A US federal court ruled this month that a state judge accused of relying on artificial intelligence to produce a judicial decision is still protected from being sued for it, because judicial immunity covers the act of deciding regardless of how the decision was reached.

The ruling is narrower than the headlines around it suggest, and the gap between what it decided and what people took it to mean is worth setting out.

What judicial immunity is

Judicial immunity is an old doctrine, and an unusually strong one. A judge cannot be sued in a civil action for acts performed in a judicial capacity — even if the decision was wrong, even if it was made in bad faith, even if it exceeded the judge’s authority, provided it was not taken in the clear absence of all jurisdiction.

The rationale is not that judges deserve protection. It is that a judge who could be personally sued by every losing party would decide cases with an eye to personal liability rather than to the law. The remedy for a bad decision is an appeal, not a lawsuit against the person who made it.

That doctrine long predates any question about software, which is precisely why it applied here without much difficulty.

What the ruling did not decide

Three things, all of which have been widely conflated with it.

It did not hold that judges may use AI. Immunity from a civil suit is not authorisation. A judge can be immune from being sued for conduct that is nonetheless improper, and this ruling says nothing about whether the conduct was appropriate.

It did not validate the decision itself. A ruling produced with AI assistance remains appealable on its merits like any other. If it misstates the law or the facts, an appellate court can reverse it.

It did not remove other accountability. Judicial conduct commissions, disciplinary boards, court rules and legislatures all retain authority over how judges work. Immunity blocks one specific remedy — a civil damages claim — and leaves the rest untouched.

The real question underneath

Stripped of the immunity issue, the substantive problem is about what a judicial decision is supposed to be.

Courts have generally distinguished between AI as a research and drafting tool — which is closer to what law clerks, legal databases and precedent-search software already do, and which many courts permit with disclosure — and AI as the decision-maker, substituting for the judgement a judge is appointed to exercise.

The line between the two is genuinely difficult to police. If a judge asks a model to draft an opinion and then signs it after a read-through, which side of the line is that on? The output is the judge’s, formally. Whether the reasoning is, in any meaningful sense, is not a question the law has good tools for answering.

There is also a practical hazard that has already produced consequences in several jurisdictions: language models fabricate case citations. Courts in multiple countries have sanctioned lawyers who filed briefs containing invented authorities. The same failure in a judgment would be considerably more serious, and harder to catch, because nobody checks a judge’s citations with the adversarial energy that opposing counsel brings to a filing.

Where the rules are being written

Because immunity forecloses the litigation route, whatever governance emerges will come from elsewhere: court-issued practice directions on AI use, disclosure requirements when generative tools contribute to a filing or a ruling, judicial ethics guidance, and in some jurisdictions legislation.

Several US federal courts and a number of national judiciaries elsewhere have already issued standing orders requiring disclosure of AI use in submissions. Extending equivalent requirements to judges themselves is the obvious next step and has been proposed in several places.

The deeper point is one the immunity doctrine itself rests on: the legitimacy of a court depends on litigants believing a person considered their case. That belief is not a legal technicality, and it is not protected by any doctrine. It has to be earned by how courts actually work — which is why disclosure, rather than liability, is probably where this lands.

Sources

• Tom’s Hardware, on the federal court ruling upholding judicial immunity, 18 August 2026