Bias, measured.

Every programme generates a bias report at three points: after the AI scoring round, after the human review round, and after the final decisions are locked. The three reports get compared so you can see whether the AI introduced bias, the panel corrected for it, or the panel introduced new bias the AI didn't have.

The report uses the demographic categories the programme collects, and only those. If your application form doesn't ask for gender, the report can't analyse outcomes by gender. We don't infer demographics from names, schools, or photos. That's precisely the kind of pattern-matching the bias report is supposed to catch.

It's not a fairness certification. We don't hand out a green tick that says “this cohort was fair.” The report tells you where the disparities are; what to do about them is a programme-design question, not a statistics question.

It's not a substitute for a real ethics committee. If your programme has compliance obligations under NDPR, POPIA, the Kenya DPA, or the Ghana DPA, the bias report feeds those reviews. It doesn't replace them.

It's also not a tool for hiding decisions behind statistics. The bias report names the panel members, the AI version, and the scoring round. If the report says the cohort had a 12-point award-rate disparity by first-generation status, it also says who reviewed which applications and what the cited evidence was.

We can't detect bias in the criteria themselves. If the programme's rubric rewards a kind of essay-writing only one demographic produces, the bias is in the rubric, not in the scoring. Catching that requires a human who's read the rubric and understands the context. FairLens helps you see the outcome; the criteria-design conversation is yours.

We can't prove causation. A score disparity might be real, might be sampling noise, might be the consequence of an upstream pipeline (who knew about the programme; who could afford the application fee). The report shows the disparity and the sample size; it doesn't claim to know which of those it is.

We don't publish “debiased models.” The phrase implies a fix that doesn't exist. What we publish is the audit trail, the disparity numbers, and the override log. Bias mitigation is a process; see the methodology page for how the platform supports it without claiming to resolve it.