A 200-patient synthetic table stayed unique after removing all identifiers
MaziyarPanahi · x · 2026-07-28
A synthetic 200-patient table can still be re-identified even after all obvious identifiers are removed: age, sex, ZIP code, and admission date alone can generate 74,550 combinations, making uniqueness the default.
The post says OpenMed 2.0 raised the threshold to k=15, but still refused to call the dataset safe. It also argues that expert review fields must remain placeholders and that a qualified expert should sign off with their own key only after the evidence bundle is complete.
Related event: Synthetic Data Remains Identifiable Even After Removing Identifiers(2 posts)→
More from Safety
- Reddit users allege Cursor sends codebase data even with telemetry turned off — Ok-Painter573 · 2026-07-28
- A Reddit post uses Irving John Good’s 1965 essay to explain the rush of capital into AGI — ProxyLumina · 2026-07-28
- Open Weights Can Evade US Export Orders, But Not China's, Analyst Says — shashib · 2026-07-28
- U.S. defense-law sanctions may hinge on an untestable chatbot claim about Kimi K3 — IridiumEagle · 2026-07-28
- Anthropic’s CEO again argues open-weight models carry serious misuse risks — The Decoder · 2026-07-28
- Microsoft open-sources a governance toolkit for autonomous AI agents — microsoft · 2026-07-28