A note before the split. This is general guidance rather than legal or professional-conduct advice. Your bar rules, your firm's policy and your regulator override everything below, and the tools change constantly, so treat the specifics as a snapshot and the judgement underneath as the durable part.
Lawyers currently hold two pieces of knowledge about ChatGPT that sit uncomfortably together. The first is the headlines, where colleagues have been sanctioned in open court for filing briefs built on convincing cases that never existed. The second is the quiet observation that half the profession is drafting with it anyway, and that the ones doing it well seem to be getting their evenings back.
Somewhere between those two facts is a line, and nobody at the firm has drawn it for you. It is drawable, and it runs along a distinction the sanctioned lawyers missed. There is a difference between work the tool produces for your review and facts the tool asserts on its own authority. The first is delegation. The second is where the sanctions came from.
The short answer: trust ChatGPT, or the AI of your choice, with the drafting and translating that fills a legal week, meaning first versions of routine correspondence, plain-English explanations of clauses for clients, structure for arguments and documents, and summaries of long material for your own orientation. Never trust it with legal research you have not verified case by case, current law, final advice, or anything carrying client-identifying facts on the consumer version, because privilege does not survive contact with a tool that trains on what you paste.
Where it earns its place
The safe territory is work you were always going to review, which in a legal practice is a great deal of the day.
| The job | What that looks like |
|---|---|
| Routine correspondence | First drafts of the status update, the document request and the scheduling exchange, arriving in seconds and taking thirty seconds of your edit |
| Plain-English translation | Clauses and statutory language explained in words a client can follow, a service clients notice and a task the tool does well |
| Orientation summaries | A deposition transcript or an opposing brief compressed into a map of where to spend your careful reading. A map rather than a substitute, since details in summaries bend |
| Structure | Organising an argument, building a chronology from a mess of dates, outlining the advice letter before you write it properly |
What to never trust it with
| Keep it away from | Because |
|---|---|
| Legal research and citations, unverified | It invents cases with perfect formatting and full confidence, which is precisely how the sanctions happened. Every authority gets opened and read before it appears anywhere |
| Current law | Its knowledge has a cutoff and no jurisdiction awareness, so recent legislation, rules and rates may be absent or wrong |
| Client-identifying facts on consumer versions | The standard tiers can use what you paste for training and retain it, and privilege is not a setting you can toggle back on afterwards |
| Final advice and filings | It carries no liability, no duty and no insurance. All three remain attached to you |
| Anything you cannot check | If you could not tell a right answer from a convincing wrong one in that area, the tool cannot be used safely there |
The privilege rule
Confidentiality is where legal use differs from everyone else's, because the stakes are professional rather than merely embarrassing. On the consumer versions of ChatGPT, Claude and Gemini alike, inputs can feed training by default and are retained for a period, which makes pasting client facts into them a disclosure question with an unpleasant answer.
The workable rule has two halves. Anything identifiable stays out of consumer tools entirely.
- Anonymise the matter into a genuine hypothetical. A supplier dispute becomes "a mid-sized company and a vendor in a services contract", which keeps everything the tool needs to help you think and nothing that could identify a client.
- Or handle it inside a firm-sanctioned enterprise tool with a proper data agreement, which is precisely what those tools exist for.
- When in doubt, the old test still works. If it could not go in a public filing, it does not go in a public chatbot.
The same instinct applies to anything the tool asserts on its own authority. There is more on catching that in how to tell when ChatGPT is making it up, and on choosing the right tool for a given task in the free finder.
About the clients with ChatGPT
There is a newer discomfort worth naming, because it is arriving in inboxes across the profession. Clients now run advice and documents through ChatGPT and come back with challenges, sometimes trusting the chatbot's read over decades of experience.
The productive response is the same verification discipline this whole piece rests on, taken calmly in the other direction. A client's AI-generated critique gets checked against the actual source like any other claim. It is usually wrong in ways that take minutes to demonstrate, and a lawyer who can say "let's look at what the case actually says" is displaying exactly the judgement the client is paying for.
The tool that unsettles the relationship from one side is the same one returning your evenings from the other. The difference between those two experiences is never the tool. It is who in the conversation verifies.
Not sure where to start?
The right tool for the matter in front of you.
The free finder tells you which AI to open for the task in front of you, in about 60 seconds, and the full "AI, sorted." reference lands in your inbox.
Try the free finder →One question in, one answer out.
Clair helps non-technical professionals know when to trust their AI, when to check it, and when to skip it.