Courts Want AI to File. It Should Be Reading the Tentative

The Courts Found the Copy Machine. The Bench Is Still Waiting

The National Center for State Courts convened a panel this week to announce that artificial intelligence can help state courts run more efficiently by automating repetitive tasks. Nobody will argue. Dockets are swollen, clerks are scarce, and a machine that routes filings, redacts exhibits and schedules hearings without sighing is a public good. I am for it.

​

But notice what the conversation is about. It is about the mailroom. The most powerful reading and reasoning instrument ever built has arrived at the courthouse, and the institution's first instinct is to hand it a stapler.

​

I understand the caution. Judges have watched lawyers file briefs citing cases that never existed, and the lesson they drew is that the machine lies. That is half of the ancient, but the current speeds, lesson. The other half is that one machine, asked one leading question, tells you what the question implies. So does one associate. So, on a bad day, does one judge.

The Reversal You Can Read Before You Sign

​We know this because we built the thing the panels are not discussing. AI Bench is seventeen simulated judicial seats: five trial judges, five appellate justices, seven Supreme Court minds. Each is a separately running agent with its own temperament and its own research tools. They get the raw record, meaning the briefs and the orders and nothing else. They do not get whose brief it is. They read not just every filing, but the case's record in full, pull the full text of every cited authority, hunt for the cases both sides missed, and vote issue by issue, party by party- a feat no human judge can achieve. Every judge must also write the best argument against its own vote.

Your Tentative Just Lost 2-3. Better to Hear It in Chambers

We proved why the blindness matters on our own appeal. Run one was the intuitive way: a memo framed the issues, and the panel predicted a clean reversal. Run two was blind, the same record with no framing, and it came back a 3-2 coin flip leaning against us. The case and the technology were identical. The only variable was our thumb on the scale. On a second matter the blind panel confirmed the strong prediction and strengthened it. The protocol deflates inflated cases and confirms good ones, which is the only kind of honesty worth paying for.

​

Every trial judge knows the moment when the first draft of a ruling starts to feel inevitable. The tentative is written, the reasons line up, and the mind does what minds do: it hunts for justifications that can be defended later. The ruling then goes to the Court of Appeal, where three strangers with a transcript and no memory of how persuasive counsel sounded that morning read it cold. Reversal is the system's stress test. It arrives two years late, at the parties' and the judge's expense.

Run the tentative past five appellate temperaments

A simulated panel moves that test to the front. It runs the draft past five appellate temperaments that were not in the room when the first impression landed. Then sends it up: seven Supreme Court minds, brilliant and unbribable by your first impression. Reruns the whole exercise on competing models, because one model in seventeen robes may share one set of blind spots.

 

Then hand the result to the commercial legal-AI platforms and watch what happens. So far they have not kept up. Shown the Bench's analysis, they concede theirs was too shallow. If the textualist or the standard-of-review disciplinarian will not sign your tentative, better to learn that on a Tuesday in chambers than in a published opinion. That does not outsource judgment. It keeps the judge from confusing the first convincing reason with the only one. Pride is not a standard of review.

Caifornia AI Rules: not delegation of decision

​California already has the framework. Rule 10.430 required every court that permits generative AI to adopt a use policy, and standard 10.80 speaks directly to judicial officers using it in their adjudicative role. The rules say how to use it carefully: confidentiality, verification, no delegation of the decision. They do not say what it is best used for. Efficiency is the easy answer because it offends no one, since a faster clerk's office threatens nobody's self-image. A panel that tells a judge the holding is a 2-3 loser on review is a different proposition, and a more valuable one.

​

The objections are real, and I have answers from practice.

​

• Privilege and confidentiality. Keep the record in a closed environment and never feed it a sealed file.

• Hallucination. Verify every citation, in both directions. 

• Abdication. The simulated justice predicts and the real one decides. AI percentages are structured self-reports, not calibrated odds, and every memo it issues says so on its face. No output leaves without a licensed human deciding what, if anything, to do about it.

 

Automating the paperwork will save hours, but testing the reasoning will save judgments.

Lawyers are already doing this. The ones who are not will be surprised by a question seventeen unimpressed readers would have asked for free. The bench should not be the last to adopt it. Automating the paperwork will save hours, and testing the reasoning will save judgments.