On a free or Plus account, no. Conversations are stored, reviewable by staff for safety purposes, and used for training unless you opt out. Team, Enterprise and API traffic are excluded from training and closer to acceptable, but confidential material still leaves your machine, so most policies require redaction first.
Short answer
On a free or Plus account, no. Conversations are stored, reviewable by staff for safety purposes, and used for training unless you opt out. Team, Enterprise and API traffic are excluded from training and closer to acceptable, but confidential material still leaves your machine, so most policies require redaction first.
Step by step
Check which plan you are on: Free, Plus and Pro are consumer plans. Team, Enterprise, Edu and API are excluded from training by default and come with data processing terms an employer can actually sign.
Ask what your obligation actually says: An NDA, HIPAA, or a client contract usually forbids sending covered material to any third party processor, regardless of that processor's settings.
Redact before you paste: Replace names, account numbers, addresses and identifiers with placeholders. The model rarely needs them to do the work you are asking for.
Use Temporary Chat for the borderline cases: It keeps the conversation out of history, memory and training, which limits how long the material persists.
Keep the record on your side: Export and convert your history so the useful output lives in your own files, then delete the conversation on the platform.
The short version: the risk is not that OpenAI is careless. It is that confidential material stops being confidential the moment it leaves the place you promised to keep it.
What the plan changes, and what it does not
| Plan | Used for training | Admin visibility | Suitable for covered data |
|---|---|---|---|
| Free, Plus, Pro | Yes, until you opt out | None | No |
| Team | No | Workspace admins | Only if your policy allows it |
| Enterprise, Edu | No | Workspace admins, compliance export | Often, with the right agreement in place |
| API | No | Yours to control | Often, with the right agreement in place |
Even on the strongest row, conversations are stored, scanned by automated safety systems, and can be preserved under legal process. That is enough to breach a strict NDA on its own, which is why "we turned off training" is not the answer a compliance team is looking for.
The categories to keep out entirely
Credentials and API keys. Patient records and anything health related. Client documents covered by a confidentiality clause. Unreleased financials. Personal data of other people, especially children. Source code your employer owns, unless they have approved the tool in writing.
Redaction is usually enough to keep working
Most prompts do not need the identifying details. "Rewrite this clause so the payment terms favour the supplier" works exactly as well with the party names replaced by A and B. Swap identifiers for placeholders, do the work, then put the real values back in your own document. You keep the productivity and remove the exposure.
Where your own archive comes in
Once the useful output exists, the safest place for it is a file you hold, not a conversation on someone else's server. Export from Settings, then Data Controls, then Export Data, and you receive a ZIP built around conversations.json. Drop the whole ZIP into ChatExports and it becomes per conversation PDF, Word, Markdown or CSV documents, converted inside your browser tab with nothing uploaded anywhere. That matters here more than anywhere else on this site: a tool that promises to help with confidentiality has no business receiving the transcripts.
With your own copy in hand, delete the conversation on the platform. Shortening how long sensitive material sits in an account you do not control is the most effective control available to you.
A one line policy that works
Nothing enters a prompt that you would not be comfortable seeing in a subpoena. Redact what you can, keep the output in your own files, and delete on the platform side.
Frequently asked questions
Can I paste a contract into ChatGPT?
Not if it is covered by a confidentiality clause and you are on a personal plan. Redact the parties and identifying terms, or use an approved Enterprise workspace.
Is ChatGPT HIPAA compliant?
Consumer ChatGPT is not. Compliance requires an agreement with OpenAI covering protected health information, which is an Enterprise and API level arrangement, not a toggle.
Does turning off training make it safe for confidential data?
It reduces one risk. Conversations are still stored, scanned and subject to legal preservation, so most confidentiality obligations are still breached by sending the material at all.
What is the safest way to use ChatGPT at work?
An approved Team or Enterprise workspace, redacted prompts, Temporary Chat for sensitive work, and your own exported archive so nothing important depends on the account.