How to do it (deciding case by case)
"Safe" isn't a yes or a no: it's a scale. You move up a level as the data becomes more sensitive and the plan more protected.
Look at the plan. Personal account (free or paid): training is on by default, so the content can feed the models. Business or enterprise account: normally training on content is excluded by contract, and here uploading is much more defensible.
Sort out the settings. Turn off training (data controls on ChatGPT, Privacy on Claude) and open a temporary chat for the document. This removes the worst case: the contract becoming training material.
Ask yourself whose data it is. A quote of yours leaving your company is your choice. A contract with a client's or an employee's data isn't only yours: handing it to an external system can violate confidentiality agreements or the rules on personal data. Here the threshold rises.
Choose the right level:
- Your quote, non-sensitive data: business account or personal account with training off + temporary chat.
- Contract with third parties' data: anonymize (placeholders instead of names and figures) before uploading.
- Data under professional secrecy or health data: local AI (Ollama, LM Studio), so the file doesn't leave your computer.
The working syntax for the anonymized version of a contract:
I'm pasting you a contract in which I've replaced the sensitive parts with placeholders
("Party A", "Party B", "amount X", "date Y").
Check: payment clauses, termination, penalties and deadlines; flag any ambiguities and gaps.
Don't ask me for the real data and don't make it up.
Text: [anonymized version]
A concrete example
A real estate agency wants the AI to check the quotes for its clients. On the free personal plan it would be risky. They switch to a business plan, where training on content is excluded by contract. For the contracts that contain clients' personal details, however, they go further: they anonymize the name and tax code before uploading, even on the business plan, because that data isn't theirs to hand over. Result: they use the AI every day without exposing the clients.
When it does NOT work (and how to fix it)
If you're convinced the paid plan is enough
A paid personal subscription doesn't disable training and isn't a business plan. Fix: turn off the switch by hand, or move to a team/enterprise plan if the company handles confidential documents regularly.
If you've already uploaded a contract on an unprotected account
The file has already passed through and, if training was on, it may already have been used. Fix: turn off training right away, delete the chat, and from now on anonymize. For the past there's no recovery: the only lever is not repeating the mistake.
If the contract is a scan or an image PDF
Uploading a PDF doesn't protect anything extra: the content is read by the system anyway. Fix: the same rule as for text applies: anonymize the data or use the local AI, regardless of the file format.
A tip from someone who actually uses it
The right question isn't "is it safe?" but "what do I lose if this ends up outside?". For a standard quote the answer is "little", and with training off you can proceed. For a contract with a client's data the answer is "their trust and maybe a penalty": there you anonymize or stay local. Calibrate the protection to the harm, not to the convenience.
Frequently asked questions
If I upload a contract, does the AI keep it forever?
Not forever, but it stays in your history until you delete it, and on a personal account with training on it can be used to improve the models. With the temporary chat and training off, the conversation doesn't stay and isn't used.
Can I upload a contract if I first ask the AI not to save it?
No, writing it in the prompt doesn't change the system's behavior: the AI follows its technical rules, not your instructions about retention. The real levers are the settings (training off, temporary chat), not a sentence in the text.
For client data the company is responsible, not me who uploads it, right?
It's the misunderstanding that costs dearly. When you upload a client's data into an external AI, you're the first point of responsibility: you handled that data and you made it leave. The company answers for it, but whoever pressed send isn't covered. That's why third parties' data is always anonymized, even when "the plan is safe".