Does deleting the ChatGPT conversation undo the exposure?

Removing a chat from your history doesn't guarantee the copy is gone. Providers may keep inputs for a while to monitor abuse, and text already used to train a model can't be pulled back out. Deletion is still worth doing, but treat it as tidying up rather than proof the data is unrecoverable. Note when you deleted it in case you later have to show what steps you took.

Is it less of a problem if I used ChatGPT Team or Enterprise instead of the free version?

Under the provider's current terms, business tiers say they don't train on your inputs by default, which lowers the reuse risk. The text still left your environment and reached a US provider, though, so retention, access and subprocessor questions remain. It also doesn't erase an exposure that already happened on a consumer account. Read your actual plan's data terms rather than assuming the paid tier makes it harmless.

What if the data I pasted was health records or other special-category data?

Special categories under GDPR (health, biometric data, or data revealing ethnicity, religion or similar) carry stricter obligations, so an exposure that might otherwise stay internal is more likely to cross the threshold for notifying the people affected. Document exactly what was involved and get legal advice before deciding it's minor. How sensitive the data is matters more than how many records were involved.

Answers

Accidentally pasted client data into ChatGPT: what to do now

The data is on OpenAI's servers now, and you cannot reach into them to pull it back. What you can do in the next ten minutes is stop it from being used for training, start the deletion clock, and record what happened while the details are fresh. Anything beyond that depends on whose data it was and what it contained. Here is the order to work in.

What actually happened to the data

When you pressed enter, the text left your machine and was stored with your account on the provider's servers. What happens to it next depends on the plan.

On consumer ChatGPT (Free and Plus), conversations may be used to train future models unless training is switched off in the data controls. On Team, Enterprise and API plans, the provider's terms say customer inputs are not used for training. Either way, deleting the chat removes it from your account. Under the provider's current terms the stored copy is then scheduled for deletion after a retention window. But an active court preservation order in ongoing US litigation currently pauses backend deletion of many ChatGPT logs, so deleting from your account is worth doing. Do not assume the stored copy is gone.

The harder question underneath is whether ChatGPT is GDPR compliant at all, given that client data was processed on infrastructure you do not control. For this incident, focus on the steps below.

What to do right now

Delete the conversation. Open the chat's menu and delete it. This removes it from your account. Under the provider's current terms it also schedules the stored copy for deletion, though an active court preservation order may suspend that on the back end.
Turn off model training. In settings, under data controls, disable the option that lets your chats be used to improve the model. On consumer plans this is on by default, and it only governs future chats, so do it now.
If the data was sensitive, file a privacy request. The provider runs a privacy request channel where you can ask for deletion of personal data, separate from deleting the chat in the interface. Use it if the paste contained health information, financial details, or anything a client would consider damaging.
Write down what happened. Record exactly what was pasted, when, from which account and plan, and whose data it was. Any breach assessment will need these facts, and they fade from memory within days.

Whether you have to report it

If the paste contained personal data of clients (names tied to matters, finances, health, anything identifying), treat it as a possible personal data breach under GDPR. Personal data reached an external provider without the safeguards your firm had set up for it. That fits the definition of a breach even though nobody attacked you.

A breach does not automatically mean notifying the supervisory authority. The duty to notify applies when the breach is likely to result in a risk to the people concerned. A single first name in an otherwise generic prompt will often fall below that line; a pasted client file with account numbers or medical details will usually sit above it. Where notification is required the deadline is short, 72 hours from becoming aware of the breach, so involve whoever handles privacy in your firm today.

Document the assessment either way. A short note recording what was exposed, how you judged the risk, and what you did about it is exactly what an authority expects to see if it ever asks.

Professionals bound by secrecy duties (lawyers, doctors, accountants) should also check their professional rules. Those duties are separate from data protection law and can apply even when the GDPR risk is low.

Preventing the next one

The lesson from this incident is not "be more careful". Care fails under deadline pressure, at 11pm, on the tenth prompt of the evening. The fix that holds is structural: real client identifiers should never be in the prompt at all.

That means replacing names, account numbers, addresses and case references before the text goes anywhere, and keeping the mapping on your side. For a one-off you can do this by hand; there is a walkthrough in how to anonymize data before using ChatGPT. For daily work it has to be automatic, because a manual step you must remember is the step you will skip on the day it matters.

How Cedreon handles this

Cedreon is an AI workspace for sensitive data, developed in Switzerland, built around that structural fix. Before any text leaves for the model provider, Cedreon screens it: names, IBANs, addresses and other identifiers are found by pattern matching plus layered best-effort machine detection, and replaced with placeholders. The originals stay in an encrypted vault on Cedreon's side, and the reply is filled back in for the signed-in user. The model provider never sees the real values.

Two caveats. Detection is probabilistic, so this is pattern masking, not a promise that nothing can slip through. And masking is pseudonymization rather than anonymization, which is why every exchange also lands in an append-only, hash-chained audit log: if you ever face the breach assessment described above, you can show exactly what left the house and what was replaced first.

Try Cedreon free and run a harmless example through the masking before you trust it with real work.

Common questions

Does deleting the ChatGPT conversation undo the exposure?: Removing a chat from your history doesn't guarantee the copy is gone. Providers may keep inputs for a while to monitor abuse, and text already used to train a model can't be pulled back out. Deletion is still worth doing, but treat it as tidying up rather than proof the data is unrecoverable. Note when you deleted it in case you later have to show what steps you took.
Is it less of a problem if I used ChatGPT Team or Enterprise instead of the free version?: Under the provider's current terms, business tiers say they don't train on your inputs by default, which lowers the reuse risk. The text still left your environment and reached a US provider, though, so retention, access and subprocessor questions remain. It also doesn't erase an exposure that already happened on a consumer account. Read your actual plan's data terms rather than assuming the paid tier makes it harmless.
What if the data I pasted was health records or other special-category data?: Special categories under GDPR (health, biometric data, or data revealing ethnicity, religion or similar) carry stricter obligations, so an exposure that might otherwise stay internal is more likely to cross the threshold for notifying the people affected. Document exactly what was involved and get legal advice before deciding it's minor. How sensitive the data is matters more than how many records were involved.