AI for lawyers and compliance teams

Where things stand (as of 2026)

AI tools are strong at drafting first-pass documents, summarising long records, comparing clauses across versions, and triaging large volumes of discovery material to focus human review time. Several legal research platforms now combine a language model with a genuine, current case law database, which meaningfully reduces (but does not eliminate) the risk of fabricated citations compared with a general-purpose chatbot.

What remains unreliable: any general-purpose AI tool asked to recall specific case law, statutes, or citations from memory. These tools generate plausible-sounding legal text, and plausible is not the same as accurate. Courts in multiple jurisdictions have sanctioned lawyers for filing briefs containing citations to cases that do not exist or that don't say what was claimed. Regulators and some courts have issued specific guidance or standing orders about disclosing AI use in filings - expectations here are moving quickly and vary by court, so check what applies in your jurisdiction and don't assume last year's rule still holds.

The mental model

A general-purpose AI tool has read an enormous amount of legal writing and can produce fluent, structurally correct-looking text - but it does not consult a live database of current law unless it's specifically built to. It cannot tell the difference between a real case and a very well-formed-sounding fake one it has generated, because at the level it operates, both are just "plausible next words." Treat every case name, citation, or statutory reference it produces as an unverified claim, not a fact.

Scenarios that work

Where it fails, and what it costs you

Fabricated citations. This is the single most publicised failure mode in the legal profession's use of AI. A model asked to support an argument will, if it doesn't have a real case to hand, sometimes produce a fake one with a plausible name, citation format, and even a fabricated quoted holding. Lawyers who filed such material without checking it have faced sanctions, cost orders, and reputational damage that followed them publicly.

Privilege and confidentiality. Entering client material into a consumer-grade AI tool without a proper data processing agreement can risk waiving privilege or breaching confidentiality obligations, depending on the tool's terms and your jurisdiction's rules. Some tools retain input data for model improvement unless you're on a contract that excludes it - read the terms, don't assume.

Jurisdiction drift. A model trained heavily on US case law and statute structure will sometimes answer a UK, EU, Australian, or other jurisdiction's question using US-flavoured assumptions, phrasing something as settled law when it's actually jurisdiction-specific or simply wrong outside the US. This is easy to miss if you're moving fast.

Clause comparison blind spots. Asking AI to compare two contract versions for changes can miss a substantive change buried in reformatted text, or flag a cosmetic change as substantive. It is a useful first pass, not a substitute for a careful redline review, especially on anything high-value.

Discovery triage errors compound at scale. A small error rate across a handful of documents is manageable; the same error rate across tens of thousands of disclosure documents can bury something case-critical. Courts still expect a defensible, documented human-supervised process behind any disclosure exercise.

Disclosure expectations to courts and regulators. Some courts now require parties to certify whether and how AI was used in preparing filings. Regulators in some professions have issued guidance on acceptable AI use. Not knowing the current rule in your jurisdiction is not a defence - check before you file, not after a challenge.

How to check the answer

1. Look up every case name, citation, and statutory reference in a primary legal database - never accept it from an AI tool's output as verified.
2. Read the actual text of any cited case or provision to confirm it supports the point being made, not just the headnote or summary.
3. Run a citator or similar check to confirm a case hasn't been overturned, distinguished, or superseded.
4. Have a second qualified lawyer review any AI-assisted draft before it goes to a client or is filed.
5. Confirm no privileged or confidential material was entered into a tool without the appropriate agreement in place.
6. Check your court's or regulator's current disclosure requirements about AI-assisted work before filing.

How to read the docs and find the truth

For the law itself, always go to the primary source: official court judgment databases, your jurisdiction's legislation website, and recognised citator services. Treat any AI-generated summary of a case or statute as a lead to verify, never as the final word.

For the AI tool, read its documentation on what data sources it actually draws from - a legal AI product's marketing page should say clearly whether it's connected to a live, current case law database or is a general-purpose model with no such connection. Read the vendor's terms on data retention and confidentiality, and check for a specific statement about jurisdiction coverage; a tool built primarily for US law will often say so if you look for it, but rarely advertises the gap loudly. Check the publication date of any legal AI vendor's claims about accuracy - courts' expectations and the tools themselves both change quickly, and a review from even a year ago may be out of date.

Your one-page playbook

ROLE: You are drafting a first-pass [document type] for a qualified lawyer to review.
CONTEXT: Governing law: [jurisdiction]. Matter type: [brief description, no
privileged specifics unless this tool is approved for confidential data].
TASK: [draft / summarise / compare / triage] the following: [attach or paste].
CITATIONS: Do not state or imply that any case, statute, or citation you produce
has been verified. If you reference one, tell me to verify it in a primary
database before use.
OUTPUT IS A DRAFT: I will independently verify all legal citations, jurisdiction
assumptions, and factual claims before this is used or filed.

Check yourself

1. What is your firm's rule on entering client-confidential material into an AI tool, and does the tool you're about to use comply with it?
2. If an AI tool gives you a case citation, what specific step do you take before it appears in a filed document?
3. Name one reason a general-purpose AI tool might confidently give you the wrong answer about a non-US jurisdiction's law.
4. What does your court or regulator currently require you to disclose about AI use in a filing, and when did you last check that requirement is still current?