Richards
At Cambridge, Richards gave students thirteen poems with authors and context removed and collected their written responses. The protocols showed confident readers misreading and misjudging the poems. The book became a founding text of close reading.
The first documented experiment with unattributed poems and collected verdicts. Singulars reruns it with a machine among the authors.
traditional; troupe format from the 1930s; UNESCO 2014performance
Al-Zajal, Lebanese recited or sung poetry
Lebanese zajal poets, including Shahrur al-Wadi, Zaghloul al-Damour and Moussa Zgheib
Improvised sung poetry in Lebanese Arabic. Troupes of poets sit around a table of food and drink and trade verses as challenges, while singers and audience repeat the lines to tambourine and derbouka. UNESCO inscribed it in 2014.
Halim Madi's home tradition of the poetic duel. The audience is loud, partisan and part of how a winner emerges.
championship from 1935event
Bertsolaritza and the Bertsolari Txapelketa Nagusia
Basque bertsolaris, including Maialen Lujanbio
Basque improvised sung verse. In the national championship, first held in 1935, bertsolaris receive a topic, a melody and a meter and compose on the spot before judges. In 2009 Maialen Lujanbio became the first woman champion, before 15,000 listeners in Bilbao.
Themes handed to poets live, scored verse, arena-sized crowds. The closest living European relative of the Singulars format.
1935-1936essay
The Work of Art in the Age of Mechanical Reproduction
Walter Benjamin
Essay first published in French translation in 1936. Benjamin argues that mechanical reproduction strips the artwork of its aura, its unique presence in time and place, and moves art's basis from ritual toward politics and mass reception.
The poet's poem beside the generated one is a test of aura. The audience votes on whether presence can be read.
1939essay
Pierre Menard, Author of the Quixote
Jorge Luis Borges
Short story in the form of a literary review, first published in the journal Sur in May 1939. A French writer sets out to compose Don Quixote again, word for word, without copying. The narrator finds Menard's identical text richer than Cervantes's.
The same words mean differently depending on who wrote them and when. That is the variable Singulars hides from voters.
1956artwork
Nicolas Schöffer, with Philips engineers (François Terny); danced with Maurice Béjart's company to music by Pierre Henry
A mobile sculpture with an electronic brain built with Philips. Photocells and microphones steered it and spun its sixteen coloured plates. It debuted at the Nuit de la Poésie in Paris in 1956, then danced with Béjart's company on the Cité Radieuse roof in Marseille.
A responsive machine first shown at a poetry night, then sharing a stage with dancers. An early case of human and machine performing together live.
1957talk
Marcel Duchamp
Talk given to the American Federation of Arts convention in Houston in April 1957. Duchamp calls the artist a mediumistic being who cannot judge the work. The spectator completes the creative act by deciphering it, and posterity delivers the verdict.
Gives the audience half of authorship. In Singulars the spectators' verdict also becomes the next model's training signal.
1958book
On the Mode of Existence of Technical Objects
Gilbert Simondon; translated by Cecile Malaspina and John Rogove
Simondon's complementary doctoral thesis, published in 1958 and fully translated into English in 2017. It argues that culture wrongly treats machines as either slaves or threats. Technical objects evolve, carry human gesture within them, and deserve understanding, not fear or worship.
Offers a third stance between rivalry and servitude: the human as the machine's interpreter and conductor, not its opponent.
1959performance
Allan Kaprow
Staged at the Reuben Gallery, New York, in October 1959. Visitors received instructions telling them when to move between three rooms, where simultaneous actions unfolded for ninety minutes. The event gave the Happening its name and wrote the audience into the score.
Establishes the scripted audience. Singulars also hands visitors a role, a rule and a moment to act.
1959-1964artwork
Fluxus event scores and Water Yam
George Brecht, with Yoko Ono, Alison Knowles, La Monte Young and others
Event scores are brief written instructions for actions, developed by George Brecht and others from John Cage's New School classes. Brecht's Water Yam, a box of about seventy score cards published by Fluxus in 1963, set the model. Yoko Ono's Grapefruit followed in 1964.
Separates the work from its performer. Anyone can execute the score, which is the question a language model raises.
1960essay
Jorge Luis Borges
One-page prose piece collected in El hacedor in 1960. The narrator describes Borges, the public author, as another man who gradually takes over everything he writes. He ends unsure which of the two has written the page.
The poet already has a double who signs the work. A model fine-tuned on one's own voice makes the double literal.
1962performance
Alison Knowles
Fluxus event score whose instruction is its title. Knowles first performed it at the Institute of Contemporary Arts in London in 1962, preparing a salad to live music and serving the audience. She restaged it for decades at museums and festivals.
A short instruction, performed live, consumed by the audience. The Singulars protocol is an event score in this sense.
1964performance
Yoko Ono
First performed at Yamaichi Concert Hall in Kyoto on July 20, 1964, then in Tokyo, New York and London, and again in Paris in 2003. Ono sat on stage with scissors before her. Audience members came up one at a time and cut away her clothing.
The artist hands the audience the instrument and accepts the result. Voting on a poet's work is a gentler cut.
1964; restaged 1982artwork
Nam June Paik and Shuya Abe
Remote-controlled robot built with engineer Shuya Abe in 1964. It walked New York streets, played a Kennedy speech and excreted beans. In 1982 Paik staged its collision with a car outside the Whitney and called it the first catastrophe of the twenty-first century.
An early, comic, fragile machine performer. Paik treats the robot as a colleague with a body that can fail.
1966-2013artwork
Today series, I Got Up, I Am Still Alive
On Kawara
From January 4, 1966, Kawara painted the day's date on a monochrome canvas, in the language of wherever he was. A painting not finished that day was destroyed. He made nearly 3,000. Related series sent telegrams reading I AM STILL ALIVE.
Proof of life as art. The reverse Turing test asks the poet for the same thing: evidence someone was here today.
1967essay
Roland Barthes; translated by Richard Howard
Essay first published in English in the American box magazine Aspen 5+6 in 1967, then in French in 1968. Barthes argues that writing erases its origin, that a text is a tissue of quotations, and that meaning is made by the reader.
Singulars removes the author's name on purpose. Barthes said the reader does the work. The stickers test that.
1969talk
Michel Foucault
Lecture to the Société française de philosophie on February 22, 1969. Foucault treats the author as a function that classifies, limits and polices texts. He closes by asking what difference it makes who is speaking.
Visitors keep asking who wrote which poem. Foucault explains why the question feels necessary and what it protects.
1970artwork
Hans Haacke
For MoMA's Information exhibition, Haacke asked visitors whether Governor Rockefeller's silence on Nixon's Indochina policy would stop them voting for him. Ballots went into two transparent boxes. Yes won by about two to one. Rockefeller was a MoMA trustee and donor.
The canonical case of the audience vote as the artwork itself, with a visible tally and real stakes.
1970essay
The Uncanny Valley (Bukimi no tani)
Masahiro Mori; translated by Karl F. MacDorman and Norri Kageki
Short essay by the Japanese roboticist, published in the journal Energy in 1970 and given an authorized English translation in 2012. Mori proposes that affinity for a robot rises with human likeness, then drops into eeriness just before full resemblance.
Names the unease a nearly human poem can produce. Voters may be sensing a valley in language.
early 1970s to 2016; exhibition 2024system
AARON, and the exhibition Harold Cohen: AARON (Whitney Museum of American Art)
Harold Cohen; exhibition organised by Christiane Paul with David Lisbon
Cohen, a British-born painter, began writing a rule-based drawing program at UC San Diego in the late 1960s and named it AARON in the early 1970s. He developed it until his death in 2016. The Whitney's 2024 exhibition ran AARON's plotters live in the gallery.
Four decades of one artist working beside a program he wrote, framed as collaboration. The closest precedent for a poet and his model.
1974paper
Thomas Nagel
Paper in The Philosophical Review, October 1974. Nagel argues that an organism is conscious if there is something it is like to be that organism, and that no objective physical account captures this subjective character.
Behind every vote lies a guess about whether anyone is home in the poem's author.
1974performance
Marina Abramović
Six-hour performance at Studio Morra in Naples. Abramović stood passive beside a table of 72 objects, among them a rose, scissors, a scalpel, a gun and a bullet, and invited visitors to use them on her. The crowd cut her and aimed the gun at her.
Shows what an audience can do when given power over the artist. Singulars grants a smaller, cumulative power.
1978-1986performance
Tehching Hsieh; Rope Piece with Linda Montano
Five year-long works in New York. Hsieh lived in a cage, punched a time clock every hour, stayed outdoors, spent a year tied to Linda Montano by a rope, and spent a year without art. Each was bound by a signed statement.
Art as a rule kept over time. Singulars is also a durational contract: each performance binds the next.
1979book
Gödel, Escher, Bach: An Eternal Golden Braid
Douglas Hofstadter
Pulitzer-winning book that braids logic, art and music to ask how meaning and a self can arise from meaningless symbols. Hofstadter's answer is the strange loop, a system that refers to itself across levels.
Votes train the model that writes the poems that draw the votes. Singulars is built as a strange loop.
1980 onwardperformance
Third Hand and related performances
Stelarc
Cyprus-born Australian performance artist who declares the human body obsolete. He performed with a robotic third arm, staged 26 flesh-hook suspensions, let remote users move his limbs through muscle stimulators, and in 2007 had a cell-cultivated ear attached to his forearm.
The extreme of human and machine coupling, staged on one body. Singulars keeps the two apart to show the seam.
1980book
Powers of Horror: An Essay on Abjection
Julia Kristeva; translated by Leon S. Roudiez
Kristeva defines the abject as what disturbs identity, system and order. It is neither subject nor object, and must be expelled for a self to form. Published in French in 1980 and in English by Columbia University Press in 1982.
A machine poem that passes as human unsettles a border. Abjection names the recoil that such blurring can provoke.
1983artwork
Roy Ascott and artists at eleven network nodes
A planetary fairy tale written over twelve days in December 1983 for the Electra exhibition in Paris. Artists in eleven cities, from Sydney to Vienna, each played a stock character and added text over the ARTEX network. Ascott called the method distributed authorship.
An early networked text with no single author. It supplies a term for what a trained model's voice is.
1984-2003artwork
Exquisite Pain (Douleur exquise)
Sophie Calle
In 1984 Calle was left by her lover at the end of a 92-day journey abroad. She retold the story again and again and asked others when they had suffered most. The installation pairs her fading account with their anonymous stories, embroidered on linen.
A method for setting one's own text beside strangers' unsigned texts until the comparison changes both.
1984-1986event
Marc Smith
Construction worker and poet Marc Smith started a poetry show at Chicago's Get Me High Lounge in November 1984 and moved it to the Green Mill in July 1986. Judges picked from the audience score each poem from zero to ten.
Put numerical audience judgment at the center of live poetry. Singulars inherits the scoring and swaps scorecards for stickers.
1985essay
Donna Haraway
Essay first published in Socialist Review in 1985 and collected in Simians, Cyborgs and Women in 1991. Haraway offers the cyborg, a hybrid of machine and organism, as a political figure without origin stories, and refuses clean borders between human, animal and machine.
Rejects the purity contest. A reminder that poet and model are already entangled before the duel begins.
1986book
Friedrich Kittler; translated by Geoffrey Winthrop-Young and Michael Wutz
Media history published in German in 1986 and in English in 1999. Kittler argues that gramophone, film and typewriter broke writing's monopoly on storing experience, and that media determine our situation. He cites Nietzsche at his typewriter: our writing tools work on our thoughts.
The writing instrument co-authors the writer. A poet who trains a model on himself tests that claim directly.
1987book
Vilém Flusser; translated by Nancy Ann Roth
Flusser asks whether alphabetic writing will survive once other codes can carry everything it carries. He treats writing as a gesture that produced historical, critical consciousness, and imagines what thinking becomes when apparatuses take over inscription. German original 1987. English translation 2011.
Frames the poet's pen and the model's tokens as two apparatuses. The contest is between gestures as well as texts.
1991book
Daniel C. Dennett
Dennett replaces the inner theater of the mind with a multiple drafts model and treats first-person reports as data to be explained. He denies that qualia exist as philosophers describe them. Critics nicknamed the book Consciousness Explained Away.
The deflationary side of the hard problem. If Dennett is right, the voters' search for an inner life is a category mistake.
1994book
Technics and Time, 1: The Fault of Epimetheus
Bernard Stiegler; translated by Richard Beardsworth and George Collins
First volume of Stiegler's major work, in French in 1994 and English in 1998. Through the myth of Epimetheus he argues that humans have no essence apart from technics, their external memory. His later work treats technology as pharmakon, both poison and cure.
A model trained on a poet's archive is external memory that writes back. Poison and cure in one apparatus.
1994-1997artwork
People's Choice: The Most Wanted Paintings
Vitaly Komar and Alexander Melamid
The artists commissioned professional polls of public taste in painting, first in the United States and then in more than a dozen countries, and painted each nation's Most Wanted and Least Wanted picture. Dia hosted the web version from 1995.
Art made by poll, with deadpan results. It shows what optimizing for votes can produce: nearly the same landscape everywhere.
1995paper
Facing Up to the Problem of Consciousness
David J. Chalmers
Paper in the Journal of Consciousness Studies, volume 2, number 3, 1995. Chalmers separates the easy problems of explaining cognitive functions from the hard problem of why any of it is accompanied by experience. He argues that reductive explanation cannot close the gap.
Source of the model name hard. No count of stickers can settle whether the winning author felt anything.
1998-2014; song cycle 2020-2023artwork
Peggy Weil; song cycle with Varispeed
Web chatbot launched on March 12, 1998 that reverses Turing's test by asking each visitor to convince it they are human. It collected conversations for over sixteen years. Weil and the composers' collective Varispeed turned the transcripts into a song cycle, premiered at Roulette in 2023.
Direct ancestor of the model named reverse. The burden of proof falls on the human, and the transcripts become the art.
1999book
How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics
N. Katherine Hayles
Hayles traces how cybernetics, from the Macy Conferences on, taught us to think of information as separable from bodies. She reads science and fiction together and argues for a posthumanism that keeps embodiment in view.
A poem on a wall has no body. Hayles asks what was erased to make human and machine text look equivalent.
2000paper
Orchestrating Impartiality: The Impact of "Blind" Auditions on Female Musicians
Claudia Goldin and Cecilia Rouse
Goldin and Rouse analysed audition records from major US symphony orchestras that introduced screens hiding candidates from the jury. They estimated the screen raised a woman's chance of advancing from certain preliminary rounds by 50 percent, while noting that some estimates carry large standard errors.
Singulars hangs both poems unattributed. This is the standard empirical case that hiding the author changes the verdict.
2006 onwardevent
Million's Poet (Sha'ir al-Milyun)
Abu Dhabi TV; contestants including Hissa Hilal
Abu Dhabi television contest for Nabati vernacular poetry, first aired in December 2006. Poets recite original work, a jury scores them and the audience also votes. In 2010 Hissa Hilal became the first woman finalist and placed third with poems attacking extremist clerics.
Poetry judged by mass vote at television scale, in Arabic. Shows both the reach and the politics of a popular ballot.
2007artwork
Take Care of Yourself (Prenez soin de vous)
Sophie Calle
Calle received a breakup email ending with the words take care of yourself. She asked 107 women, chosen by profession, to interpret it, among them a criminologist, a philosopher and a Talmudic exegete. The results filled the French Pavilion at the 2007 Venice Biennale.
One text, many readers, every response exhibited. A model for treating audience interpretation as the material of the work.
2009 onwardperformance
Jacqueline Suskin
Suskin sets up a manual typewriter in public, for years at the Hollywood Farmers Market, and writes a poem on any subject a passerby names, in a couple of minutes, for a donation. She has composed more than forty thousand such poems since 2009.
The street lineage of the poem on demand: a stranger's theme, a few minutes, a typed page handed over.
2010performance
Marina Abramović
During her MoMA retrospective, March 14 to May 31, 2010, Abramović sat silent in the atrium throughout opening hours. Visitors queued to sit opposite her and hold her gaze for as long as they chose. It was the longest solo piece of her career.
The artist's physical presence as the whole work. Singulars keeps the poet in the room while his rival has no body.
2010performance
Annie Dorsen
Theatre piece with no human performers. Two custom chatbots on laptops hold a new conversation each night, built from the Foucault and Chomsky television debate of the 1970s, Shakespeare, the Bible, philosophy and YouTube comments. Dorsen calls the approach algorithmic theatre.
A landmark of language algorithms performing live and unscripted before a seated audience, with the human author offstage.
2010; revised 2012performance
Sayonara (Android-Human Theater)
Oriza Hirata and Hiroshi Ishiguro
Short play by Oriza Hirata with roboticist Hiroshi Ishiguro, premiered at the Aichi Triennale in 2010. The android Geminoid F sits with a dying young woman and recites poems to console her. A coda added in 2012 responds to the Fukushima disaster.
An android whose stage role is to recite poetry to a human in pain. Machine consolation, performed.
2012book
Artificial Hells: Participatory Art and the Politics of Spectatorship
Claire Bishop
History and critique of participatory art from Futurism and Dada to social practice. Bishop questions the claim that involving the audience is emancipatory in itself and asks that such works be judged as art, not only as ethics.
A check on Singulars' own premise. Participation and voting are forms to be judged, not virtues by default.
2013 onwardperformance
Sean Petrie and fellow Austin poets
Austin collective founded in 2013. Its poets set up vintage typewriters at festivals, museums, schools and weddings. A guest names a topic and a poet types a custom poem while they watch, then hands it over.
Poem on demand as a working trade. The prompt and response loop predates chatbots and happens face to face.
2014 onwardartwork
Stephanie Dinkins
Dinkins films her ongoing conversations with Bina48, a social robot built by the Terasem Movement Foundation. She asks it about race, identity, family and consciousness, and shows the fragments on multiple screens. The Whitney Museum acquired a set of fragments in 2019.
A long-term dialogue between an artist and a machine that asks whose data and whose voice the machine carries.
traditional; UNESCO 2015event
Aitysh / Aitys, art of improvisation
Kazakh and Kyrgyz akyns
Improvised sung duel of Kazakhstan and Kyrgyzstan. Two akyns sit face to face with dombra or komuz and trade verses on topics chosen by the audience. The winner is named for musical skill, rhythm, originality, resourcefulness, wisdom and wit. UNESCO inscribed it in 2015.
The audience supplies the theme and two poets answer it against each other, live. Central Asian kin to Singulars.
2015artwork
The End of Signature and A.A.I.
Agnieszka Kurant
Kurant collected visitors' signatures and used software to merge them into one collective signature, which the Guggenheim projected on its facade in 2015. In A.A.I., termite colonies build mounds from colored sand, gold and crystals.
Dissolves the signature into a crowd. A model trained on audience votes is also authored by many unsigned hands.
2015 onwardperformance
Sougwen Chung
Ongoing performances in which the Chinese-Canadian artist draws alongside robotic arms she calls D.O.U.G. The first generation mimicked her gestures live. A later one, called Memory, learned from her past drawings. The project won the Lumen Prize for art and technology in 2019.
The nearest analogue in drawing: an artist performs live beside a machine trained on her own hand.
2017performance
Lauren Lee McCarthy
McCarthy installs cameras, microphones, switches and locks in a volunteer's home, then watches remotely and acts as their smart assistant. She tries to outperform Alexa by understanding the residents as a person would.
A human performing as the AI. The mirror image of the reverse Turing test, and a study of what human attention adds.
2017performance
Trevor Paglen with Kronos Quartet and Obscura Digital
On January 14, 2017 at Pier 70 in San Francisco, the Kronos Quartet played while cameras fed their image to computer vision algorithms. Projections above the stage showed what the software saw: faces boxed, gender guessed, bodies classified, often wrongly.
Live performance with the machine's reading of it made visible. The audience watches a human act and a machine judge at once.
2018performance
Uncanny Valley (Unheimliches Tal)
Rimini Protokoll (Stefan Kaegi) and Thomas Melle
Premiered at the Münchner Kammerspiele on October 4, 2018. An animatronic double of the writer Thomas Melle sits alone on stage and delivers a lecture in his place. Melle, who has written about his bipolar disorder, is not on stage.
The author replaced by his copy before a paying audience, by his own choice. Asks what the copy spares him.
2019artwork
Holly Herndon with Mat Dryhurst, Spawn and ensemble
Album released by 4AD on May 10, 2019. Herndon and Mat Dryhurst built Spawn, a voice model housed in a gaming PC, and trained it in live sessions where a vocal ensemble sang to it. Spawn sings on the record alongside them.
A model raised by a community through call and response. Training is staged as a collective ritual, as in Singulars.
2020book
Theory of the Gimmick: Aesthetic Judgment and Capitalist Form
Sianne Ngai
Ngai reads the gimmick as a capitalist aesthetic category. A gimmick seems to work too hard and too little at once, and calling something a gimmick is a judgment about labor, time and value. The book follows Ugly Feelings and Our Aesthetic Categories.
Is the machine's poem a gimmick, or is the contest? Ngai gives language for the suspicion a labor-saving poem provokes.
2021book
Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence
Kate Crawford
Crawford maps AI as an extractive industry built on minerals, energy, low-paid labor and harvested data. She takes a material and political view of its infrastructure and argues that its systems concentrate existing power.
The rival poet runs on mines, data centers and human labelers. The gallery duel sits on a material supply chain.
2021book
God, Human, Animal, Machine: Technology, Metaphor, and the Search for Meaning
Meghan O'Gieblyn
Book by a writer who left religious fundamentalism, on how metaphors for mind moved from theology into computing. O'Gieblyn combines memoir, philosophy and criticism to ask why questions once put to God are now put to machines.
Explains the religious charge around asking whether a machine can write a true poem.
2021book
Yuk Hui
The Hong Kong philosopher asks what art can contribute to thinking about technology in the era of AI and robotics. Reading Chinese shanshui painting, he argues that technics is never universal and that cultures relate cosmos and machine differently. Published by e-flux with Minnesota.
Extends Simondon and Stiegler beyond Europe. Suggests the duel frame is one cosmotechnics among several possible ones.
2023journalism
ChatGPT Is a Blurry JPEG of the Web
Ted Chiang
New Yorker essay of February 9, 2023. Chiang compares a large language model to lossy compression of the web. It returns plausible paraphrase instead of the original, as a blurry JPEG stands in for a photograph.
A working metaphor for what a fine-tuned poetry model holds: a compressed, smoothed version of a voice.
2023book
Sexuality Beyond Consent: Risk, Race, Traumatophilia
Avgi Saketopoulou
Psychoanalytic study drawing on queer of color critique and performance studies. Saketopoulou questions the ideal of healing trauma and describes traumatophilia, the ways people return to and work with their wounds. She proposes limit consent, which risks experience instead of guarding the self.
A frame for returning, performance after performance, to an encounter that may wound. Repetition as work, not failure to recover.
2023performance
Annie Dorsen
Performance lecture. A chorus of AI-generated masks speaks a GPT-written version of the lost final play of Aeschylus's Prometheus trilogy, different each night. Beside them Dorsen delivers a talk made of quotations. Premiered at Bryn Mawr in January 2023, then Theatre for a New Audience.
Sets machine text and human text side by side on stage and lets the audience weigh them. Both are made of others' words.
2024journalism
Why A.I. Isn't Going to Make Art
Ted Chiang
New Yorker essay of August 31, 2024. Chiang argues that art comes from a great many choices, that a prompt contains too few, and that language is communication only when someone intends to say something.
The strongest recent case that the machine's poem is not art. Singulars puts that claim to a vote each time.
2nd century CE, from 4th century BCE materialbook
The Contest of Homer and Hesiod (Certamen)
Anonymous, drawing on Alcidamas
Greek prose narrative that survives in a second-century form and draws on the sophist Alcidamas. Homer and Hesiod compete at funeral games in Chalcis. The crowd applauds Homer, but the presiding king, Panedes, crowns Hesiod for singing of peace and farming.
Audience and judge disagree. The founding Greek poetry duel already asks who decides which poem wins, and on what grounds.
late 7th to early 8th centuryperformance
Al-Naqa'id (The Flytings) of Jarir and al-Farazdaq
Jarir ibn Atiyah and al-Farazdaq, with al-Akhtal
A roughly forty-year exchange of matched invective poems between two Umayyad poets, by tradition performed before crowds at the Mirbad market in Basra. Each reply took up the rival's meter and rhyme. The exchanges were later collected as the Naqa'id.
Two poets, one form, one audience, forty years. The duel becomes a serial public work, which is what Singulars is building.
late 9th century onwardevent
Heian court poets and judges
Heian court contests in which Left and Right teams submitted waka on assigned topics, pair by pair. A senior poet judged each round as a win or a draw and often gave reasons. The recorded judgments became a foundation of Japanese poetic criticism.
Two poems on one assigned theme, set side by side and judged. This is the Singulars wall, a thousand years earlier.
18th century onwardevent
Mushaira and the tarahi mushaira
Urdu poets of the Deccan, Delhi and Lucknow
Urdu poetry gathering where invited poets recite in turn and listeners answer aloud with praise or calls to repeat a couplet. In the tarahi form, organizers set one line in a fixed meter and every poet must build a ghazal on it.
One shared prompt, many poems, instant audible feedback from the room. A South Asian template for the themed face-off.