Pasted personal data into ChatGPT by mistake: what to do now

What should you do first?

Stop entering further data, write down what happened, and inform your data protection officer straight away. In parallel, find out which account and plan were involved and what data is in the chat. Whether the incident has to be reported to the supervisory authority is for the controller and the data protection officer to answer, with Article 33(1) GDPR in mind: if the incident is a personal data breach, the controller notifies without undue delay and, where feasible, not later than 72 hours after having become aware of it, unless the breach is unlikely to result in a risk to the rights and freedoms of natural persons.^[2]

If you want to settle beforehand which inputs are acceptable at all, read Is ChatGPT GDPR compliant?. This page is about the case where it has already happened.

What are the steps in the first hour?

The steps below are a practical order, not a statutory scheme. Note the time of each step; the notes help with the documentation later.

1. Stop typing. No more messages in the same chat, and no “correction” that repeats the data. Do not try to ask the AI to “forget” the data either; that does not change what was transmitted.
2. Tell your data protection officer. At once, by phone or message, without forwarding the data itself. If you have no data protection officer, tell management and whoever handles data protection in your organisation.
3. Clarify account and plan. Was it a private account, a consumer plan (Free, Go, Plus, Pro) or a workspace in ChatGPT Business, Enterprise or Edu? That determines what the provider says it does with inputs (see below). Also note whether the account uses a company or a private address.
4. Identify the data. Which people and which kinds of data are in the prompt or the uploaded file? Go through the whole file, including headers, comments and metadata. Count how many people are affected. Check whether special categories, such as health data, are included.
5. Secure, then delete. Note date, time, plan and extent before you delete the chat, and agree with your data protection officer whether a copy of the chat is needed as evidence. A screenshot creates one more copy of the data, so take one only if agreed and store it protected. Then delete the chat from the history and note when you did it.
6. Switch off training. In a consumer plan, turn off data sharing (“Improve the model for everyone”) under Settings, Data controls. This applies to the future and does not delete saved chats.^[4]
7. Identify others involved. Did anyone forward the text or file, paste it into another tool or leave it in a shared chat? Ask the person who reported it, without blame. People who report mistakes openly should be able to keep doing so.

What do deleting and switching off training do, and what do they not do?

Both limit future consequences, but neither undoes the transmission. According to OpenAI’s help page “Data controls in ChatGPT”, data sharing is on by default in personal workspaces with Free, Plus or Pro. It can be turned off under Data controls; turning it off does not delete saved chats, and the history stays.^[4] If someone gives feedback with a thumbs up or down, the page says the entire associated conversation may be used to train OpenAI models.^[4]

A temporary chat does not appear in the history and, according to OpenAI, is not used to improve its models, but it may be retained for up to 30 days for safety purposes.^[4] Data entered in a temporary chat cannot therefore be treated as gone immediately.

What deleting a normal chat does at the provider in detail (periods, backups) is not something the sources we checked tell us. If the answer matters for your assessment, ask the provider; with a business plan, use the contact named in your contract. For business plans OpenAI states that by default it does not use data from ChatGPT Business, Enterprise, Edu or the API for training or improving its models.^[5] For a data processing agreement OpenAI names ChatGPT Business, Enterprise and the API; for the consumer plans it names none.^[6]

Why can you not simply take it back?

Because a sent input cannot be recalled, and you cannot see what happened to it afterwards. The sender has no technical way to observe where a provider processes, copies or stores a prompt. What you have is the provider’s documentation and, on a business plan, your contract, and for the risk assessment that is the material you work with: which plan, whether inputs are used for training, what the provider says about retention. These statements have documented limits. OpenAI, for example, says temporary chats may be retained for up to 30 days and that turning data sharing off does not delete saved chats.^[4]

That is not a finding that anything has gone wrong in your case, and it is not an argument for alarm. It explains why the first hour is about facts and why the note should record what you could and could not establish. It also explains why the durable fix sits before the send button: what staff enter is the one thing the company controls with its own means. See where your data goes when you use an AI assistant for the general picture.

Is it a “personal data breach”, and what follows if it is?

That depends on the facts; the definition is in Article 4(12) GDPR. A personal data breach is a breach of security leading to the accidental or unlawful destruction, loss, alteration, unauthorised disclosure of, or access to, personal data transmitted, stored or otherwise processed.^[1] Whether input into an AI chat amounts to “unauthorised disclosure” is assessed case by case. It depends, among other things, on whether the provider acts as a processor under a contract, whether the input was permitted by your internal rules and which legal basis exists. We do not answer that for your case here; it belongs to the assessment by your data protection officer.

If there is a breach within the meaning of Article 4(12), these rules apply:

• Notification to the supervisory authority (Article 33(1)). The controller notifies the competent supervisory authority without undue delay and, where feasible, not later than 72 hours after becoming aware of it, unless the breach is unlikely to result in a risk to the rights and freedoms of natural persons. If the notification comes later, the reasons for the delay accompany it.^[2]
• Content of the notification (Article 33(3)). At least: the nature of the breach including, where possible, the categories and approximate number of data subjects and records; the name and contact details of the data protection officer or other contact point; the likely consequences; the measures taken or proposed. If not everything is available at once, the information may be provided in phases (Article 33(4)).^[2]
• Documentation (Article 33(5)). The controller documents every breach, with the facts, its effects and the remedial action taken, so that the supervisory authority can verify compliance. This applies even where the controller does not notify.^[2]
• Communication to the people affected (Article 34). Where the breach is likely to result in a high risk, the controller communicates it to the data subjects without undue delay, in clear and plain language. Article 34(3) lists exceptions: appropriate protection measures that make the data unintelligible to unauthorised persons, such as encryption; follow-up measures after which the high risk is no longer likely to materialise; or disproportionate effort, in which case a public communication or similar measure is used. The supervisory authority can also require the communication.^[3]

Where a processor is affected, Article 33(2) requires it to inform the controller without undue delay.^[2] For everyday inputs at work, the company rather than the individual employee will usually be the controller. That is why a report has to travel up the organisation quickly: the 72 hours run from the controller becoming aware, and when that happens in your case is for your data protection officer to assess.

Which factors count for the risk assessment?

The GDPR gives no checklist. Article 33 refers to a “risk” to the rights and freedoms of natural persons and Article 34 to a “high risk”.^[2][3] As orientation for the conversation with your data protection officer, these questions help:

• Plan and contract. A consumer plan with data sharing switched on, or a business plan with a data processing agreement and no training by default?^[4][5][6]
• Kind of data. Only names and business e-mail addresses, or also account details, ID numbers, credentials, salaries, appraisals?
• Special categories. Health data, trade-union membership, religious beliefs and the other categories in Article 9(1) GDPR.^[9] They are in principle given particular protection.
• Amount. One person or a whole customer base? A file with many records weighs differently from a single entry.
• Professional secrets. Client, patient or customer data of professionals bound by secrecy is also subject to rules such as § 203 of the German Criminal Code, which also protects the secrets of legal persons.^[10] For lawyers, tax advisers and doctors, the professional chamber or legal advice belongs at the table.
• Identifiability. Was the text pseudonymised? Pseudonymised data stays personal data for you as key holder, but the risk to the people affected may be lower. The DSK also points out that removing names and addresses is regularly not enough, because context can identify people.^[7]
• Reach in time. How long was the content in the chat, who had access (including shared accounts), was it passed on?

On enforcement: as of 13 November 2026 we found no decision by a supervisory authority fining a company for its employees’ use of ChatGPT. That does not mean nothing can happen. The duties in Articles 33 and 34 apply however often authorities enforce them, and supervisory authorities have published guidance on AI use.^[7][8]

What goes into the internal incident note?

Document every incident, including those you do not notify, with the facts, the effects and the remedial action.^[2] The template below is a starting point with blanks, not a finished form; adapt it to your organisation and record the reasoning for each decision. Do not put the affected people’s real data into the note where type and number are enough.

The documentation is meant to enable the supervisory authority to verify compliance.^[2] Write it so that a third person can follow what happened without further questions.

How do you prevent the next case?

With clear rules, a company account and a way of keeping sensitive data out in the first place. The DSK notes that without clear rules employees use AI services on their own initiative, and recommends documented instructions with examples of what is allowed and prohibited.^[7] It also advises providing work devices and accounts instead of sending employees to private accounts.^[7] The Hamburg authority’s checklist recommends, among other things, business accounts, not entering personal data where the provider may use inputs for its own purposes, and checking the results.^[8]

After an incident is a good moment to update the rule: discuss the case (without names) with the team and look at which hurdle was missing. The wider basics are in use AI at work without leaking data.

What is the next step?

Call your data protection officer now, start the note and enter the time at which the controller became aware. Everything else, the risk assessment, the notification and the communication to the people affected, can only be judged with those facts. This is general information and not legal advice.

This article is general information and not legal advice. It does not replace an assessment of your own case by your data protection officer or a lawyer. Last updated: 13 November 2026.