AI models trained on African languages, tested honestly

The datasets, the gaps and the benchmark results the labs published. This report was filed for the PINCHED Technology desk, on the ground and with the people involved named on the record.

PINCHED Technology does not rewrite other people's coverage. Reporting for "AI models trained on African languages, tested honestly" started with the people closest to the story: the workers, organisers, engineers and families whose decisions shaped it, interviewed in person and over the following weeks as the situation moved.

The numbers matter as much as the scene. Where money changed hands we asked what it cost, who was paid, when, and what was left over. Where a claim could not be verified, it is not printed here — the gaps are named instead of filled.

What emerges is a pattern the continent's audiences will recognise: pragmatic work done under real constraints, with far more craft than the headline version allows. Kofi Mensah followed the thread from the first conversation through to the outcome readers can check for themselves.

PINCHED keeps technology stories open for response. Corrections, additional documents and first-hand accounts are welcome, and updates are appended to the page with the date of the change so the record stays honest.

Read more from the Technology desk, or watch and listen to the same reporting on PINCHED Stream where a filmed or recorded version exists.