Can your employer see what you type into ChatGPT?

The question underneath the question

When somebody asks whether their employer can see what they type into ChatGPT, they are usually asking two different things at once and have not separated them yet.

The first is whether the AI company hands conversations over to the employer. The second is whether the employer can see it by some other route, with the AI company playing no part at all.

These have different answers. The second one is the answer most people get wrong, and it is the one that actually costs people their jobs.

Question one: whose account is it

This turns entirely on which account you are signed in to. Not on which plan you pay for. Not on whether it is a browser or an app. On whose workspace the account belongs to.

A ChatGPT Enterprise or Edu workspace. Assume yes, and plan accordingly. OpenAI announced compliance tooling for Enterprise in July 2024, and the current OpenAI Compliance Platform gives an organisation a record of time-stamped interactions that includes conversations, uploaded files, workspace GPT configuration and metadata, memories, and workspace users. It exists to feed the eDiscovery and legal-hold systems companies already run, and the launch partners were exactly the names you would expect in that world: Microsoft Purview, Palo Alto Networks, Zscaler, Relativity, Smarsh.

Two details are worth carrying with you. Only a workspace owner, not any administrator, can grant access to conversation messages. And the compliance logs themselves are retained for thirty days, so an organisation that wants a long archive has to be exporting to its own store.

A ChatGPT Business workspace. Here OpenAI's own documentation says two different things, and you should read both before deciding what you believe.

Its enterprise privacy page states that workspace admins "can view, access, export, and delete end user conversations in the workspace."

Its help centre, describing the same product, states that "by default, admins and owners cannot see all private member chats," and elsewhere that each user keeps their own chat history unless they choose to share a specific chat.

The reconciling words are almost certainly by default and all, which would mean the capability exists but is not switched on, or is not self-serve. OpenAI has not published how the two statements fit together, and neither page acknowledges the other. The honest position for anyone on a Business workspace is: possibly, and you cannot tell from the outside.

Your own personal account, free or paid. No. Every administrative channel described above is scoped to a workspace, and a personal account is not a member of one. There is no product surface through which an employer asks OpenAI for your conversations.

That last answer is true, and it is also the one most likely to get somebody into trouble, because of what it does not cover.

Question two: the channel that has nothing to do with OpenAI

Everything above is a statement about OpenAI. None of it is a statement about your employer.

If you are typing into a company laptop, the relevant question is not what OpenAI logs. It is what the laptop logs. Managed devices routinely carry endpoint monitoring, managed browser profiles, extensions that inspect page content, and data-loss-prevention rules that fire on patterns like a customer list or a block of source code leaving the machine. Company networks can inspect traffic. None of this requires OpenAI's cooperation, and none of it is affected by whether you used a personal account.

A personal ChatGPT account on a company laptop is not private. It is a personal account being used inside somebody else's monitored environment. The account boundary and the device boundary are different boundaries, and only one of them is yours.

This is the single most common mistake in this whole subject, and it is worth saying plainly: the account decides what the AI company can see. The device decides what your employer can see.

Microsoft 365 Copilot is the far end of the scale

If ChatGPT Business is ambiguous, Microsoft 365 Copilot is the opposite. It is documented, deliberate, and complete.

Per Microsoft's own Purview documentation, Copilot prompts and responses are stored in the user's mailbox. From there they are searchable through eDiscovery, can be added to review sets and exported, and are captured in the unified audit log along with which service was used and references to the files that were accessed. Communication Compliance policies cover them. Retention policies can keep or delete them.

This is not a loophole or an oversight. It is the product working as designed, for customers who need exactly that. But it means the mental model of "the chat is between me and the assistant" is simply wrong for Copilot inside a corporate tenant, and it is wrong by a wide margin.

Google Gemini is a split picture, and recently changed

Gemini in Google Workspace has two mechanisms that are easy to confuse.

Audit logs carry metadata, not content. Google's July 2025 update describes what is logged: the action a user performed within a Workspace application, the app Gemini was used in, and which features were used, for example "help me write." That is adoption telemetry. It is not your prompt.

Google Vault now covers Gemini app content. As of June 2026, Vault supports retention rules and litigation holds for the Gemini app. That is a content preservation and discovery surface, and it is a different thing entirely from audit logging.

Any article written before mid-2026 that tells you Workspace admins only see Gemini metadata was accurate about audit logs and is now misleading as a general claim. Both mechanisms exist. They do different jobs.

Deleted does not always mean deleted

OpenAI's stated policy is that deleted conversations are removed from your account immediately and scheduled for permanent deletion within thirty days, unless retention is required for security or legal obligations.

That final clause did a lot of work during 2025. In the New York Times litigation against OpenAI, a preservation order required output log data that would otherwise have been deleted to be retained instead. The circulating versions of this story are both wrong, in opposite directions.

What the stipulation and order filed on 9 October 2025 actually says: the blanket going-forward preservation obligation terminated as of 26 September 2025. Data already segregated before that date remains preserved, except for users in the EEA, Switzerland and the United Kingdom. And going-forward preservation continues for accounts linked to 85 specified domains listed in an appendix, regardless of where the user is.

So neither "OpenAI keeps everything forever" nor "the order was lifted, thirty days again" is correct. The appendix listing those 85 domains is not public in the version on the record, which means there is no way for an ordinary reader to check whether their employer's domain is on it. That is an unsatisfying answer, and it is the true one.

Working out your own answer

Four questions, in order. They take about a minute.

• Whose account? Personal, or a workspace your employer owns. If it is a workspace, assume the conversations are retrievable and stop there.
• Whose device? A company laptop or phone puts everything in scope regardless of the account, through channels that have nothing to do with the AI vendor.
• Whose network? Same logic, narrower reach.
• Which product? Microsoft 365 Copilot in a corporate tenant is the most visible by design. ChatGPT Enterprise is visible by design. ChatGPT Business is undocumented. A personal account on your own hardware is the only combination where the honest answer is no.

The useful habit is not paranoia. It is noticing that "can they see it" has never been one question, and answering the right one first.

If what prompted the search was a worry about something already typed, the more practical companion piece is what actually happens to the company data you paste into an AI tool, which deals with where that text goes after it leaves your screen.

This question is one chapter of Using AI at Work.