This article describes general uses and issues; it does not constitute legal advice and does not address any specific procedure. The practices discussed vary considerably by country and jurisdiction.
Few fields are as exposed as the justice system to both the benefits and dangers of these tools. Law is made of text, which AI handles well. It demands absolute accuracy on references, which it does very poorly.
What has already gone wrong
Let's start there, because it is documented and instructive. Several courts in different countries have sanctioned lawyers who filed briefs citing court decisions that did not exist.
The mechanism is exactly what we described in our article on hallucinations. A legal citation has a very regular form: a court name, a date, a number. Producing something that looks like a citation is easy for a model. Producing a citation that exists is an entirely different problem.
The trap is formidable because the output is perfectly credible to anyone who does not check. A busy professional who recognises the format can miss it.
What actually works
The picture would be unfair without the uses that deliver real value.
Mass document analysis. In litigation involving tens of thousands of pages, identifying relevant documents used to take weeks of legal work. This is the clearest gain, and it is already widely adopted.
First-draft drafting. Structuring a deed, preparing a template, reformulating. The lawyer keeps control of the substance and saves time on the formatting.
Summaries for the litigant. Explaining in plain language what a decision says. This is a real need, as decisions are written in a language inaccessible to the people they concern.
Access to justice. This is perhaps the most socially important issue. Many people give up asserting their rights because they do not know they have them, or because a lawyer is too expensive. A tool that explains a procedure and helps draft a claim fills part of that gap, as we noted for accessibility.
Those who would benefit most from these tools are those least able to verify what they produce. A person without a lawyer who receives an erroneous explanation has no way of knowing it, and may base a legal step on it. The tool that democratises access simultaneously creates a category of people exposed to an error they cannot detect. That is not a reason to give up; it is a reason to insist on verifiable sources rather than answers.
The question of principle
Beyond uses, a fundamental question arises, and it is deeper than a question of accuracy.
A court decision is not merely a correct line of reasoning. It is an act taken by someone who answers for it, in the name of an institution, within a framework where the parties have been heard. This dimension of responsibility is constitutive of the act of judging.
This is exactly the void we described in our article on responsibility. A system that produced a legally impeccable decision without anyone answering for it would not be a court decision; it would be a calculation.
There is also a difficulty specific to law: legal reasoning must be reasoned. One must be able to challenge a decision by attacking its reasoning. Yet we explained in our article on interpretability that an AI produces a plausible justification after the fact, not a report of its actual computation. A rationale that does not match the actual reasoning strips the right of appeal of its substance.
Where the reasonable line lies
A criterion emerges, and it matches the one we proposed for delegation in general: distinguishing what prepares a decision from what makes it.
Sorting documents, summarising a case file, proposing a structure, flagging a precedent to verify: these are preparation tasks, where errors can be caught by the professional. Assessing credibility, evaluating damages, handing down a sentence: these are acts of judgment, where responsibility cannot be delegated.
The line is not technical; it is institutional. And it deserves to be defended explicitly, because economic pressure will naturally push to move it.
What to take away
The justice system offers a textbook case of what we repeat week after week: technical capability arrives well before questions of responsibility are settled, and it is this asymmetry that creates the risk.
The good news is that this field reacts quickly, because it is structurally organised around contestation. A lawyer sanctioned for a fabricated citation sets precedent within weeks. Few sectors have such a rapid correction mechanism, and that is probably what will best protect litigants.