Roughly 480 people a month type "ChatGPT for lawyers" into Google, and most of the answers they find are either breathless or dismissive. Neither is useful. ChatGPT is genuinely good at a specific slice of legal work, genuinely bad at another slice, and the line between them is not where most people assume it is.
The failure mode worth avoiding is not "using AI". It is using one general-purpose tool for every task, including the ones where its architecture guarantees it will fail.
What ChatGPT is actually good at
The tasks where a general model performs well share a property: you already know the answer, and you need help expressing it.
- Rewriting and compressing. Turning a dense paragraph into something a client can read, or cutting a brief to fit a page limit without losing the argument.
- Structuring an argument. Give it your position and ask for the strongest counterarguments. It is a competent sparring partner precisely because it has no stake in your conclusion.
- First-pass translation. Moving a draft between French, German, Italian and English. It will not catch jurisdiction-specific terms of art, but it gets you 80 percent of the way.
- Explaining unfamiliar territory. A quick orientation on a technical subject outside your practice area, before you talk to the actual expert.
- Drafting the boring parts. Meeting agendas, interview question lists, client update emails you have written a hundred variations of.
On this slice, the risk is low and the time saved is real. Use it.
Where it fails, and why the failures matter
The tasks where it performs badly also share a property: they require the model to know something it was never given.
Citations. Ask for supporting case law and you will often get plausible references that do not exist, complete with docket numbers and dates. This is not a bug that will be patched. A general language model predicts likely text; it does not retrieve documents from an authoritative database. Courts in several jurisdictions have now sanctioned lawyers who filed briefs containing invented citations.
Current law. The model's knowledge has a cutoff, and web browsing is not a substitute for a maintained legal database. For Swiss law in particular, coverage is thin and revisions are easy to miss.
Your matter. ChatGPT has never seen the file. It does not know what the counterparty conceded in the third email, what the client said in the intake call, or which of the four drafts is operative. Everything it produces about your case is produced from what you paste into the box.
Confidentiality. Which is the constraint that decides the rest. We wrote about this at length in Is ChatGPT confidential?, but the short version: on a consumer plan your conversations may train the model, and on every plan the data leaves Swiss jurisdiction. Professional secrecy is a legal obligation, not a preference, and the nFADP adds documented requirements for transfers abroad.
The stack that actually works
Firms getting real value are not choosing between ChatGPT and a legal tool. They are running both, with a clear rule about what goes where.
Tier one, the general assistant. ChatGPT or equivalent, used only on anonymised or already-public text. Rewriting, structuring, explaining. No client names, no matter facts, no document contents.
Tier two, legal-specific AI on your own files. A system that has actually read the matter, retrieves from real sources, cites what it finds, and runs somewhere your professional-secrecy obligations survive. This is the tier where drafting, case analysis and research belong.
Tier three, your judgement. Not optional, and not delegable. Every output from tiers one and two is a draft that a qualified person approves, edits or rejects.
The mistake is collapsing tier two into tier one, because tier one is free and already open in another tab.
Where Whisperit fits
Whisperit is the tier-two layer for firms that cannot send client data abroad. It is hosted in Switzerland, processed under Swiss jurisdiction, and built so that privileged material stays inside a system your firm can defend to a regulator.
Practically, that means dictation that turns into structured legal documents, case files the assistant has actually read, research that cites its sources, and contract review that links every finding back to the paragraph it came from. If you are weighing the options in this category, our Harvey AI alternative comparison sets out how the main tools differ on hosting, price and fit for European practice, and Whisperit vs Claude vs ChatGPT covers the general-model comparison directly.
A one-page policy you can actually enforce
Most firm AI policies fail because they are too long to remember. This fits on a card:
- General assistants are for text that contains no client information. If in doubt, it contains client information.
- Anything touching a matter goes through the firm's approved, Swiss-hosted platform.
- Every citation is verified against the official source before it leaves the firm.
- Every AI output is reviewed by the responsible lawyer, who remains accountable for it.
- When a new tool appears, it gets assessed before use, not after.
The question worth asking
"Is ChatGPT good for lawyers" is the wrong question, because it treats a category of work as a single decision. The better question is which tier a given task belongs to. Once a firm can answer that consistently, the tooling choice stops being controversial and starts being obvious.