Pasting Client Data into ChatGPT: The GDPR Reality Check
Published:
It depends on your plan and on what is actually in the paste. On a personal account there’s no contract sitting behind your client’s data, so when someone asks to see the paperwork, there isn’t any, and that’s true whether or not anything ever leaks. A business plan with a signed data processing addendum puts that contract in place, though it doesn’t finish the job by itself. The better news: most of the time you don’t need the client’s real data at all. This page walks through the artefacts a data protection officer will look for, what your plan does with your text by default, and the two habits that make most of the question go away. It’s education, not legal advice: for a binding answer about your situation, ask your DPO or a data protection lawyer.
The honest answer: it depends on the plan and the data type
Two dials decide this, and people usually only look at one.
The first dial is the data. A client brief with the company name, three contact names and next year’s budget is personal data plus confidential material. A description of the same problem with names stripped out is neither. Same task, same quality of answer, completely different legal weight. Most of what you actually want help with (structure, tone, logic, a first draft) lives in the second version.
The second dial is the plan. Consumer ChatGPT and business ChatGPT are different products wearing the same interface. The business tiers can sit behind a signed contract that names OpenAI as your processor. The personal tiers can’t, because OpenAI’s DPA is offered for ChatGPT Business, ChatGPT Enterprise and the API, not for a personal subscription. That single fact is why the freelancer on a personal Plus account and the agency on a Business workspace get different answers to the identical question.
So the honest verdict, in the shape we use everywhere on Skuto:
- ⬣ Client personal data on a personal plan, no contract: avoid. There’s no processor agreement, so there’s nothing to point at when someone asks.
- ▲ Client data on a business plan with a signed DPA: workable, with homework. Transparency, a lawful basis and your own records are still your job.
- ● Genuinely anonymised or placeholder text on any plan: fine. If nobody can be identified from it, there’s no personal data and no processor question. “Genuinely” is doing real work in that sentence, and the section below explains why.
What GDPR actually asks for
Strip away the acronyms and GDPR is asking you to answer three questions in writing before the fact, not after an incident.
Who is the controller here? If a client hands you their customer list to work on, they are usually the controller and you are their processor. When you then feed that list into ChatGPT, you have hired a sub-processor. Your contract with the client almost certainly has a clause about that, and it usually requires their prior authorisation. Read that clause before you read anything else. It is the shortest path to a real answer about your specific situation.
Is there a contract with the vendor? GDPR wants processing by a processor governed by a written agreement. That’s what OpenAI’s Data Processing Addendum is for. It isn’t automatic. Someone at your company has to request it, name your legal entity and sign it. Until someone does that, the paperwork doesn’t exist, whatever your subscription costs.
A worked example, because the abstraction hides the point. You run a three person agency in Munich. A retail client sends you a spreadsheet of survey responses so you can write the campaign copy. Those respondents are the client’s data subjects. The client is the controller, you are the processor, and OpenAI would be the sub-processor the moment that spreadsheet enters a chat window. Three artefacts decide how that looks when anyone asks: the sub-processor clause in your client contract, an executed DPA with OpenAI, and a line in the client’s own privacy notice. Miss any one and you can’t show your work. Notice that none of this required a leak to become a problem.
Do the people know? Transparency is the duty regulators keep landing on. If your privacy notice doesn’t mention that AI tools process client material, and your client contract doesn’t either, the gap is the finding. It’s also the least expensive thing on the list to fix: it’s a paragraph.
None of this is us telling you that you’re compliant or that you aren’t. We’re describing the artefacts a data protection officer will look for. If your organisation has a DPO, this is exactly the conversation to bring them, and bringing it early is much cheaper than bringing it late.
Plan by plan: what happens to your text by default
Here’s where prose ages badly. Training defaults, retention windows and plan names change, so we keep them in one dated dataset instead of freezing them into a sentence you might read two years from now. The shape has been stable, and the shape is what you need:
Personal plans (Free, Go, Plus, Pro). Your conversations feed model training unless you switch it off yourself, per OpenAI’s Data controls help page. Paying for a personal plan does not change that. You can opt out in Settings under Data Controls, and a Temporary Chat sidesteps both training and history. Neither gives you a contract.
Business plans (ChatGPT Business, Enterprise, the API). OpenAI states on its enterprise privacy page that business data is excluded from model training by default, and these are the tiers a DPA covers. Admin controls, workspace retention settings and a signed agreement come with them.
For the current per-plan wording, straight from each vendor’s own terms and dated, our paste checker (linked at the end of this page) shows you the live table. That’s the only place we let those facts live. It’s the only place we can keep them honest.
The safer patterns that make the question disappear
Every experienced person we know converges on the same two habits.
Anonymise, don’t summarise. Anonymisation means the person is genuinely no longer identifiable, and it’s stricter than it sounds. Deleting the surname while leaving “the CFO of the only ceramics exporter in Faenza” identifies someone perfectly well. Rare job titles, exact dates, order numbers and small-town details re-identify people faster than names do. Aim for the description a stranger couldn’t reverse.
Work in placeholders. Rewrite the input as roles and tokens: [CLIENT], [FINANCE LEAD], [AMOUNT], [DATE]. Ask for the draft. Paste the real values back in your own document afterwards. It takes about thirty seconds and it’s the whole ballgame, because the model was never helping with the identifiers anyway. It was helping with the argument.
Two smaller ones that pay for themselves. Never upload the original file when the extracted paragraph will do, because attachments carry metadata, headers and everything else on page nine that you forgot was in there. And write down what you did: which tool, which plan, what was stripped. A one-line note per project is the difference between a story and a record.
If you want the mechanics of file uploads specifically, we go deeper in our guide on uploading documents to ChatGPT, and the general boundary list lives in the never-paste rundown.
What regulators have actually said
This is not hypothetical in Europe, and the record is short enough to read.
In March 2023 Italy’s Garante ordered a temporary halt to OpenAI’s processing of Italian users’ data, citing the absence of adequate notice and of a lawful basis for the mass collection of personal data used to train the model. ChatGPT went dark in Italy for weeks. In December 2024 the same authority closed its investigation with a fine of 15 million euro, finding among other things that OpenAI had processed users’ data for training without an adequate legal basis, had fallen short on transparency, and had not notified a March 2023 breach. It also ordered a six month public information campaign explaining how people can object to their data being used for training. In March 2026 the Court of Rome annulled that fine on jurisdiction grounds: once OpenAI set up its EU base in Ireland, the GDPR one-stop-shop rule made Ireland’s Data Protection Commission the lead authority. The court did not rule on the complaints themselves.
At the European level, the EDPB’s ChatGPT Taskforce published its report in May 2024, coordinating the national investigations. Its tone is the useful part: preliminary, unresolved on several points, and explicitly not a clean bill of health for anyone.
Read those together and the pattern is clear. Regulators have gone after the vendor for the training corpus and for transparency. Nobody has yet built a case around one freelancer pasting one brief. But the reasoning that reached OpenAI (no lawful basis, no adequate notice) is the same reasoning that reaches you, and “the tool made it easy” has never been a defence anywhere.
The practical takeaway isn’t fear. It’s that the careful version of this costs you a signature, a paragraph in your privacy notice, and a placeholder habit. Run the next thing you were about to send through our check-before-you-paste tool first, and take whatever it flags to your DPO.