AI in legal practice: what it can do and where it goes wrong

Generative artificial intelligence has moved quickly from novelty to fixture in legal practice. Firms are using it to draft and summarise documents, to review large volumes of contracts, to prepare first cuts of correspondence and to interrogate disclosure in litigation. The appeal is obvious: much legal work is language work, and tools that generate fluent text at speed promise real gains in productivity for tasks that once consumed junior lawyers’ hours.

The risks, however, have proved just as real, and the most visible is fabrication. Large language models generate plausible text rather than retrieve verified facts, and they can produce citations to cases that do not exist, quotations that were never written and statements of law that are simply wrong, all delivered with complete confidence. In law, where an argument stands or falls on authority, that failure mode is particularly dangerous.

The courts in England and Wales have made the point in unambiguous terms. In a judgment handed down in June 2025 in the linked cases of Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank ([2025] EWHC 1383 (Admin)), the Divisional Court, led by Dame Victoria Sharp, addressed instances in which lawyers had put fictitious authorities before the court that appeared to have been generated by AI tools. The court stressed that such tools are not capable of conducting reliable legal research, that answers which look coherent may be entirely incorrect, and that lawyers have a professional duty to check what they put before a court against authoritative sources, whether they use AI themselves or supervise others who do. It warned that submitting false material generated this way could expose those responsible to sanctions ranging from a wasted-costs order to contempt proceedings or referral to the regulator.

Fabrication is not the only concern. Confidentiality and legal professional privilege are put at risk if client information is entered into tools that may store or reuse it. Data protection obligations apply to personal data fed into AI systems. Bias in a model’s outputs, the opacity of how a result was reached, and questions of who is accountable when an AI-assisted piece of work goes wrong all bear on a lawyer’s core duties. The regulators have acknowledged both sides: the Solicitors Regulation Authority has examined the technology’s implications in its risk work, recognising the efficiency it can bring while flagging the professional obligations that do not change simply because a machine did part of the work.

The consistent message from bench and regulators is not that lawyers should avoid these tools, but that responsibility cannot be delegated to them. A generative model can produce a draft; it cannot be accountable for it. The duties to act competently, to check accuracy, to preserve confidentiality and to be honest with the court remain exactly where they were, resting on the human professional who signs the work. Used within those limits, AI is a powerful assistant. Trusted beyond them, as several lawyers have now discovered in public, it is a liability.