What Not to Put Into an AI Tool

AI assistants are easy to talk to. That's exactly the problem. Because a chatbot answers in a friendly, conversational way, it can feel like a private chat with a helpful person who's on your side. So people relax, and they type things they'd never write on a public form — bank details, a relative's diagnosis, a work document marked confidential.

Here's the plain truth: an AI chat is not a private conversation. When you type something into a tool like ChatGPT, Claude, Gemini, or Copilot, your words usually travel over the internet to the company's servers, where they may be stored, may be reviewed by people, and — unless you've turned it off — may be used to help train future versions of the tool. Researchers at Stanford have pointed out that companies routinely feed user inputs back into their models, and that the privacy policies explaining this are often too dense for anyone to actually read. General-purpose chatbots also aren't bound by health-privacy laws like HIPAA, so medical details you share don't get those protections.

None of this means AI is dangerous to use. It means a little awareness goes a long way. Keep this short list out of any AI tool:

• Identifying numbers — Social Security number, driver's license, passport, account numbers.
• Financial details — credit card or bank account numbers, logins to financial sites.
• Passwords and usernames — a chatbot is not a password vault, ever.
• Medical records and lab results — if you want help understanding them, remove names and ID numbers first.
• Sensitive personal combinations — full name + home address + birthday together is a gift to scammers.
• Other people's private information — a family member's or client's details aren't yours to paste.
• Confidential work material — client data, contracts, trade secrets, internal plans.

These aren't hypothetical worries. In 2023, employees at Samsung were reported to have pasted confidential company code into ChatGPT — and once it was entered, it was no longer under the company's control. That same year, a software bug briefly let some ChatGPT users see snippets of other people's chat histories. Breaches and mistakes happen; the safest data is the data you never shared.

So what is fine to share? Plenty. General questions, anything already public, hypothetical situations, and drafts where you've stripped out the private specifics. A good rule of thumb: would I be comfortable putting this on a postcard? If yes, type away. If no, leave it out or remove the sensitive parts first.

You don't need to be afraid of these tools. You just need to know the few things to keep to yourself.

Keep reading