There is a specific kind of error that generative AI produces that no human professional would ever make on their own. A tired lawyer might misremember a holding. A rushed associate might cite the wrong page. But no human lawyer, however overworked, sits down and invents an entire court case out of nothing, complete with a plausible name, a fabricated docket number, and a quotation that was never written. That is a distinctly machine failure, and in 2026 it has been showing up in courtrooms across the country with enough regularity that judges have stopped being surprised and started reaching for the sanctions playbook.
The technical term for it is a hallucination. When a large language model like ChatGPT is asked for legal authority, it does not look anything up. It predicts the most statistically likely sequence of words that resembles a citation, and if no real case fits, it confidently produces one that does not exist. To a lawyer who does not check, the output is indistinguishable from real research. To the opposing counsel who does check, and to the judge who reads the brief afterward, it is a fabrication filed under penalty of professional responsibility. The cases stacking up this spring show what happens when that gap closes in open court.
A Federal Appeals Court Draws The Line At $2,500
In a February order, the U.S. Court of Appeals for the Fifth Circuit ordered a Dallas attorney to pay $2,500 after she submitted a brief riddled with citations to cases that did not exist. The court did not treat this as a simple slip. Its show-cause order enumerated 16 instances of fabricated quotations and 5 additional serious misrepresentations of law or fact, a level of contamination that suggests the brief was not merely assisted by AI but substantially generated by it.
What pushed the panel from concern to sanction was not the original error. It was the response. The attorney first denied using AI at all, then shifted to saying she had only used it to clean up her prose, and characterized the fabricated citations as having come from publicly available sources. The court was blunt about the cost of that approach, finding that when confronted with a serious ethical misstep, she had misled, evaded, and violated her duties as an officer of the court. The panel made clear she could have avoided the penalty entirely had she simply come clean. That is the recurring lesson of the year, the hallucination is the misstep, but the cover-up is what gets punished hardest.
The $110,000 Penalty That Killed A Lawsuit
If the Fifth Circuit fine reads as a warning shot, the Oregon case reads as live fire. In April, a U.S. magistrate judge in Oregon imposed a total of $110,000 in fines and attorneys' fees against two lawyers in a lawsuit over control of a family winery and tasting room. The plaintiff, an Oregon woman, had sued her two brothers, alleging elder abuse and wrongful enrichment, and the suit sought $12 million in connection with the family business.
The filings told a different story than the pleadings. Across three court documents, the judge found 23 fabricated legal citations and 8 false quotations, all generated by artificial intelligence. The lead attorney, who was not licensed to practice in Oregon and relied on a local lawyer for a procedural role, had attempted what the court characterized as a cover-up when the bogus material was uncovered. The judge's response was not limited to money. He dismissed the underlying lawsuit with prejudice, meaning it cannot be refiled. A client's $12 million claim was extinguished not because it lacked merit on the facts, but because her lawyers handed the court a brief built on cases that were never decided. The financial penalty was reported as the largest of its kind handed down by an Oregon federal judge, dwarfing the state appellate court's previous high-water mark of $10,000.
Even Wall Street's Elite Are Not Immune
The comforting theory about AI hallucinations in legal filings was that they were a problem of solo practitioners and small firms without the resources to verify, lawyers cutting corners under pressure. April demolished that theory. Sullivan & Cromwell, one of the most prestigious law firms on Wall Street, apologized to a federal judge for submitting a filing in a Manhattan bankruptcy case that contained inaccurate citations and other errors generated by artificial intelligence.
In a letter dated April 18, the co-head of the firm's global restructuring group acknowledged that the errors included AI hallucinations, instances in which the model invents case citations, misquotes the law, or generates legal sources that do not exist. The firm attached a ledger of the problems that ran roughly three pages and totaled around three dozen errors, including misquotations of the U.S. bankruptcy code, wrong case citations, and inaccurate summaries of other cases. The mistakes were caught not by the firm's own review but by opposing counsel at another firm. The letter did not disclose which AI program produced the filing. When a firm of that caliber, with that level of institutional review, still files three dozen AI errors in a single bankruptcy motion, the lesson is unambiguous, this is not a competence gap among the under-resourced. It is a workflow failure that scales with how much you trust the tool.
When The Sanction Is Education, Not A Fine
Not every court reaches for the checkbook, and the Maine case shows a different model emerging. In a lawsuit against a Bath boarding school, in which a former attendee alleges that students were subjected to forced labor and abuse over a span of decades, a claim the school denies, a U.S. district judge sanctioned the plaintiff's attorney for using artificial intelligence in a filing that produced inaccurate and fabricated legal citations.
The school's lawyers had flagged the problem in a December filing, pointing out that the plaintiff's November brief referenced two cases that could not be located, mischaracterized a third, and grossly misquoted a Maine statute. The attorney took responsibility in an April declaration, writing that family pressures had kept her from her usual line-by-line verification process. The judge's sanction was notably forward-looking rather than punitive. The attorney was ordered to complete a legal education course on AI in the legal field and to build procedures at her firm to prevent the same failure from recurring. She was not fined, and she remains on the case, which will continue. It is a sanction designed to fix behavior rather than simply punish it, and it may become the template for first offenses where the lawyer owns the mistake instead of hiding it.
The Pattern These Cases Form
- The hallucination itself is rarely the deepest problem. In the cases that drew the heaviest penalties, the original fabrication was compounded by denial, evasion, or an attempted cover-up, and that is what moved judges from leniency to maximum sanction.
- The exposure is not just professional embarrassment. It ranges from a $2,500 fine to a $110,000 penalty to the outright dismissal with prejudice of a $12 million claim, meaning a client can permanently lose a case because of a tool the client never touched.
- Firm size and prestige offer no protection. An elite Wall Street firm filed roughly three dozen AI errors in one motion, the same category of failure that sanctioned solo and small-firm practitioners elsewhere.
- Opposing counsel is now the de facto verification layer. In multiple cases, the fabricated citations were caught not by the filing lawyer or the court but by the other side, which has every incentive to run every citation to ground.
Why This Keeps Happening, And Why It Will Continue
The uncomfortable truth underneath these stories is that the technology is working exactly as designed and the lawyers are misunderstanding what it is. A chatbot is a fluency engine, not a research database. It is optimized to produce text that reads like authoritative legal writing, and it is extraordinarily good at that surface. The citations look right. The quotations sound judicial. The formatting is impeccable. Everything about the output is engineered to pass a quick glance, which is precisely the failure mode for a busy professional who treats a glance as verification.
The defense that keeps appearing in these cases, that the lawyer was under time pressure or personal strain, is true and beside the point. The tool does not get safer when you are tired. It gets more dangerous, because exhaustion is exactly when a lawyer is most likely to accept a confident-looking citation without pulling the actual case. The judges handing down these sanctions are not technophobes lashing out at progress. They are enforcing the oldest rule in the profession, that the name signed at the bottom of a brief is a certification that someone read and verified what is above it. AI did not change that rule. It just made it far easier to violate without noticing.
The volume of these cases in 2026 suggests the bench has reached a collective decision. The grace period, in which courts treated AI hallucinations as an honest novelty and let lawyers off with a stern warning, is closing. What replaces it is a simple and unforgiving standard, if you file it, you own it, and the chatbot is not a co-author the court will accept. The lawyers learning that lesson the expensive way are not the last ones. They are the early warning. For more on how the same failure mode is playing out beyond the courtroom, see our running coverage of AI hallucinations, the broader pattern of AI lawsuits and legal fallout, and the latest from the ChatGPT Disaster blog.