AI confidentiality: the three-question test

The three-question test

1. Does this tool train its models on my inputs or outputs?
2. Does this tool retain my inputs after the session, and for how long?
3. Has this tool's vendor signed a data processing agreement that survives in court?

If the answer to question one is yes, do not paste anything that a client would object to seeing in a future model output. If the answer to question two is unbounded retention, treat the tool as a public archive. If the answer to question three is no, the tool is a personal productivity tool, not a professional one.

Tool-by-tool, as of April 2026

OpenAI ChatGPT

• Free and Plus: trains on inputs by default. Toggle in settings to opt out.
• Team and Enterprise: does not train on inputs. DPA available.
• API: does not train on inputs. DPA available.

Anthropic Claude

• Free: may train on inputs.
• Pro and Team: does not train on inputs by default.
• API: does not train on inputs. DPA available.

Google Gemini

• Consumer Gemini: trains on inputs.
• Workspace Gemini: does not train on customer content. DPA available.

Notion AI

• Does not train on workspace content. Notion has a published DPA.

This list changes. Verify before relying on it. The /ai/ethics page has the duty framework that sits behind the test.