When Judge Xavier Rodriguez prepares for a hearing at the U.S. District Court for the Western District of Texas, he begins by feeding the relevant case documents into an AI tool. Within seconds, he has a timeline of the case and a summary of the arguments each party has made. His clerks used to spend an hour on this.
"It does this instantly," Rodriguez told the Washington Post.
Rodriguez is not an outlier. A study published this week by Northwestern University, surveying 112 federal judges, found that more than 60 percent had used at least one AI tool in their judicial work. Around 22 percent use AI daily or weekly. The Los Angeles County Superior Court announced a pilot programme with Learned Hand -- a legal AI startup named after a federal judge celebrated for the precision of his reasoning and the clarity of his written opinions; he died in 1961 -- in March. Thomson Reuters and LexisNexis have contracts to supply AI tools to the federal judiciary. The Michigan Supreme Court is already running Learned Hand's system, as are trial courts in ten states.
Artificial intelligence is not approaching the courtroom. It is already there, reviewing briefs, drafting decisions, and identifying weaknesses in legal arguments. The question before us is not whether this will happen. It is what it means that it already has.
The Case That Is Being Made
The argument for AI in the judiciary is straightforward.
Courts are overwhelmed. Dockets are long. Judges in many jurisdictions operate without adequate research staff. A tool that can reduce the time required to process a summary judgment motion -- sorting through hundreds of pages of deposition transcripts, identifying the relevant testimony, flagging the arguments that require attention -- is not a luxury. It is an efficiency intervention in a system that needs one.
The equal treatment argument is made less often but matters more. Human judges are subject to fatigue, to implicit bias, to the quality of the legal representation before them. A defendant with an overworked public defender and a defendant with a well-resourced legal team do not receive identical treatment in practice, whatever the law says in theory. AI tools, applied consistently, do not vary their analysis based on how well the brief was written or how expensive the suit. The defendants in cases processed through the LA County pilot programme were not consulted on the software procurement decision.
A third argument is made less often, yet the present moment makes it newly relevant. Courts are increasingly asked to evaluate evidence generated by AI: synthetic media, deepfake video, AI-assisted forensic analysis. The judges best equipped to assess the reliability of AI-generated evidence are those who understand, from daily practice, what AI tools can and cannot do. Familiarity is not a conflict of interest. It is a qualification.
A Note on Optimisation
If judges use AI to evaluate legal arguments, lawyers will optimise their arguments for AI evaluation. This is not a prediction. It is a description of how optimisation works wherever algorithms make consequential decisions: in search rankings, in hiring software, in content moderation, in credit scoring. The argument that scores well with the system is not necessarily the argument that is true. It is the argument the system was trained to find persuasive.
A brief written to persuade a human judge and a brief written to perform well with a LexisNexis AI assessment are different documents. They may reach the same conclusions. They will not use the same language, the same structure, or the same selection of precedents. Lawyers who understand this will adapt. Lawyers who do not will be at a disadvantage that has nothing to do with the merits of their case.
None of the 112 judges in the Northwestern study were asked whether they could tell the difference.
The Risks Are Real and Known
The risks are real and should not be minimised.
In the past year, hallucinations appeared in the written opinions of two sitting federal judges. Judge Henry T. Wingate of the Southern District of Mississippi and Judge Julien Xavier Neals of the District of New Jersey both published documents containing citations to cases that do not exist, fabricated descriptions of plaintiff claims, and false attributed quotes. Both cases involved staff who had used AI tools without adequate verification. Both judges corrected the record. Judge Neals subsequently banned generative AI from his office entirely. He is now the only judge in the study to have done so. No one is currently measuring whether his decisions are better or worse as a result. The question has not been asked.
A 2024 Stanford University study found that the legal AI tools offered by LexisNexis and Thomson Reuters, though more reliable than general-purpose chatbots, still produced errors in 17 to 33 percent of queries. Applied to the volume of federal cases processed annually, this represents a substantial number of incorrect legal conclusions delivered with high confidence to the people responsible for deciding them. The courts describe this as a manageable risk. The people whose cases are being decided were not asked to characterise it.
These are not arguments against adoption. They are arguments for rigorous verification standards alongside the tools themselves. The judges using AI most carefully are also the most insistent on this distinction. "I don't rely on it for the decision," Judge Christopher Patterson of Florida's 14th Judicial Circuit told the Post. "Don't let it replace your judgment with the judgment of the tool."
Rodriguez describes AI as "an extra pair of eyes" -- useful for preparation, subject to review, not authoritative on its own.
The Question That Remains
Eric Posner, professor of law at the University of Chicago, has argued that judges "simply cannot play with a technology that is not fully understood and is known to hallucinate." It is a serious objection. It is also, increasingly, a description of a ship that has already left the harbour.
Consider the full arc of a case processed through the current system. The plaintiff's brief was drafted with AI assistance. The defendant's brief was drafted with AI assistance. The judge used AI to evaluate both. The judge used AI to draft the ruling. The ruling entered the legal databases. The legal databases will train the next version of the AI tools that lawyers use to draft briefs, that judges use to evaluate arguments, and that courts use to draft rulings. The humans were present at every stage. They described the results as efficient.
This is the house that Jack built.
The 112 judges in the Northwestern study were not asked whether AI should be used in courts. They were asked how they were already using it. The answers covered the full arc of judicial work: from the first review of a new filing to the drafting of a final order. The tools are in the workflow. The workflow is producing decisions. The decisions are the law.
What remains to be determined is not whether AI will shape the outcomes of legal proceedings. It is under what conditions, by what standards, and who decides what it is trained to believe.
The Prompt asked a legal AI research tool to summarise the Northwestern University study cited in this article. It returned the correct figures and cited two supporting cases. One of the cases does not exist. The other was correctly cited. The tool described its confidence as high.
Sources: Northwestern University study, April 2026. Washington Post, Daniel Wu, April 3, 2026 (via Merkur.de). Stanford University study on legal AI tools, 2024.