Privacy primer: what OpenAI, Anthropic and DeepMind actually do with your data
niloofar_mire · x · 2026-09-28
Privacy researcher Niloofar Mireshghallah wrote a primer on frontier-lab data policies (checked Sept 9, 2026), prompted by the privacy fallout of OpenAI's Navier–Stokes announcement.
Key points:
- Walks through OpenAI, Anthropic and Google DeepMind policies: training defaults, retention, feedback exceptions, and what deleting a chat actually removes.
- Context: OpenAI claimed a Navier–Stokes result using 10,000 agents, drawing accusations from mathematician Buckmaster; OpenAI's initial note that it "cannot rule out" de-identified usage data helping its models sparked concerns about log access, which it later denied in an update.
- Her core argument: fixating on whether anyone read the logs misses the point — the valuable secret can be a single bit, namely that "someone already made this approach work with an AI," which alone shifts where others spend time and compute. Hence Clio-style aggregation and LLM privacy filters are insufficient.
Related event: Privacy Policies of Four Major AI Labs Examined Amid Navier–Stokes Dispute(2 posts)→
More from Safety
- METR has only ~40 staff and is talent-starved as top researchers flock to frontier labs — TheZvi · 2026-09-28
- Investigation: Italian cyberweapons firm Dataflow is deeply tied to Israel's intelligence elite — cyb3rops · 2026-09-28
- Chamath: banning open-source AI in the US could tank the stock market — rohanpaul_ai · 2026-09-28
- Anthropic Bets $1B on Accenture for Embedded Evaluators as Zvi Flags the Catch — Don't Worry About the Vase (Zvi) · 2026-09-28
- LLMs can reconstruct documents from structural metadata alone, engineer finds — ChuckDBrooks · 2026-09-28
- How OpenAI, Anthropic, DeepMind and xAI actually handle your chat data: a policy primer — niloofar_mire · 2026-09-28