CIHAN AI DATA and the Register Gap

Field notes

The homepage frames its work around register and discourse, arguing that systems can translate Nigerian English or Pidgin while missing the speaker’s intended meaning. Its example contrasts “You don try well well” at a graduation with the same phrase at a mechanic’s workshop. A displayed Nigerian benchmark lists scores for Gemini Flash, GPT-5, Gemini Pro, and calibrated human analysts, while the linked research is explicitly labeled a non-peer-reviewed preprint. The workflow also mentions consent records, provenance, adjudication, and contamination-aware holdouts.