Is putting client data in ChatGPT a reportable data breach?

A colleague pastes a client’s letter into ChatGPT to tidy up the wording. Nothing was hacked and nobody outside the firm saw it. Data protection law sees a disclosure: the client’s details now sit with a company your firm may have no contract with.

This guide separates the two questions people merge: did a breach happen, and do you have to report it. The first mostly turns on which account was used. The second carries a deadline, so my advice is to answer it first. For OpenAI’s half of the story, see how ChatGPT handles what you type.

Does pasting client details into ChatGPT count as a disclosure?

Yes, under EU and UK law. The GDPR and UK GDPR define a personal data breach as a breach of security leading to the “unauthorised disclosure of, or access to” personal data. No attacker is needed. The European Data Protection Board calls this a confidentiality breach, and its guidelines cover “unauthorised or accidental” disclosure alike.

Australia’s regulator says it more bluntly. In its guidance on commercial AI products, staff at an insurer enter a customer’s claim details into a public chatbot, and the OAIC concludes that the insurer “is disclosing the information” to the chatbot’s owners. The reasoning travels well beyond Australia.

What turns a paste into a breach is the word unauthorised. A business account your organisation chose, used as your rules allow, is processing you decided on. A personal account opened over lunch is not, and I’d treat that one as a breach until someone shows otherwise.

Is it safe to put personal data into ChatGPT? It depends on the account

The plan someone signs in with moves the answer more than anything else. OpenAI’s pricing page shows one line for Free, Go, Plus and Pro: content is used to train its models, with an opt-out available. Its ChatGPT Business page says the opposite: “Your data is never used for training our models.”

Training is only one part. A personal account also leaves the firm without a contract with OpenAI and without any admin view of what was sent. That gap is what makes the disclosure unauthorised. The opt-out doesn’t close that gap, so my line is simple: a personal plan never sees a client file.

The API is different again. OpenAI’s developer documentation says data sent to it is not used to train models, and that abuse-monitoring logs are kept for up to 30 days unless the law requires longer. A US practice holding patient records has one more gate first: whether a business associate agreement covers the product.

Accidentally put sensitive data in ChatGPT? The first hour

In EU and UK law the 72 hours run from awareness, not from the paste. The EDPB places awareness at the point where you have “a reasonable degree of certainty” that personal data was compromised, and allows a short investigation before that. So write the time down as soon as you know.

Deleting the chat is sensible, but it doesn’t undo the disclosure: the text has already left. Removing a conversation from ChatGPT properly explains what deletion reaches and what it doesn’t. Do it after you’ve noted what was in the chat, not before.

If an outside IT provider manages your accounts, ask them today which account was used and who signed in. I’d insist on one named person owning the decision from here, because a group chat produces five versions of the same facts.

Do you have to report it? The test changes with the regime

Whether a paste must be reported is a different question from whether it happened. Most regimes ask about the risk to the people in the text, not about how the data left. The thresholds differ, and so do the clocks.

Geography is less of an escape than people hope. Article 3(2) of the GDPR reaches firms outside the EU that offer services to people inside it, much as the EU AI Act follows the people it affects. A Toronto firm offering its services to clients in Lyon is in scope for their data.

I lean towards reporting the borderline case rather than arguing months later that it was too small to count. Pick the example closest to where you and the people in your text are.

Can an employee be personally responsible for a GDPR breach?

Under the EU and UK GDPR the organisation answers. Article 24 puts the duty to have “appropriate technical and organisational measures” on the controller. The EDPB adds that employees with access to personal data are generally not controllers, but persons acting under the controller’s authority, which is Article 29’s wording.

That cuts both ways. The firm can’t hide behind the employee who pasted the letter, because writing the rule and enforcing it was the firm’s job. A ready-made AI policy closes that gap in an afternoon, and I’m against letting any team near a chatbot until it’s signed.

Professional confidentiality is a separate duty, and it usually attaches to the professional personally. It also covers a client company’s trade secrets and figures, where data protection law has nothing to say because no individual is identified.

Taking the names out first: when text stops being personal data

Recital 26 of the GDPR draws the line. Pseudonymised data that can be linked back to a person “by the use of additional information” is still personal data, while anonymous information falls outside the regulation. A placeholder swap you can reverse sits on the first side, as de-identified versus anonymised sets out.

That doesn’t make redaction pointless. If ChatGPT only ever sees placeholders, a leak or a long retention exposes nothing that points to a person, and the risk to the people in the text falls sharply. Risk is the very test Article 33 applies.

Context can still give someone away. A dismissal at a small named employer identifies the claimant to anyone in that town, so I’d read every redacted text once before it goes. For choosing software, a survey of redaction tools by type is the place to compare.

If it were my firm: three rules before anyone pastes again

I’d say no to a ban, because people use their own phones instead, where the firm sees nothing. Business accounts with a signed rule are the floor. Taking identifiers out before the text leaves is what keeps a slip from turning into a reportable event, and the rest of our guides cover each step.

That last step is what Nonimo does before ChatGPT reads a word: it swaps client names, emails, numbers and other identifiers for placeholders, then puts the real values back into the answer. Where your own software calls the API, the same swap can run on your own server.

Sources

• GDPR, Regulation (EU) 2016/679, on EUR-Lex: Articles 3(2), 4(12), 24, 29, 33 and 34; Recitals 26 and 30.
• EDPB, Guidelines 9/2022 on personal data breach notification (PDF), version 2.0, 28 March 2023: awareness, confidentiality breaches, the internal register.
• EDPB, Guidelines 07/2020 on the concepts of controller and processor, version 2.1, on the EDPB website: employees act under the controller’s authority.
• UK GDPR Article 33, legislation.gov.uk: the UK’s 72-hour rule.
• OAIC, guidance on privacy and commercially available AI products, updated 17 January 2025: the insurer example.
• OAIC, when to report a data breach: the three criteria and the quick assessment.
• FTC, Data Breach Response: A Guide for Business: breach notification laws in all states, DC, Puerto Rico and the Virgin Islands; the HIPAA and FTC health breach rules.
• 45 CFR Part 164, Subpart D, 2024 edition (govinfo.gov): HIPAA breach definition and presumption, 60-day notice to individuals, notice to HHS.
• ChatGPT pricing: personal plans train on content, with an opt-out; Enterprise has no training on business data by default.
• ChatGPT Business: “Your data is never used for training our models.”
• OpenAI, Your data, API platform documentation on developers.openai.com: API data not used for training by default; abuse-monitoring logs kept up to 30 days.
• ABA Formal Opinion 512 (PDF), 29 July 2024, copy hosted by the National Conference of Bar Presidents: informed consent under Model Rule 1.6.

Accessed 10 November 2026.