Study: 96% of ChatGPT memories are written without user asking, 52% infer your psyche
APPSO · wechat · 2026-11-01
A paper accepted at ACM Web Conference 2026, The Algorithmic Self-Portrait, used GDPR data-access rights to analyze the full ChatGPT memories and chat logs of 80 real users (1,058 conversations, 22,971 turns, 2,050 memory entries) — the first external deconstruction of the AI memory black box.
Key findings:
- 96% of memories were written unilaterally by the system: only 4.1% of entries trace back to explicit user requests; the rest were saved because ChatGPT judged them "useful."
- Beyond facts: 28% of memories contain GDPR-defined personal data and 7% contain special-category data (61% of those health-related); 52% contain inferences about the user's mental states — desires (73%) and intentions (16%) dominate, e.g. "fears failure in major career decisions."
- Largely faithful: 84% of memories have direct textual grounding in chat history; only 4% are unsupported or contradictory.
The team also replicated ChatGPT's memory extraction with fine-tuned open-source models (60–70% semantic similarity) and built AttributionShield, which predicts what sensitive memories a query would trigger and offers de-identified rewrites before sending. In evaluation, 94–100% of rewrites significantly reduced personal attribution while 87–91% preserved answer usefulness — suggesting privacy and quality aren't a trade-off.
The article closes with advice to audit stored memories in settings, plus a ready-to-use prompt for checking what the AI knows about you.
More from Safety
- Binance's $7B hack reportedly started with a tampered Homebrew install during vibe coding — xuanalogue · 2026-11-04
- Ternary Logic proposal: enforcing AI safety at the architecture level — MoralLogs · 2026-11-04
- Cambridge's David Krueger: A global AI pause needs coordination, not trust — DavidSKrueger · 2026-11-04
- Security advisory: run AI agents in managed cloud environments, not employee endpoints — amasad · 2026-11-04
- CrowdStrike acquires AI observability startup Fiddler to secure agents — SumitGup · 2026-11-04
- AI safety researcher Jeff Ladish warns RSI is nearing and scarier than any mountain lion — JeffLadish · 2026-11-04