AI for lawyers: what it's safe to use, and what has already gone wrong

Start with the cautionary tale

No profession has a sharper warning than law, and it is no longer a story about American courts. In June 2025 the Divisional Court in England and Wales heard two matters together under the Hamid jurisdiction, the court's inherent power to enforce the duties lawyers owe it: Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank. In Ayinde, five fabricated authorities were put before the court in judicial review grounds, one of them presented as a Court of Appeal decision. When copies were requested, they could not be produced.

The court found the conduct improper, unreasonable and negligent. It declined to bring contempt proceedings against the junior barrister involved, taking account of her seniority and her self-referral, but referrals to the Solicitors Regulation Authority and the Bar Standards Board followed and the judgment was delivered as a warning to the whole profession.

The pattern is not new. In the 2023 Mata v. Avianca case in New York, two lawyers were fined after filing a brief citing cases ChatGPT had invented, complete with fabricated quotes. What changed in 2025 is that the same failure reached the English courts, and the regulators.

The lesson isn't "don't use AI". It's that a language model will produce a confident, correctly formatted, entirely fake citation if you let it, and that the only defence is verification. For a legal team that reflex is the whole point of training. Everything else is secondary.

Where AI actually helps a legal team

The dependable wins are in reading and first-draft work, always under a lawyer's review:

• Document summarisation: condensing long contracts, filings, or discovery into a first-pass summary a lawyer then checks.
• Clause surfacing and review: finding the indemnity, termination, or liability language across a stack of agreements so review is faster.
• First-draft standard language: drafting routine clauses, NDAs, and letters from your own templates for a lawyer to refine.
• Due-diligence triage: grouping and flagging documents so people spend time on judgement, not sorting.
• Plain-language explanation: turning dense terms into something a client or colleague can follow, with the lawyer confirming accuracy.

What these share: the model handles volume and first drafts, and a qualified human owns every output that leaves the building. Legal research is on the list only with a hard rule attached, which is the next section.

The guardrails that are non-negotiable

For a legal team, the guardrails aren't best practice, they're professional duty:

• Confidentiality and privilege: privileged or client-confidential material never goes into a consumer tool. Use only systems with the right enterprise terms and data protections.
• Verify every citation: every case, statute, and quote checked against the primary source before it's relied on. This is the rule that would have prevented every sanction above.
• Competence and accountability: the lawyer, not the tool, is responsible for the work. AI assistance doesn't lower the standard of care.
• Disclosure: knowing each court's and regulator's current rules on disclosing AI use, which are changing quickly.

A provider who can't speak to privilege and citation-verification has not trained lawyers before. This is the part that makes AI training for a legal team different from any other.

General-purpose tools or legal-specific ones

The question every legal team reaches is whether to use a general assistant like Claude or ChatGPT, or a purpose-built legal product. The honest answer is that they solve different problems and most firms end up with both.

General-purpose models are strong on the language work: summarising, restructuring, drafting from your own precedent, explaining a dense clause in plain English. They are weak precisely where lawyers need most reliability, which is knowing whether an authority exists and says what it appears to say. They have no reliable connection to a case law database, so they generate what a plausible citation looks like.

Legal-specific research products are built against actual sources and are correspondingly better on that one axis. They are not immune from error, and they cost considerably more. Neither category removes the obligation to check, which is why the tooling decision matters less than the habit around it.

The practical consequence for training: teach people which category they are holding at any moment. Most of the reported failures come from someone using a general assistant for a task that needed a source of truth, not from anyone being reckless.

What good looks like

Effective legal AI training runs on your own matters and templates (suitably walled off), and it spends as much time on when not to trust the model as on how to use it. People leave faster at first-pass review and drafting, and instinctively suspicious of any unverified citation.

That's how our hands-on AI training is built, sized from a single team to a firm-wide rollout. The free AI proficiency assessment is a quick way to see where your lawyers stand before you scope it.