A Small Line Item, A Real Precedent
Japan's Supreme Court has asked for roughly ¥60 million (about $400,000) in its fiscal 2027 budget request to formally test generative AI as a drafting aid for judges handling civil trials — the court's first-ever budget line specifically for AI use in litigation. The money would fund a system that lets judges and court staff feed records from already-concluded civil cases into an AI model to produce draft summaries of the record and lists of contested points (争点整理), the organizational work that normally consumes hours of a judge's preparation time before a hearing.
That framing matters. This is not a proposal to have AI decide anything. It is a proposal to see whether AI can compress paperwork.
What the Pilot Already Found
The court isn't starting from zero. A committee of six judges with civil litigation experience began testing generative AI on mock case materials in January 2026, specifically to explore whether it could help organize arguments and evidence — with fact-finding and judgment explicitly walled off from the experiment from the outset. The results, as the court itself has acknowledged, were mixed in an instructive way: AI could genuinely speed up the drafting of summaries, but it also risked reinforcing a judge's existing read of a case by quietly omitting facts or points of contention that cut against it, and it could produce answers that sounded authoritative while being flatly wrong.
That second finding is the more dangerous one. A judge who catches an AI hallucinating a case citation has caught an error. A judge who doesn't notice that the AI's summary of a 400-page record silently dropped the plaintiff's strongest exhibit has absorbed a bias without knowing it. The FY2027 budget request also earmarks funds for a system to prevent leakage of personal information and trade secrets from the records fed into the model — a recognition that court filings are among the most sensitive documents an AI system could ever touch.
The Case for Moving Slowly
It's worth taking the cautious view seriously before dismissing it. Judicial reasoning is supposed to be transparent enough that a losing party can identify exactly which facts and arguments a court weighed, and appellate courts depend on that transparency to review decisions. An AI-drafted summary that subtly reshapes which facts are salient — even without ever touching the verdict — can distort that record in ways neither party nor a reviewing court can easily detect, because the distortion happens upstream of any decision a judge signs their name to. Automation bias, where a human reviewer defers to a machine-generated draft simply because it looks polished and complete, is a well-documented failure mode in professions from radiology to aviation; there's no reason courts would be immune. Critics who want Japan's judiciary to move deliberately here, rather than treat AI adoption as inevitable, are not being reactionary — they're describing a real and specific risk that the pilot itself surfaced.
Why the Design Answers That Case
But look at what the Supreme Court actually built in response to its own findings, rather than what a generic "courts adopt AI" headline implies. Fact-finding and judgment remain off-limits by design, not as an afterthought. The test uses closed records from cases that are already finished, so no live litigant's outcome depends on the experiment. The budget itself is tiny — ¥60 million is a rounding error against Japan's judicial budget — which tells you this is a measurement exercise, not a procurement for rollout. And the court is publishing its own risk findings (bias reinforcement, convincing-but-false output) rather than burying them, which is exactly the behavior you want from an institution before it hands any tool more responsibility.
This is what proportionate regulation of judicial AI should look like: identify the narrow task where AI plausibly saves time (summarizing a record, listing disputed points), test it against real risk criteria before scaling it, and keep the boundary around judgment itself explicit and enforced rather than assumed. Japan's courts are chronically stretched — civil trial judges routinely manage dockets running past a hundred active cases — and reflexively banning any AI assistance because a poorly-designed version might fail somewhere else would leave that strain unaddressed for no offsetting safety gain. The alternative to this pilot isn't a bias-free status quo; it's judges continuing to do the same drafting work unaided, with no data at all on where AI would help or hurt.
What to Watch
The FY2027 request has to survive Japan's normal budget process, which runs through cabinet approval in December and Diet passage by spring — so the actual test likely doesn't start until deep into 2027. The real test of Japan's approach won't be whether the pilot expands; it will be whether the guardrails identified in the six-judge trial — the exclusion of fact-finding, the omission-bias warning, the confidentiality system — survive contact with a legislature and a public eager for courts to move faster. If Japan holds that line as the program grows, it will have built a genuinely exportable model for judicial AI. If the line quietly erodes once the technology looks more capable, the bias the pilot already flagged won't have gone away — it will just be running at scale.