What you should never tell an AI
Never paste other people’s personal data, passwords or keys, anything covered by confidentiality, medical histories, bank details or internal documents you do not own. The short rule: what you type into a chat leaves your house, so do not paste anything you would not say out loud in a crowded lift.
The never list
- Other people’s personal data: names with addresses, phone numbers, identity documents, medical histories.
- Credentials: passwords, access keys, tokens, recovery codes.
- Bank details, yours or a client’s.
- Documents under a confidentiality agreement, or a client’s internal material.
- Anything a contract obliges you not to share, however harmless it looks to you.
Anonymising in twenty seconds
Most of the time the AI does not need the real details to do the job. Swap them out before you paste:
- Proper names become labels: "the client", "company A", "the supplier".
- Identifying figures become ranges: "an invoice of roughly three thousand".
- Addresses, phone numbers and document numbers: out, no replacement.
- Exact dates become relative ones: "two months ago".
The output is just as useful, and what left your computer no longer identifies anybody. You put the real names back when you paste the answer where it belongs.
The three-question filter
- Does this identify a specific person?
- Did somebody give it to me expecting it to stay put?
- Would I mind if it turned up in a screenshot?
If any answer is yes, anonymise before pasting. You do not need to memorise a policy: those three questions settle nearly everything.
And one thing about the settings
Most tools let you stop your conversations being used to train their models, and some have it switched on by default. Finding that setting once is half an hour well spent. It does not replace the list above, though: the best protection is still not pasting what did not need to leave your desk.
The question is not whether the tool is safe, but whether that information needed to leave your computer at all.