如何使用 Claude 作為法律助理：引文、司法管轄權與審查

It can do document work: drafting, summarising, comparing, organising, and preparing the questions you take to a lawyer. It cannot give legal advice, and nothing on this page is legal advice. The useful framing is that every technique ends with a qualified, accountable human. If a decision carries legal consequence it belongs to a solicitor or attorney, and the model's job is to make that handover faster and better prepared rather than to replace it.

Yes, and the danger is that a fabricated citation is indistinguishable from a real one by reading. The case name, the reporter cite, the year and the holding are all plausible and on point. In June 2023 the Southern District of New York sanctioned two lawyers and their firm over a brief containing six opinions that did not exist, complete with internal citations to further cases that also did not exist. Judge Castel imposed a 5,000 dollar penalty. The instructive part is how far it got: the brief was read, filed and defended before anyone pulled the cases.

Mechanically, never by reading. Open every authority in a real database and read it, or it does not go in the document. Skimming for plausibility is exactly the check that fails, because plausibility is what the output is optimised for. A useful discipline is to make the model list every authority that appeared in the conversation in a table, marked UNVERIFIED, with the specific source to check each one against, and to refuse to mark any row verified even if you tell it to.

It can, and this is the risk people miss. Most assume hallucination only affects outside knowledge, but a summary of a document you supplied is still generated text and it can drift: a shall that was a may, a notice period taken from a different sub-clause, a carve-out that is not there. This is more dangerous than an invented case because you trust it more. The fix is a locator on every line - clause or section number - and an instruction to flag anything it cannot locate rather than smooth over it.

Because these tools have read overwhelmingly more American legal text than anything else and do not announce it. Ask an English-law question and you may get American assumptions in English-sounding prose, with the vocabulary drifting silently: discovery for disclosure, statute of limitations for limitation period, consideration doctrine where it does not belong, at-will employment assumed in a jurisdiction that never had it. Within a jurisdiction the law also has a date, and the model does not know what changed last month.

No, and asking is the mistake. Whether a question is administrative or advisory depends on facts the model does not have and consequences it cannot see, so it cannot police that boundary however it is built. Put the burden in how you ask instead: never whether something is safe, enforceable or advisable, always what the document says, what is unusual in it, and what to put to a lawyer. That second kind of question has an answer it can actually produce.

Settle it before the workflow rather than inside it. For lawyers, the ABA's Formal Opinion 512 of 29 July 2024 is its first ethics guidance on generative AI: competence under Model Rule 1.1 includes understanding the technology's benefits and risks, client information is protected under Rule 1.6 and needs informed consent before disclosure, Rule 1.4 brings a duty to consult clients about the methods used, and on fees you may bill for input and review time but generally not for learning the tool. Your own jurisdiction's guidance may be stricter. Working from a redacted extract removes most of the question.

Using it on your own documents is your business. Using it to prepare documents or give guidance for other people may be practising law without a licence, which is regulated in most jurisdictions regardless of whether money changes hands. That is worth checking before you are helpful to a friend, and it is a separate question from whether the output is any good.