AI and the Law: An Office Assistant, Not a Court Judge

In June 2023 a judge in New York fined two lawyers and their firm five thousand dollars, because they had submitted in a court filing six judicial opinions that did not exist, produced by ChatGPT. The fine was not for the error alone, but because they insisted on the accuracy of what they had submitted when asked about it, which the judge took as bad faith; he also ordered them to write to each judge to whom an invented opinion had been attributed, and the suit was then dismissed as time-barred (ABA Journal, 2023). The episode sums up the tool's limits: skilled at phrasing, unable to realise that it is wrong.

Legal research: faster, but not more truthful

The risk is not confined to general tools. Stanford researchers built a set of more than 200 open-ended legal queries and tested specialised legal-research tools marketed as grounded in verified references. Two of the tools erred on more than 17% of queries, and a third on more than 34% (May 2024). The errors took two forms: an incorrect answer, and an answer that cites a source which does not support what was said. It is a study released before peer review, and it does not mean every legal task will fail at this rate. Its team acknowledged that these tools reduce errors compared with general models, and did not mean to single out any one company. But it suffices to justify a single rule: no reference is relied on before it is opened and read.

Does the algorithm predict rulings?

In 2016 researchers trained a model on 584 cases from the European Court of Human Rights, split evenly between cases in which a violation was found and cases in which it was not, and it reached an average accuracy of about 79%. The best performer was reading the facts of the case, not its legal arguments. The authors themselves warn against over-reading the result: the fact sections were written by the court itself and may carry implicit judgments, and the sample consisted of cases that reached the merits stage, so it should not be generalised to lower courts. The model recognises patterns in rulings already issued; it does not weigh justice in a new case.

In the Kingdom

A study by Al Tamimi & Company (April 2024) says that the Ministry of Justice's Najiz platform offers more than 160 electronic services, and that the e-litigation project allows the exchange of memoranda, remote hearings and electronic judgments. It describes the use of AI in the judiciary as an area of planned investment within Saudi Vision 2030, and gives no figures for its actual use. It lists risks: bias when data are lacking, leakage of the parties' secrets, effects on the right to a fair trial and on judicial independence, and it calls for transparency and for reasons to be given for results.

What remains for the human being?

Judging, in the Sharia, is a heavy trust and a personal responsibility for which the judge himself is answerable. A tool, however good, bears no responsibility, hears no claimant and weighs no human circumstance; and the justice of a ruling is not measured by its match with earlier patterns but by its fairness to the person who holds the right in his case. As for practical steps for the lawyer and the organisation that uses it:

• Verify every regulation, precedent and article number at its official source before submitting it.
• Do not enter clients' names or confidential case details into a general tool.
• Treat the tool as a first draft, and leave review and signature to the specialist.
• Record where you used it, so that each party bears responsibility for what it submits.

For more on the causes of confident error, see the article «Large Language Models». For bias in data, see «Algorithmic Bias».