Confidential documents and ChatGPT

Is it safe to upload confidential documents to ChatGPT?

It depends on which ChatGPT you use and what is in the document. On the personal plans (Free, Plus and Pro), your conversations can be used to improve OpenAI's models unless you turn that off. On the business plans (ChatGPT Business, Enterprise and Edu), OpenAI says it does not train on your data by default. On every plan, the document is sent to OpenAI and kept with the conversation, and nobody in your organisation decides what each person uploads.

For a document holding clients' or employees' personal details, or one covered by a confidentiality agreement, use a business plan at the least. Better still, use a workspace where the organisation sets the rules and personal details are masked before any model sees them.

What happens to a file you upload

The file goes to OpenAI's servers, where its text is read so the model can answer you. It stays with the conversation until you delete the conversation.

OpenAI's help pages say a deleted chat is removed from your account at once and from its systems within 30 days, unless it must be kept for security or legal reasons. A Temporary Chat is not used for training and is deleted after 30 days, though it may be reviewed for abuse in that time.

Personal plans and business plans

The difference that matters most is who controls the settings.

Free, Plus and Pro: conversations may be used to improve the models unless the person turns off "Improve the model for everyone". The setting belongs to each person, not to the company.
Business, Enterprise and Edu: OpenAI says it does not use inputs or outputs for training by default, and that the organisation owns its data. Administrators get workspace controls.
The API: OpenAI says API data is not used for training by default either.

The real risk is the account, not the plan

A company can buy a business plan and still have staff upload contracts to the personal account on their phone, because it is quicker. Then the company's documents sit under one person's settings, outside anything the company can see or change.

This is why banning AI rarely works. People reach for it because it helps. What works is an approved place that is as quick to use, with the rules set by the organisation rather than by each person.

Five questions before you upload

Does the document contain personal details of clients, patients or employees?
Does a contract or confidentiality agreement restrict sharing it with third parties?
Which account are you signed in to: your company's, or your own?
If it is a personal account, have you turned off training?
Would you be comfortable explaining the upload to the person the document is about?

What an organisation can add

Forge, by GradTensor, is a workspace for this. It uses OpenAI's and Anthropic's models underneath; what differs is what happens before the model is called, and who decides.

Personal details such as email addresses, phone numbers, card numbers and ID numbers are masked before the prompt leaves, and restored in the answer your person reads.
An administrator can make masking a rule that no member can switch off.
Access is granted per project, so a client's documents are seen only by the people working for that client.
Every AI call is recorded, without its content, so the organisation can see what was used and by whom.
The organisation chooses the model provider and can use its own key.

A supplier contract, two ways

1A finance analyst needs the payment terms and penalty clauses from a 40-page supplier contract.
2In a personal ChatGPT account, the whole contract, with the signatories' names, emails and bank details, goes to OpenAI under the analyst's own settings, and the company has no record that it happened.
3In Forge, the analyst uploads it to the supplier project, which only the finance team can open. Email addresses, phone numbers and account numbers are masked before the question reaches the model.
4The answer cites the contract it came from, and the organisation's record shows that the analyst asked a question in that project, with which model, and that masking applied.

What masking does not solve

The rest of the document's text still reaches the model. Masking removes identifiers, not the commercial content around them.
Names and street addresses are not detected. Keep that in mind for documents that are mostly about people.
Forge cannot stop someone uploading a document to their own ChatGPT account. It gives them a better place to do it.

Sources, checked 12 October 2026

Questions

Is ChatGPT Plus or Pro safe for confidential documents?

They are personal plans. Unless you turn off "Improve the model for everyone", your conversations may be used to improve OpenAI's models, and the choice belongs to you rather than to your employer. Paying for Pro does not change that.

Does ChatGPT Business or Enterprise train on my data?

OpenAI says that by default it does not use data from ChatGPT Business, Enterprise or Edu, or from its API, for training. The data is still sent to OpenAI and stored with the conversation.

What should I do if I already uploaded a confidential document?

Delete the conversation, which OpenAI says removes it from its systems within 30 days unless it must be kept for legal or security reasons. Turn off training for the future. If the document held someone else's personal data or was under a confidentiality agreement, tell whoever is responsible for it in your organisation.

Is a Temporary Chat safe for confidential information?

It is better than a normal chat on a personal plan: OpenAI says it is not used for training and is deleted after 30 days. The document is still sent to OpenAI, and may be reviewed for abuse during that time.

How can a company stop confidential data leaking into AI tools?

Write down what may and may not be uploaded, give people an approved tool that is as easy as the one they use now, and make the protections rules rather than habits. Blocking sites alone pushes the habit onto personal phones.

A sample AI policy for employees →