Researchers analyze 2,050 real ChatGPT memories: 96% saved without users asking
petesena · x · 2026-10-24
A study from Max Planck, Ruhr University Bochum, IIT Kharagpur and TU Delft — accepted at ACM Web Conference 2026 — analyzed 2,050 memory entries donated by 80 real ChatGPT users who pulled their data under GDPR. Key findings: 96% of memories were created unilaterally by the system, shifting agency away from users; 28% contain GDPR-defined personal data and 52% contain psychological insights; 84% faithfully reflect user context. The authors also propose Attribution Shield, a framework that flags sensitive memory inferences and suggests query reformulations to protect personal information.
More from Safety
- Claude Code spotted a malicious extension stealing cookies, saving $6,800 in crypto — Arindam_1729 · 2026-10-24
- Open-weight models won't reach Mythos-level bio capabilities by 2028, argues AI commentator — teortaxesTex · 2026-10-24
- Krzysztof Wos: Big Tech should spend billions in tokens to AI-patch every open source library — AccBalanced · 2026-10-24
- LiteLLM 1.107.0 supply chain compromise: credential theft triggered by just starting Python — yuridiogenes · 2026-10-24
- The meme moment when AI researchers admit the sandboxed VM had internet access — tekbog · 2026-10-24
- ThePrimeagen Recaps a Brutal Week: OpenCode Hacked, PHP, Node, Ruby, curl Supply Chain Chaos — burny_tech · 2026-10-24