On Friday afternoon the other side's writ arrives: twenty-eight pages, and you want to know before the weekend what the argument rests on. You open ChatGPT on your phone, the account you use at home to look up recipes, and paste the text. Ten seconds later you have the gist in five points.
What happened to those twenty-eight pages depends on the account. Your duty of confidentiality does not. In a free or personal account the text can be used by default to improve the model unless you switch that off, there is no data processing agreement, the chat sits in your history and nobody at the firm can manage any of it. A business account changes those three things. It changes nothing about the fact that the text went to a third party, and whether that fits this matter remains your judgement and your firm's decision. What you can do: take out everything that makes the matter and the client recognisable before you paste. The legal question survives that without difficulty.
What happens to a writ in a free account?
Five things, and none of them is a hack.
The text can be used for training. On personal ChatGPT accounts the setting to improve the model is on by default; you have to switch it off yourself. The text sits in your chat history, on an account the firm cannot see and that you take with you when you change firms. Memory can retain what you pasted: a client name, a matter, a position, and that context can resurface in a later chat. There is no data processing agreement, because consumer terms do not come with one. And a shared link to the chat is an access path for anyone who holds it.
The full account of these settings is in ChatGPT and work data. For a lawyer the sum is simpler than for most readers of that piece: a writ in a free account is case information at a third party without any agreement, and that is exactly what the duty of confidentiality is meant to prevent.
What does a business account change, and what does it not?
| Aspect | Free or personal account | The firm's business account | What it means for your professional duty |
|---|---|---|---|
| Training on input | On by default, you switch it off | Excluded by default | Less spread, but the text has still been sent |
| Data processing agreement | None | Yes | Settles the GDPR side, not your duty |
| Administration | Nobody at the firm | The firm manages members and settings | The firm can make agreements and check them |
| History and memory | With you, on your own account | Inside the firm's environment | What sits there is still case information |
| Who decides whether it is allowed | You, alone | The firm, per tool | The decision should be visible when you open the tool |
With the business variants, input is excluded from training by default, as OpenAI describes it, and a data processing agreement sits underneath. That is the GDPR side, and it is required as soon as the text contains personal data. The difference per vendor is in free versus business AI accounts.
The duty of confidentiality is a different layer. It sits with you as a lawyer, it covers everything the client has entrusted to you, and it does not pass to a vendor, not even with a signed agreement. The privilege that goes with it protects the client because what he tells you stays with you. Whether that protection holds when a copy of his story sits with an American vendor is a question with no general answer; it depends on the matter, the tool and what exactly was shared. That is why the business account is a precondition and not a licence, and why the firm decides per tool what does not go in even with a business account.
What do you take out before you paste?
Everything that makes the matter or the client recognisable. In a writ that is more than you think.
The names of parties, their lawyers, witnesses and experts. Addresses, dates of birth, national identification numbers in family and criminal matters. The case number and the firm's own reference. Company registration numbers, bank account numbers and amounts that make the matter unique, such as a purchase price or a claimed loss. And the passages about health, criminal offences or family circumstances, because those are special category data, with or without a name attached. If you are unsure about a detail that looks harmless on its own, such as a job title or a place name, quasi-identifiers explained helps: three harmless details together often point to one person.
Put a placeholder in their place: [claimant], [defendant], [amount], [date]. The structure of the argument stays intact, and the tool still understands your question, because "on what ground does [claimant] argue that the agreement was terminated" is as answerable as the version with names. How to do that quickly and consistently is in the prompt redaction guide. If it is the whole document as a PDF, cleaning the document works better than cutting and pasting; see redact a PDF before it goes into an AI tool.
An employment lawyer receives a writ in a dismissal case in which the employee's medical history is set out over two pages. She takes out the names, the personnel number and the medical passage, replaces them with [employee], [employer] and [medical note], and asks the tool in the firm's business account which grounds for dismissal are put forward and how they are supported. The answer is as useful as it would have been with the full text. What the writ said about the employee's health never left the firm.
What does the firm decide, and what stays with you?
Whether ChatGPT is allowed, and with which account, is your firm's decision. A firm that has recorded nothing about it has not banned the tool; it has simply not decided yet, and the careful line until then is that only public information goes in. You should see that decision at the moment you open the tool, not in a protocol you once read. What such a list looks like is in which AI tools may employees use.
BeeSensible shows at firm level which AI tools are opened, with no user id and no page content, and the firm records a decision per tool in the AI Tools module. Open a tool something has been decided about, and you see that decision in the browser, with the approved alternative as the first button; continuing is always possible. Paste case information into a supported AI tool, and party names, matter numbers, amounts and company registration numbers are marked as you type, and you choose whether to mask, remove, or leave them. This is what the firm's overview looks like:
