Is it safe to put confidential documents into ChatGPT?

Whether it’s safe to put confidential documents into ChatGPT depends on the plan you use, how an admin has set it up, and what the document is. The same goes for Claude and Microsoft 365 Copilot.

In short:

• On ChatGPT Free, Plus and Pro, OpenAI may train on your chats unless you turn off “Improve the model for everyone”. Rating a reply with a thumbs up or down can still send the whole conversation for training.
• ChatGPT Business, Enterprise and the OpenAI API don’t train on your data by default, and OpenAI removes deleted chats from its systems within 30 days.
• On Claude Free, Pro and Max, if you allow training, Anthropic may keep de-identified chats for up to 5 years. If you don’t, its existing 30-day retention period applies. Claude Team, Enterprise and the API are outside that rule and aren’t trained on by default.
• Microsoft 365 Copilot doesn’t use prompts, responses or Microsoft Graph data to train foundation models, and it only shows people content they already have permission to open.
• Keep client files off personal accounts entirely. A business plan with a signed data processing addendum (DPA) suits most internal documents. Anything a contract says must stay on your own systems belongs on a self-hosted model.

What does each plan keep and train on?

We read every page linked in this table on 12 November 2026. These terms change, so check the vendor’s page before you rely on a cell.

Plan	Trains on your chats by default?	How long chats are kept	Admin export and delete	Zero data retention	Data residency
ChatGPT Free, Plus, Pro	Yes, unless you opt out (OpenAI)	Until you delete them, then removed within 30 days. Temporary chats kept up to 30 days (OpenAI)	No admin. Each person deletes their own	No	No
ChatGPT Business	No (OpenAI)	Admins set retention. Deleted chats removed within 30 days	Admins can view, export and delete members’ chats	Not listed for ChatGPT	Not listed (OpenAI)
ChatGPT Enterprise	No	Admins set retention. Deleted chats removed within 30 days	Audit log through the Compliance API	Not listed for ChatGPT	10 storage regions for new workspaces. Inference in the EU, US or UAE
OpenAI API	No, since 1 March 2023 (OpenAI)	Abuse monitoring logs up to 30 days	Not applicable	With OpenAI’s approval	Set per project, 10 regions
Claude Free, Pro, Max	Your choice (Anthropic)	Up to 5 years, de-identified, if you allow training. 30 days if you don’t	No admin. Each person deletes their own	No	Not listed
Claude Team	No (Anthropic)	Kept until deleted, then removed within 30 days (Anthropic)	Primary Owner exports organisation data	No	Stored in the US
Claude Enterprise	No	Custom period, minimum 30 days. Default is indefinite (Anthropic)	Export, plus custom retention	Claude Code only, with approval	Stored in the US. US-only inference on usage-based billing
Claude API	No	Deleted within 30 days	Not applicable	With Anthropic’s approval (Anthropic)	Inference in the US or global. Storage in the US only
Microsoft 365 Copilot	No (Microsoft)	Stored with your other Microsoft 365 content. Admins set retention in Purview	Purview and Content search. People can delete their own history	Not listed	EU Data Boundary for EU customers

Some retention sits on top of every row. OpenAI and Anthropic both keep chats longer when the law requires it. Anthropic keeps chats that its safety systems flag for Usage Policy violations for up to 2 years, and on every Claude plan it keeps feedback you send with a thumbs up or down for up to 5 years.

Does ChatGPT train on my data?

On the plans for individuals, it can. OpenAI’s help page says that on Free, Plus and Pro “we may use your content to train our models.” You can opt out under Settings > Data controls by turning off “Improve the model for everyone”. The same page warns that if you rate a reply after opting out, “the entire conversation associated with that feedback may be used to train our models.” Temporary chats aren’t used for training, but OpenAI may keep a copy for up to 30 days.

OpenAI’s enterprise privacy page, updated 8 January 2026, says it doesn’t train on data from ChatGPT Business, Enterprise, Edu or the API by default, or on data pulled in through connected apps. On Business, workspace admins “can view, access, export, and delete end user conversations.” On Enterprise, admins set the retention period and get an audit log through the Compliance API.

On the API, OpenAI may keep inputs and outputs for up to 30 days to spot abuse. Zero data retention is available for eligible endpoints if OpenAI approves it, and some endpoints, such as Files and Assistants, still store data with it switched on.

How long does Claude keep your chats: five years or 30 days?

That depends on the plan and on one setting. Anthropic’s consumer terms update of 28 August 2025 applies to Claude Free, Pro and Max, including Claude Code used from those accounts. If you allow Anthropic to use your chats for training, it extends retention to five years. If you don’t, “you’ll continue with our existing 30-day data retention period.” The same post says the update doesn’t apply to Team, Enterprise, the API, Amazon Bedrock or Google Cloud’s Vertex API.

Anthropic’s privacy center adds that if you turn the setting off later, Anthropic stops using your old and new chats in future training. Chats already used in a finished or running training run stay there. Incognito chats are never used for training.

For Team and Enterprise, Anthropic says it won’t use inputs or outputs for training by default. Owners can stop staff sending feedback with the “Rate chats” setting under Organization settings > Data and Privacy, and Enterprise owners can set a retention period of 30 days or more.

What stays inside your tenant with Microsoft 365 Copilot?

Microsoft’s Copilot privacy page says prompts, responses and data accessed through Microsoft Graph “aren’t used to train foundation LLMs.” Copilot stores prompts and responses under the same contractual commitments as your other Microsoft 365 content, and admins can search them and set retention in Microsoft Purview.

With Copilot, the bigger risk is your own sharing settings. It “only surfaces organizational data to which individual users have at least view permissions.” If a salary spreadsheet is shared with the whole company in SharePoint, Copilot can find it for anyone who asks, so fix the sharing before you switch Copilot on.

The same page says web searches go to Bing, and admins can turn web search off. If you allow Anthropic’s models inside Copilot, they “are currently excluded from the EU Data Boundary.”

Is ChatGPT Business GDPR compliant? What the vendors’ DPAs say

We are not lawyers, and nothing here is legal advice. Whether your use of a tool meets the GDPR depends on what you put into it and on your own lawful basis, as well as on the vendor’s terms. Here is what each vendor offers.

• OpenAI says it can sign a DPA “for their use of ChatGPT Business, ChatGPT Enterprise, and the API.” Its DPA, effective 1 January 2026, makes OpenAI Ireland the contracting party for customers in the EEA or Switzerland, covers UK data with the standard contractual clauses as amended by the UK Addendum, and requires OpenAI to return or delete customer data when the contract ends.
• Anthropic’s DPA, with standard contractual clauses, is built into its Commercial Terms. Accepting those terms means accepting the DPA.
• Microsoft says Copilot was added to the data residency commitments in its Product Terms on 1 March 2024, and that it follows the commitments in its Product Terms and Data Protection Addendum.

None of these DPAs covers a member of staff using a personal account. A written policy has to close that gap, and our one-page AI acceptable use policy is a free place to start.

Which plans can keep data in the EU, UK or elsewhere?

OpenAI’s data residency article blocks automated readers, so we opened it in a browser on 12 November 2026. It says eligible API customers and new ChatGPT Enterprise or Edu customers can store content at rest in Australia, Canada, Europe (EEA and Switzerland), India, Japan, Singapore, South Korea, the UAE, the UK or the US. Inference residency, which keeps the model’s processing in one region, covers Europe, the US and the UAE. ChatGPT Business isn’t on the eligibility list. Some data, such as billing details and workspace names, can still sit outside the region you choose.

Anthropic’s privacy center says its commercial data “is stored in the US.” Its API data residency page offers US-only inference at 1.1 times the standard price on Claude 4.6 and later models, and lists “us” as the only storage location for now. Claude has no EU storage option.

For EU customers, Microsoft says Copilot is an EU Data Boundary service. Its Advanced Data Residency and Multi-Geo offerings include data residency commitments for Copilot.

Which settings should you change on day one?

• Sign the DPA before anyone uploads a client file.
• On ChatGPT Business or Enterprise, set a retention period, decide whether to allow response feedback (Workspace settings > Allow response feedback), and allow only the connected apps you need.
• On Claude Team or Enterprise, turn off “Rate chats” if you don’t want conversations kept for up to 5 years as feedback, set a custom retention period on Enterprise, and verify your email domain.
• On Microsoft 365 Copilot, review SharePoint sharing, add a Purview retention policy for Copilot, and decide on web search and third-party models.
• Ban personal accounts for work. For anyone who uses ChatGPT or Claude privately, show them where the training setting is.

We compare Claude Team and ChatGPT Business side by side in Claude Team vs ChatGPT Business.

Which confidential documents can go into ChatGPT?

This is our rule of thumb. Your contracts and your regulator come first.

Document	Personal account	Business plan with a DPA	Self-hosted model
Public or marketing material	Yes	Yes	Yes
Internal notes with no personal data	No	Yes	Yes
Client contracts under NDA	No	Only if the NDA allows it	Yes
HR files, health or financial records about people	No	Only after a privacy review	Yes, after a privacy review
Anything a contract or regulator says stays on your systems	No	No	Yes
Passwords, API keys, card numbers	No	No	No

Credentials don’t belong in a prompt on any model, including one you host.

When is a self-hosted model the better answer?

Run a model yourself when a contract, a regulator or your own policy says documents can’t leave your hardware, or when you need storage outside the US and your provider doesn’t offer it. Self-hosted chat front ends such as Open WebUI and LibreChat give staff a familiar chat window over a model you run, and we compare how each one handles permissions in Open WebUI vs LibreChat.

Self-hosting adds setup and upkeep, so use it for the documents that need it and keep a business plan for everything else.

If some of your documents can’t leave your own systems, see how our self-hosted AI pilot works.