In 2023, a researcher at HEC Paris began maintaining a public database of cases in which generative AI produced hallucinated content submitted to courts. The database started with a handful of entries. By early 2026, it documented more than 1,200 cases globally—with five to six new cases added every day. The sanctions began as warnings. They have escalated into fines, case dismissals, license suspensions, and bar charges that could result in disbarment. The pace is not slowing.
What follows is a documented record of where professional accountability for unverified AI output stands as of May 2026. It covers legal practitioners and the courts addressing their conduct. The thread running through every case is the same: a professional used AI-generated output, submitted it without independent verification, and was held accountable for what it contained.
The Court Record—Attorney Sanctions
The Regulatory Trajectory
The pattern across every case is consistent. The attorney used an AI tool. The AI produced confident, formatted, plausible-looking output. The attorney submitted it without verifying what it contained. The court, bar, or regulator held the attorney responsible for the content of what they filed—not the tool, not the vendor, not the platform.
Courts have been explicit that this obligation predates AI and does not require specific AI-related rules to enforce. The California Court of Appeal stated it plainly in its first published opinion on AI hallucinations in court filings: "No brief, pleading, motion, or any other paper filed in any court should contain any citations—whether provided by generative AI or any other source—that the attorney responsible for submitting the pleading has not personally read and verified." The court published the opinion specifically as a warning. The same duty applies to every AI-generated work product that enters the litigation workflow—not just citations, but findings, timelines, biographical details, and causation conclusions.
The sanctions are scaling with the volume of cases. What began as warnings and modest fines has moved to case dismissals, five-figure sanctions, license suspensions, and bar charges. The Oregon $110,000 sanction and Nebraska license suspension in April 2026 represent a clear signal from the judiciary and state bar systems that the threshold for serious consequences has lowered significantly.
California is moving toward codifying the obligation. The state legislature advanced a bill in early 2026 requiring attorneys to take reasonable steps to verify the accuracy of AI-generated material, correct any hallucinated output, and remove biased or harmful content—while making clear that these duties apply to work produced on their behalf by others, including AI tools. Legal analysts have suggested the bill could become a template for other states.
The Judge Is Also Using AI
The assumption underlying most attorney AI misconduct cases is that the fabrication will eventually be discovered—in discovery, at oral argument, or when opposing counsel notices something wrong. What has changed in 2026 is when and how that discovery happens. Judges are now using the same class of AI tools to process the filings submitted to them.
A Northwestern University study published in April 2026, co-authored by a sitting federal judge, found that more than 60 percent of federal judges reported using AI tools at least once in their judicial work, with around 22 percent using AI daily or weekly. One Texas federal judge described to MIT Technology Review feeding court filings into an AI tool before every hearing to generate case timelines, identify the claims each party is making, and surface questions for counsel. In at least one case before him, both sides submitted filings referencing fabricated cases—and his AI-assisted review of the filings caught it.
This is not AI detection in the technical sense. Courts are not running filings through hallucination-detection software—those tools remain unreliable and brittle in real-world conditions. What judges are doing is more straightforward and more effective: they are reading the cases cited in the briefs. When a cited case does not say what the filing claims it says—or does not exist at all—that is the detection. AI tools that help judges generate timelines and identify inconsistencies across a large filing accelerate that process. The attorney who submitted unverified AI output is now facing a judge who may have processed that filing through AI before the hearing began.
The practical implication is significant. An attorney who relied on AI-generated content without independent verification may walk into a courtroom where the judge has already identified the problem. The window between submission and detection—which once extended through the life of the case—is narrowing. Courts are not waiting for opposing counsel to catch hallucinations. They are catching them in preparation.
The Medical Record Review Parallel
The same principle that produced these sanctions applies to every AI-generated work product used in litigation—including AI-generated medical record reviews. The tool is different. The output format is different. The professional context is different. The obligation is identical.
An AI platform that reviews 1,424 pages of medical records and produces 4,500 pages of output is generating content that will inform how a case is evaluated, built, and litigated. If that output contains fabricated findings, mischaracterized records, or biographical details unsupported by the source material—and no physician has independently engaged with the underlying records to verify what the AI produced—the attorney who relied on that output to make case decisions has the same exposure as the attorney who filed a brief with fabricated citations.
The tool is not the problem. AI-assisted review, directed and verified by a physician who has independently engaged with the underlying records, is faster and more thorough than linear review alone. The problem is reliance on AI output that no one in the workflow has verified—and the professional exposure that creates for the attorney who acted on it.
The platform does not bear the professional responsibility. The attorney does. Courts are establishing that clearly, case by case, in a growing body of sanctions and disciplinary decisions that will eventually constitute the standard of care for AI tool use in litigation.
It does not end because the output is formatted to look like a professional work product. It does not end because the platform markets itself as a replacement for human review. It does not end because the attorney is under time pressure, the records are voluminous, or the AI is expensive.
Every case documented above involved output that looked authoritative. That is what AI produces. It produces confident-sounding, well-formatted content that carries no internal signal of where it diverged from the source material. The divergence only becomes visible to someone who knows what to look for—a judge, opposing counsel, or a physician—and has independently engaged with the underlying records.
That is what the courts are requiring. That is what the bar associations are enforcing. And that is what the 1,200 documented cases have established as the applicable standard—not just for legal research, but for every AI-generated work product that enters the litigation workflow.