Where what you write in a chat goes
A message sent to a chat goes to the provider company's servers. From there its fate depends on the service's policy and your settings: how long it's kept, who can see it, whether it goes into training future models.
Basic facts for making weighed decisions:
- Major providers have privacy settings — using dialogues to train models can usually be turned off (at Anthropic, see Settings → Privacy).
- Deleting a chat from the interface and deleting it from the servers are different things; providers have retention periods.
- Business plans (Team/Enterprise) give stricter guarantees: company data doesn't go into training by default.
- Correspondence can be disclosed by court order — like email or a messenger.
The working frame: treat a chat with AI like correspondence with an external contractor. A useful one, but external.
What not to send
Red zones — don't paste into public AI services:
- Access secrets: passwords, API keys, wallet seed phrases, private keys. Never, in any form — including "just for a second, to fix the config."
- Full personal data — your own and especially others': passports, card numbers, medical diagnoses with names. For others' data you're liable by law.
- Others' secrets under NDA: a client's code, contracts, an employer's internal documents without their consent. Many companies explicitly forbid this through internal policies.
Yellow zone — case by case and with anonymization:
- Your own financial situation, work conflicts, business plans — often worth it (the value of the consultation is high), but without names and identifying details.
Anonymizing in 10 seconds: replace names with "Company A," "Client B," remove or round amounts, cut out contract numbers. For a "check this contract for risks" task the model works perfectly well with an anonymized version too.
Secrets in files
A common leak happens not through text but through an attachment: the uploaded file turns out to have more in it than you thought.
- A screenshot — with your open email in the background, a token in the address bar, names in an adjacent tab.
- A "sales" spreadsheet — with a salaries tab you forgot about.
- A config "for debugging" — with keys and passwords inside.
- A document export — with edit history and colleagues' comments.
The habit: before uploading, open the file with your own eyes and look at what's actually in it. A screenshot — crop it to the area you need.
Key takeaways
- A chat with AI = correspondence with an external contractor; check the privacy settings once, by hand.
- Absolute prohibitions: access secrets, full personal data, others' material under NDA.
- Your own sensitive material — allowed consciously and anonymized.
- Files and screenshots are the main channel for accidental leaks: watch what you upload.