The three AI privacy switches

Are you sure your data isn’t training someone’s AI?

If your people are on personal logins, the honest answer is: you can’t from where you sit. The switch is theirs, and you can’t see it.

What to do, in order.

1. Find out whose login your people are on. Price isn’t the test: a personal plan can cost more than a business seat. The test is whether IT issued the login, can see it, and can revoke it. If a person signed up themselves, with any email (even a work email), it’s a personal login and the switch is theirs.
2. Shut the door. A business plan without single sign-on still has a hole: the workspace, with the firm’s chats and files, opens from any laptop at home. Managed logins through SSO come before every setting below.
3. Until both are done, have anyone on a personal login turn the switch off today. It has a different name in each:
   • ChatGPT: Settings, Data controls, “Improve the model for everyone” off.
   • Copilot (personal account): “Training on conversation activity” off. In the app updated since 18 August, Microsoft’s page names no training switch; look under Settings, Privacy, and ask.
   • Gemini (personal account): “Keep activity” off.
   • Claude: Settings, Privacy, “Help improve our AI models” off.
   • Grok: Settings, Data, “Improve the model” off.
4. On a business plan the vendor says it doesn’t train by default, so look at the other switch: what the assistant is allowed to read. That’s an admin console setting. I check mine on a schedule, and on the last pass it was on by default for mail, calendar, Drive, Chat and Tasks. Tasks had appeared in the console three days earlier, already on. Nobody at the firm had switched any of them on. The schedule is why it lasted three days and not a year. Turn off every source nobody has approved, and screenshot the console and each switch, dated.
5. Don’t shorten retention to “be safe.” At my firm, compliance sets the period against the recordkeeping rules, not the privacy page; for a broker-dealer or adviser, chats about business can be records.

Then ask your AI to check your work. Here’s the prompt I give mine:

“I’m using you at a regulated financial firm, on a work account. First say which product and plan you think I’m on, and how current your information about its settings is. Then list every setting that controls: (a) whether what I type is used to train or improve you, (b) which of my other apps and data you can read, (c) how long my prompts and your answers are kept, (d) which of these an admin sets and which I can change myself. For each one: where it is, what it’s set to by default, what a cautious firm would set it to, and what evidence I should capture. Finish with the settings you’re not sure about, so I check those first.”

Three rules. Nothing confidential in the prompt. Run it in a work account, with whoever holds the admin console beside you. Check every answer against the console before you believe it: you’re asking the AI where to look, not to be right.

Written for compliance and risk readers in regulated financial firms. Informational, not legal advice; Mike Bidun is not a lawyer.