This is the lesson people skip and the one that ends careers. Read it before you paste your next real document.
The mental model that keeps you safe
Pasting something into a public chatbot is not like typing it into a private notepad. It is closer to handing a document to an outside contractor in another country, one you have not signed a contract with, whose staff may read it, and who may keep it.
That framing gets you to the right answer nearly every time. Would you post this document to a company you have never dealt with, with no confidentiality agreement? If no, do not paste it.
Why, specifically
Three separate risks, and only one is about the technology.
Retention and human review. On consumer tiers, conversations are typically stored, and some are read by people for quality and safety work. Many providers also train on consumer conversations by default. Business and enterprise tiers usually promise no training and shorter retention — usually, and only if your organisation actually bought that tier and configured it. Assume the free one you signed up for on your phone is the contractor with no contract.
Your duty, regardless of what the provider does. This is the part people miss. Even if the provider deleted it instantly, the disclosure already happened. A lawyer's duty of confidentiality, a doctor's duty to a patient, a signed NDA, GDPR in Europe, HIPAA in the US, India's DPDP Act, Nigeria's NDPA, Brazil's LGPD — these bind *you*. Sending client data to an unapproved third party is a breach at the moment of sending. There is no undo, and the provider's privacy policy is not your defence.
Leakage into somewhere you did not expect. Rarer than people fear, but the reason it is worth avoiding rather than reasoning about.
The list
Do not paste into a personal or unapproved AI account:
- Client, patient, or student records. Names, case details, diagnoses, grades, notes. A named patient's history is the clearest possible violation.
- Anything under an NDA. The whole point of an NDA is that you do not pass it on.
- Personal data about identifiable people — employees, customers, applicants. CVs. Complaints. Disciplinary files. Salary lists.
- Credentials. Passwords, API keys, database strings, card numbers. Ever, anywhere.
- Unpublished financials, M&A, pricing strategy, unfiled tax positions.
- Trade secrets and proprietary formulations. Source code where your employer has said no.
- Anything a journalist could put in a headline next to your employer's name.
What to do instead — this is the practical part
Most of the value survives with a little care.
Strip the identity, keep the structure. "A 54-year-old patient with these three symptoms" instead of a name and file number. "Client A" instead of the company. "A tenant in a two-year lease" instead of the address. The reasoning help is nearly always in the shape of the problem, not the identity.
Ask about the category, not the case. "What are the usual conditions for hardship relief in a case like this?" rather than pasting the file.
Use the approved tool. If your employer has a licensed enterprise version, use it — that is what it is for. If they have not, ask. "Is there an approved tool for this?" is a career-enhancing question and takes one email.
When in doubt, do not. The value of any single AI-assisted task is maybe twenty minutes. The cost of one confidentiality breach is your registration, your job, or a regulatory fine measured in percentages of turnover. Those are not close.
One more thing
If you have already pasted something you should not have, tell someone today. Nearly every organisation handles an early self-report far better than a discovery six months later. This gets worse with silence, always.
Before you move on