LLM Providers' Zero Data Retention Claims Are Riddled With Exceptions
SlopTheCannon · reddit · 2026-11-05
A hands-on discussion of "zero data retention" (ZDR) across LLM providers:
- OpenAI's ZDR requires prior approval and it "reserves the right to make certain models ineligible"
- Anthropic stores nothing post-response except for legal compliance, but Fable 5 needs 30-day retention and can't do ZDR at all
- Gemini caches inputs in memory by default; ZDR may mean disabling caching
- Google and Amazon Bedrock's ZDR is a policy, not a guarantee
- Routers like OpenRouter look unfit for enterprise or sensitive data
The author's core question: app vendors, routers, model providers and clouds each carry their own ZDR claims and exceptions—should you just treat the weakest link as the actual policy?
More from Safety
- VeriSpoof: one audio LLM unifies deepfake detection, attribution and source verification — unimelb-hf · 2026-11-05
- Debate: did a $4B California ballot push for a single AI lab mislead the public? — NathanpmYoung · 2026-11-05
- Orinyx: an open-source MCP for independently verifying AI agent outputs against evidence — Practical_Green5775 · 2026-11-05
- Alignment lacks an engineering science: behavioral training leaves red-teaming tasks wrongly blocked — JacquesThibs · 2026-11-05
- Researchers Undo SFT Behavior by Subtracting Weight Updates from Final Checkpoints — burny_tech · 2026-11-05
- Op-ed: up to 30% of STEM PhD abstracts show AI use, and AI detectors are rarely wrong — Afinetheorem · 2026-11-05