tools

Wikitongues

2026 - ongoing
Wikitongues, AI Strategy Lead
Wikitongues AI platform landing page - chat, annotate, and research entry points for endangered-language work

I am the AI (Strategy) lead at Wikitongues, a nonprofit sustaining and documenting every language in the world. There are about 7,000 of them. Roughly half are endangered, and fewer than 5% are represented in AI at all. The Wikitongues AI initiative builds AI that speaks endangered languages authentically, and, more importantly, the first community-owned benchmark for how well frontier models actually speak one. When you ask ChatGPT, Gemini, or Claude to hold a conversation in an endangered language, someone has to be able to say how well it did, and who gets to decide. The benchmark is the lever. It is how you push the labs toward the languages they have been ignoring.

The first pilot is Igala, a Yoruboid language with two to three million speakers in Kogi State, Nigeria, built with community lead Agnes and the Igala Wikimedians. Off-the-shelf models fail badly at Igala. They confuse it with neighbouring languages and hallucinate fluency they do not have. The platform has three roles. Learners chat and get culturally appropriate answers with a confidence score. Annotators, who are fluent speakers, do pairwise A/B comparisons, rubric scoring, and free-text edits. Researchers read the leaderboards, the rubric scores, and the inter-annotator agreement. Behind it runs a translator, then a reviewer, then an orchestrator that scores confidence: high-confidence answers auto-pass, low-confidence ones route to human annotators. The rubric has eight parts - authenticity, cultural knowledge, idiom and metaphor, register and honorifics, lexicon, grammar and morphology and tone, orthography, and dialectal fidelity. The data stays community-owned, and it feeds a fine-tune, RAG, and DPO flywheel.

Wikitongues AI platform project page - the Igala pilot with its rubric, roles, and community-owned benchmark
the project

We raised $25,000 for the first three months, with collaborators and support from Google Research's Impact Lab, DAIR (Timnit Gebru), Georgia Tech (through a Hawaiian-language tie-in), and academic advisors at NYU, JHU, and Sydney. The public launch, and the first-ever Igala model leaderboard, land at the Wikimedia Foundation conference in Ghana on October 4-5, 2026.

Wikitongues AI platform timeline - milestones from the Igala pilot to the public leaderboard launch
the timeline