Most legal writers—call them Type 1—don’t care much about the punctilios of Bluebook practice. Those with more bibliographic conscientiousness—Type 2—have long noticed their “sloppiness.” There is, after all, an elaborate code governing citation: a catalogue of mandatory abbreviations in Table T6 and a hierarchy of signals (no signal, see also, cf.), each carrying a distinct and calibrated claim about what the cited authority actually does. Yet outside law reviews and the occasional fastidious chamber, much of this apparatus is ignored.

Then came the hallucinations, and with them a sudden, unwelcome spotlight on citation practice. Artificial-intelligence systems have shown a disreputable tendency to invent authorities—cases that don’t merely fail to support a proposition but don’t exist at all. This summer alone, several federal circuits confronted the problem in published opinions. The possibility of filing a brief with phantom authority now unsettles even experienced lawyers. A simple Westlaw search—“artificial” within the same paragraph as “hallucinate!”—returns a growing body of decisions that read like warnings: 475 of them as of late July, and steadily climbing.

Whether that fear is actually changing behavior, rather than simply generating headlines, is harder to know than it might seem. There’s no bar survey tracking how carefully lawyers now check their citations, and firms have little incentive to publicize a shift in habits prompted by embarrassment. What can be said with more confidence is that the incentive structure has changed unmistakably, and incentives tend to drive behavior far more reliably than exhortation ever does. Whether the incentive has yet translated into a broad change in practice, or is still working its way through the profession one sanctions order at a time, is the more honest question—and one this essay means to raise rather than settle.

Judge Guido Calabresi of the Second Circuit once described to me the first thing his law clerks do with every brief, and it isn’t what most lawyers might expect. They check the citations—not for substance, not yet, but for reliability. Each authority gets a kind of internal grade: is it quoted accurately, and does it stand for what the lawyers claim? Only after that threshold assessment does the judge turn to reading the arguments. He’s measuring trustworthiness before engaging with reasoning, and that order tells you something fundamental about how judges actually read.

Most lawyers don’t realize that this triage occurs, or at least they don’t fully internalize its consequences. They imagine judges reading their prose with sustained, good-faith attention, weighing each argument on its merits. In reality, many judges are quietly grading the lawyer before they ever assess the merits of a case. A brief marred by misquotations or dubious citations earns a skepticism that no later elegance can fully overcome. In retrospect, Calabresi’s practice seems like a clear-eyed response to a persistent professional failing—one that AI has simply made impossible to ignore.

Large language models, when asked for authority, will sometimes fabricate it wholesale—producing plausible case names, credible-sounding holdings and confident explanations that collapse under the slightest scrutiny. The results have been predictable and increasingly public: sanctions, withdrawn filings, and reputational damage that spreads well beyond the lawyers immediately involved.

Yet the machine’s defect has an unexpected virtue, at least in its present, relatively crude form. Its errors are so stark, so easily exposed, that they force a discipline that lawyers long managed to evade. It’s one thing to shade a holding or rely too heavily on a headnote; it’s another to cite a case that simply doesn’t exist. The latter cannot be rationalized or explained away. It demands verification at the most basic level. And once lawyers begin checking for existence, it’s not much further to checking for accuracy, context, and scope.

The real shift, though, lies in how these mistakes are perceived. A decade ago, a bad citation suggested carelessness—regrettable, but hardly career-defining. Today it suggests something more troubling: that the lawyer may have relied on a machine’s output without exercising independent judgment. Carelessness implies haste; uncritical reliance on fabricated authority implies abdication. The profession has begun to recognize the distinction, and it’s reacting accordingly, often with severity.

The courts, for their part, have provided a steady stream of object lessons. The early case Mata v. Avianca (S.D.N.Y. 2023) may become the canonical example: a brief supported by wholly invented cases, followed by sanctions and enduring notoriety. The court deadpanned: “A fake opinion is not ‘existing law.” No kidding. It is also, in retrospect, the easy case: A wholesale fabrication is far easy to catch.

The pattern hasn’t remained confined to a single courtroom or jurisdiction. In Lnu v. Blanche (June 3, 2026), the Ninth Circuit encountered briefs riddled with nonexistent authorities and misattributed quotations, compounded by counsel’s repeated denials that AI had been used at all—denials the court found not merely mistaken but knowing or reckless. The court’s message was unmistakable: fabrication is serious misconduct, but a lack of candor about its source compounds the offense. Notably, the court went further still, holding that simply swapping a hallucinated citation for a real one, without disclosing that the earlier citation had been invented, doesn’t cure the original false statement to the tribunal. If the machine misleads you, the expectation is not just disclosure eventually, but disclosure of the right thing.

Other circuit courts have expressed muted exasperation. In Lichfield v. Kubler (July 27, 2026), the Tenth Circuit concluded its opinion with a paragraph that probably mortified counsel: “We appreciate that counsel filed a Notice of Errata the day after oral argument. But it should not have taken questioning at argument to elicit such a response. Submission of a brief containing fabricated or hallucinated citations is a divergence from the standards expected of attorneys appearing before this court. Any such error should be corrected explicitly and without prompting.”

The personal dimension matters more than anything else. Courts increasingly assess sanctions against the lawyer who signed the filing, not against the client. Financial penalties have a way of focusing attention, especially when they come paired with reputational consequences that can’t be insured against or easily forgotten. A lawyer who has paid such a sanction is unlikely to treat citation-checking as a dispensable step again.

For decades, judges, bar associations, and legal-writing authorities have urged lawyers to double-check their citations before filing. The advice was sound, widely disseminated, and often ignored. Verification takes time, and time is always in short supply. The marginal cost of checking a citation often seemed to outweigh the marginal benefit, particularly when the consequences of error were relatively modest. AI has altered the calculus with surprising speed, at least for the crudest form of the problem: outright invention.

The current wave of sanctions is largely a story about obvious hallucinations—cases that don’t exist at all, quotations that appear nowhere in the source. As legal-specific AI tools improve and lawyers grow more practiced at smoothing over a model’s output, the failures most likely to survive into a filed brief will revert to the older forms: a real case stretched past its holding, a quotation trimmed just enough to change its meaning. Those are exactly the failures Type 1 lawyers tolerated long before any machine was involved. A discipline imposed by the starkness of an error may not automatically extend to errors that stop being stark. Whether the current reckoning produces a lasting culture of verification, or simply drives the old shoddiness further underground, isn’t yet known.

The effects extend beyond individual habits to the structure of legal practice itself. Law firms have long relied on layered delegation: responsibility for checking citations trickled down to junior lawyers and staff. That model depends on trust at each link, and AI introduces uncertainty into all of them. Any participant in the chain may have consulted a chatbot, and the resulting output may carry errors that aren’t immediately apparent. The rational response isn’t to add more layers, but to reintroduce verification at the top.

None of this makes hallucinations desirable in any straightforward sense. They remain a defect, sometimes a costly and embarrassing one, and nobody would design a system to behave this way if given the choice. But defects can have disciplinary effects that virtue alone fails to produce, at least for a while. Speed cameras don’t make drivers more principled; they make them more careful by attaching a tangible cost to inattention—for exactly as long as the cameras keep catching people. Whether AI hallucinations are playing a similar role in legal writing, or merely a temporary one, may depend on how long the errors stay easy to catch.

If the more optimistic reading holds, the profession may extract a real benefit from an otherwise unwelcome development: lawyers reading the cases they cite, quoting them with greater precision, and taking personal responsibility for the accuracy of what they file. But the honest version of that hope comes with a caveat that the courts themselves have not yet had occasion to test: a discipline built on the fear of being caught is only as durable as the odds of getting caught.

Bryan A. Garner. (Photo by Karolyne H.C. Garner)

Bryan A. Garner is the author of The Winning Brief, Garner’s Modern English Usage and Legal Writing in Plain English.