Someone Has to Check an AI Draft Before Anyone Relies on It

In August, a federal bankruptcy judge in Houston asked a lawyer how 32 flagged citations had found their way into his filing. The first was Brasher v. Stewart, which the lawyer’s June response to a motion to dismiss presented as a 1985 Dallas appeals decision on bad faith. The court found that it “does not exist as cited.” Its volume and page number lead to a different 1985 case, and the quoted language does not appear there either, according to the memorandum opinion Chief Bankruptcy Judge Eduardo V. Rodriguez signed on November 4.^[1] His table marks 26 of the 32 as misrepresentations of authority; the other six cite four nonexistent cases, two of them twice. At the hearing, the lawyer said his client, who is not a lawyer, had provided what she thought was correct, and “I accepted it.” He said he had checked some of the cases. She testified that she had used generative AI and, asked whether she had verified the cases, said she had no access to the major paid legal databases at the time.

The response looked like legal research, and much of it was not. AI has made professional-looking work nearly free to produce. That does not make it useful. When nobody checks such work before it goes out, the checking falls to whoever receives it. In Houston it fell first to a defendant’s lawyers, who flagged the citations in a July reply,^[1] and then to the judge. Where accuracy matters, someone has to verify the claims and the cases before anyone relies on the work, and to ask whether it does its job. A document can be accurate and still answer no useful question. When machines produce work faster than people can evaluate it, who is responsible for quality control? Courts have answered for filings. The Southern District of Texas, the federal judicial district that includes Houston, said in a general order in May^[2] that any lawyer or self-represented litigant who signs a filing “will be held responsible for the contents of that filing under Rule 11, regardless of whether generative artificial intelligence drafted any portion of that filing.” Rule 11^[3] makes presenting a paper to the court a certification, to the best of the signer’s knowledge after “an inquiry reasonable under the circumstances,” that its legal contentions are warranted by existing law or by a nonfrivolous argument for changing it.

Decisions like this one now number in the hundreds. Damien Charlotin^[4] maintains a public database of them, limited to those in which a court or tribunal found, or implied, that a party relied on hallucinated material; allegations alone do not qualify. On December 1 it listed 620 decisions worldwide, 416 of them in the United States. The party using AI was a lawyer in 254 and a self-represented litigant in 348. The database’s file archived on November 18,^[5] which includes the Houston decision, holds 562 decisions dated through November 14: 13 from 2023, 50 from 2024 and 499 from 2025, including 89 in October alone. The database says it does not track the “necessarily wider” universe of fake citations, so these counts are a floor, and part of the climb may reflect more people looking.

Decisions in the AI Hallucination Cases database, cumulative, 2023 to 2025

Sanctions orders document what unchecked citations cost. They cannot show how much the sanctions deter. In June 2023, a federal judge in Manhattan fined two lawyers and their firm $5,000^[6] for submitting opinions an AI tool had invented and then standing by them, a sum he judged “sufficient but not more than necessary” to deter. In July of this year, a federal judge in Alabama publicly reprimanded three lawyers, removed them from the case and referred the matter to the state bar,^[7] writing, “If fines and public embarrassment were effective deterrents, there would not be so many cases to cite.” In September, a California appeals court sanctioned a lawyer $10,000^[8] and published its opinion as a warning that no filing should contain citations the responsible attorney “has not personally read and verified.” The Alabama judge wanted a sanction strong enough “to make this misuse of AI unprofitable.” That is the right test, and a hard one to pass. If a shortcut lets a lawyer take on more clients, it pays on every filing, while a penalty lands only on the filings someone catches, and only after a judge or opposing counsel has done the checking the signer skipped.

Offices face the same transfer without a docket. In a Harvard Business Review article published September 22, Kate Niederhoffer^[9] of BetterUp Labs, Jeffrey T. Hancock of Stanford and four co-authors named it workslop: AI-generated work that passes for good work but lacks the substance to move a task forward. In their online survey of 1,150 full-time U.S. desk workers,^[10] conducted in September, 40% said they had received some in the previous month, and those who had encountered it estimated, on average, that 15.4% of what they receive at work qualifies. They reported spending an average of one hour and 56 minutes^[11] dealing with each instance. From those time estimates and self-reported salaries, the authors estimate the cost at $186 per affected employee per month, and at more than $9 million a year for a 10,000-person employer if 41% of its workers are affected. These are self-reports from one survey, and the dollar figures are only as good as the recall behind them. Workslop, the authors write,^[9] “transfers the effort from creator to receiver.”

Recipients notice, and many think less of the sender. Asked how it feels to receive workslop, 53% reported being annoyed, 38% confused and 22% offended.^[11] Forty-two percent saw colleagues who sent it as less trustworthy, and 37% as less intelligent.^[12] According to the HBR article,^[9] 34% of people who receive workslop tell teammates or managers about it, and 32% of those who have received it say they are less likely to want to work with the sender again. The authors published these percentages from the same survey without saying how many people answered each question; only the last two state a base, people who had received workslop. We are trading short-term convenience for long-term credibility.

Reactions to AI-generated workslop among U.S. desk workers, 2025

The strongest objection is that slop is a phase the tools will outgrow. On this view, each generation of tools will make fewer errors, filters will learn to catch the rest, and people will learn which outputs to trust. Carelessness already carries a price: judges sanction the worst filings, and colleagues think less of those who send workslop. Requiring a full check of every AI-assisted draft would spend the time the tools save on mistakes the next version may not make, and it would slow the experimentation that teaches people to use the tools well. Nothing needs to change, the argument concludes, because the problem will shrink as the technology matures.

I share part of that hope. We have adapted to new media before, and this may be an awkward teenage phase before the tools become genuinely useful. But this transition is moving faster than earlier ones. The tracker’s November 18 file^[5] holds nearly seven times as many decisions dated in October 2025 alone as in all of 2023. I lean toward the view that abundant low-quality content is the new normal, because anyone with an internet connection can generate as much as they want, and no earlier machine turned out professional-looking bad content at infinite scale. Filtering systems will have to get better, and I expect them to. Until they do, sanctions and lost trust arrive only after someone else has done the checking. Adapting does not have to wait for them. The Fifth Circuit, the federal appeals court covering Texas,^[13] declined in June 2024^[14] to write a special rule on AI in briefs. Filings must be “carefully checked for truthfulness and accuracy as the rules already require,” the court said, and “‘I used AI’ will not be an excuse for an otherwise sanctionable offense.”

Access is the harder case. In more than half of the tracker’s decisions,^[4] the party using AI was a self-represented litigant, and for someone who cannot afford a lawyer, a chatbot can help turn a grievance into language a judge can follow. The Houston client testified^[1] that she lacked access to the paid databases. Assisted drafting helps people like her, and demanding that an expert redo every AI-assisted filing would erase the advantage. But an invented case can land the user in more trouble than expected, and the correction can come too late. In a November 18 opinion,^[15] the judge dismissed her complaint with prejudice on grounds that did not involve the citations. He also denied as untimely her request to amend that June response to “clarify and correct any citation or errors,” finding that the inaccurate citations had been “entirely in Plaintiff’s control.” Proportionate review asks less than re-performance: confirm that each cited case exists and says what the filing claims before it reaches a court. A draft marked unverified gives its reader an honest task. An unchecked filing presented as finished gives the court and the other side a hidden one.

Nobody had to research Brasher v. Stewart to cite it. There was nothing to find. Untangling the 32 flagged items took a reply brief, a show-cause order, a hearing and a 20-page opinion,^[1] which found that the lawyer “failed to ensure that the cases were real” before he signed. On November 4, the judge ordered him to reimburse one defendant’s lawyers for their reasonable and necessary fees and costs and to obtain six hours of continuing education on generative AI. The judge also wrote that the lawyer would be referred to the chief judge who signed the May order^[2] and to the State Bar of Texas’s chief disciplinary counsel for possible disciplinary action. The opinion calls confirming the accuracy of cited case law “a basic, routine matter.” It would have been the cheapest step in the whole affair.