AI Signals & Reality Checks: Privacy-by-Design Meets Agents, and Safety Filters Go Formal

AI Signals & Reality Checks (Feb 6, 2026)

Recency rule: Everything below is from the last ~48 hours (New York time). Links are to primary sources when possible.

1) Signal: “Your chats are evidence” is becoming a mainstream legal posture—and vendors are reacting with stronger crypto language

OpenAI published a pointed note arguing against The New York Times’ request for a large tranche of ChatGPT conversations, framing it as a privacy overreach and describing mitigations (scrubbing, secure review environment). The most strategic line is forward-looking: OpenAI says it’s accelerating a roadmap that includes client-side encryption for ChatGPT messages, explicitly aiming for a world where private conversations are “inaccessible to anyone else, even OpenAI.”

Why this matters as a signal:

• Discovery pressure is now a product requirement. “We store it, but we protect it” is no longer enough if courts can compel broad production.
• Client-side encryption is also a business boundary. It draws a hard line between “model provider” and “custodian of user truth.”
• But encryption collides with safety operations. If content is unreadable to the vendor, safety workflows must shift toward on-device detection, client attestation, narrow human escalation, or user-controlled disclosure.

Reality checks:

• Expect “privacy posture” to diverge by tier: consumer chat vs enterprise/workspace vs regulated verticals.
• “Client-side encryption” will be judged by details: key ownership, recovery, multi-device sync, and what metadata still flows.

Source: OpenAI, “Fighting the New York Times’ invasion of user privacy” (Feb 2026). https://openai.com/index/fighting-nyt-user-privacy-invasion/

阅读中文版本 →