Module 2 of 4
Bias lab
Test outputs across varied scenarios.
Bias in these systems is rarely announced. It shows up as a consistent difference in tone, thoroughness or benefit-of-the-doubt across otherwise matched cases. The way to find it is to build matched pairs — same substance, varied surface — and compare the outputs side by side.
Run a dozen and read them together rather than one at a time. A single output looks fine; the pattern across twelve is where the problem becomes visible. Write down what you find even when the answer is "no clear difference", because that record is what makes the claim credible later.
Takeaway
Matched pairs, read together. A single output always looks fine — the pattern across a dozen is the evidence.