AI Systems Evaluation & Threat Research
We evaluate AI systems under the conditions they meet in service, across the markets of Africa, and publish the method and the data with every result.
Across Africa, AI systems now take part in deciding who receives a loan, how a patient is triaged, and whether a household is registered for support, and little is known in public about how they behave in service.
A single accuracy figure describes the average case, so a model that scores well overall may still decline most of the applicants in a district where income is earned in cash. We measure where performance is weakest, and publish the harness, the notebook, and the results file alongside every finding.
Latest publications
The first evaluations are under way, and an entry appears here once the organisation responsible for the system has had its right of reply and that window has closed.