Oozing Into the Courts: AI Sludge Embarrasses Attorneys (Again)

The AI absurdity from the legal trenches continues, where the citations are fake, the sanctions are real, and the attorneys are still figuring out how to turn the thing off.

The Government Cites a Ghost Case. A Judge Notices.

Let’s start with the federal government, which apparently has access to every legal database known to humankind and still managed to cite a Sixth Circuit case that does not exist.

In an immigration habeas case out of the Western District of Michigan, government counsel cited Taylor v. Hott, 724 F. App’x 387, 392 (6th Cir. 2018), for the proposition that district courts lack jurisdiction to review an immigration judge’s bond denial.^[1] Sounds authoritative. Sounds real. It is neither.

The judge looked for Taylor v. Hott and instead found Atkins v. CGI Technologies & Solutions, a case about commercial arbitration that has absolutely nothing to do with immigration bond determinations. The judge “was unable to identify a Sixth Circuit case with the caption Taylor v. Hott, or any federal case containing the quoted language.”^[2] The judge’s conclusion: the citation “was likely produced by generative artificial intelligence.”^[3]

The court declined to impose sanctions—this time—but delivered a message that should be tattooed on the forearm of every attorney who has ever typed a legal question into a chatbot:

It should be obvious that any attorney who uses AI must scrupulously review its work product to ensure that the cited cases exist and that the citations accurately and fairly represent the underlying case law. The duty of candor towards this tribunal demands no less.^[4]

The government’s response to this public embarrassment? Silence. Checks out.

The Seventh Circuit’s Astonishing Tolerance for Astonishing Errors

Over in the Seventh Circuit, the court was confronted with what a dissenting judge called “an astonishing number of erroneous and even hallucinated citations” in the appellee’s brief in a bankruptcy interest-rate dispute.^[5] The majority’s response? A gentle reminder to “exercise care and diligence.”^[6]

In dissent, Judge Hamilton was having none of it. He noted that the sloppy errors by the creditor’s attorney “made this court’s work more difficult than it should have been” and argued that “[s]anctions like those we have imposed in other recent cases involving hallucinated citations from generative artificial intelligence would be appropriate here.”^[7]

The majority declined. The attorney got a stern look and a wagging finger.

In sum, using phantom quotations, false citations, and misattributed authority can result in a strongly worded paragraph that the attorney can frame and display in their office. The deterrence value of this outcome is left as an exercise for the reader.

Texas: Where the Sanctions Are Actually Real

Meanwhile, in the Southern District of Texas, a bankruptcy judge was not in a forgiving mood.

An attorney filed motions to quash discovery requests that contained, per the court’s ruling, “fabricated quotations, non-existent citations, and invented legal authority apparently generated by artificial intelligence and submitted [the filing] to the Court without verification.”^[8] In the ruling, the judge chronicled the AI errors (covering seven pages).^[9] The attorney’s explanation? He used a prior template and did not verify the citations.^[10]

The judge found violations of Federal Rule 11(b) applied through the bankruptcy rules, and imposed sanctions accordingly.^[11] The result: nearly $30,000 in attorney fees awarded against the attorney and his clients, plus a mandatory six hours of CLE specifically focused on the use of Gen AI in the courts.^[12]

Six hours of CLE. Roughly, the time it takes to watch the extended cut of a legal thriller. One hopes the irony is not lost on the attorney.

Your Expert’s AI Prompts Are Now Discoverable. Surprise.

In what may be the most consequential development of the week for litigators who thought their experts’ AI use was safely behind the curtain, a Connecticut magistrate judge (in a docket entry) ordered the plaintiff to produce the generative AI prompts used by its expert in preparing her report.^[13]

The court’s reasoning was straightforward: an expert's methodology is discoverable under Rule 26; the AI prompts that shaped which documents the expert ever reviewed are part of that methodology.^[14] The court treated AI prompts “as no different in principle from the formulas, code, or other methodological steps an expert is routinely required to disclose.”^[15]

The order is currently stayed pending objection, so the last word has not been written.^[16] However, the logic is intuitive enough that other courts are likely to follow. If you are sponsoring an AI-assisted expert and you have not yet had a conversation about prompt preservation, you are already behind.

$1.5 Billion for the Authors. $101 Million for the Attorneys. Math Checks Out.

In cheerier news (unless you are an AI tech bro), the Northern District of California granted final approval of the $1.5 billion class action settlement in Bartz v. Anthropic, the copyright case brought by authors whose books were allegedly used to train Anthropic’s AI models without authorization.^[17]

Class members will receive approximately $3,000 per work.^[18] Class counsel will receive $101,561,111.^[19] The court applied the lodestar method rather than the percentage-of-recovery method, noting that a 25 percent fee on a $1.5 billion fund “would yield windfall profits for class counsel.”^[20] The resulting multiplier of 3.75 was deemed reasonable.^[21]

The settlement is described as “the largest copyright class action settlement in history.”^[22] Anthropic purported that the pirated datasets were never actually used in its commercially released models.^[23] Thus, the books trained nothing, the authors got $3,000 each, and the attorneys got nine figures. The AI economy is working exactly as designed.

The Benchslap Taxonomy Now Includes AI Hallucinations

In Knife Rights v. Bonta, Judge VanDyke wrote a colorful dissent about the Ninth Circuit’s Second Amendment jurisprudence, while pausing to enumerate the categories of attorney misconduct that earn a judicial “benchslap.”^[24] The list includes the “ostrich-like tactic” of ignoring adverse precedent, asking to suspend a murder trial to attend an Ernest Hemingway look-alike contest, and “citing A.I.-hallucinated cases.”^[25]

In the context of verifying case citations, the judge recommended: “if you’re a lawyer and you find yourself citing a Ninth Circuit decision that seems to vindicate the Second Amendment right, you should definitely double- (and triple-) check that it’s not an A.I.-hallucinated case.”^[26]

The fact that AI hallucination has now achieved canonical status in a federal appellate court’s taxonomy of sanctionable conduct tells you everything you need to know about where we are in 2026.

The Wrapper Debate: What Are Law Firms Actually Buying?

The key question is what’s in the wrapper?^[27] On the business side of legal AI, the debate over “thin wrappers” versus “commercial-grade wrappers” around LLMs continues.^[28] The argument: anyone can slap a user interface on GPT-4 and call it a legal research tool, but what law firms actually need is security compliance, workflow integration, accountability infrastructure, and institutional trust.

The Hershey bar analogy is apt: you’re not just buying chocolate, you’re buying the brand, the hygiene standards, the regulatory compliance, and the assurance that if something goes wrong, someone will answer the phone.^[29] Since Mata, an assurance that “someone will answer the phone when the citations turn out to be fictional” seems like a reasonable minimum requirement for any legal AI product.

Sam Altman Wants You to Own a Piece of the Robot That Took Your Job

Finally, in the category of policy proposals that are either visionary or deeply cynical, depending on your mood: OpenAI CEO Sam Altman has reportedly proposed that AI companies donate shares to a public wealth fund, distributing returns to citizens as a kind of AI dividend.^[30]

The math, even under bullish assumptions, suggests each American might receive roughly $3,600 per year in today’s dollars.^[31] That is, coincidentally, approximately what a class member in Bartz v. Anthropic received per book. Whether owning a fractional share of the machine that disrupted your industry constitutes meaningful compensation is a philosophical question best left to economists.

The Trump administration, which has already taken equity stakes in 30 companies worth $26.7 billion, finds the idea appealing.^[32] Critics note that government equity stakes in AI companies create obvious conflicts of interest and do nothing to address actual job displacement.^[33]

But it would, as the President reportedly mused, “be a beautiful thing.”^[34] In 2026, beautiful things and functional things are not always the same.

The article's underlying lesson is that the legal profession is moving past the question of whether AI will be used and into the more consequential questions of how, by whom, and with what safeguards. The biggest developments are not the hallucinated citations themselves, which are becoming almost routine, but the emerging accountability mechanisms: sanctions, discovery of AI prompts, scrutiny of expert methodologies, and billion-dollar litigation over training data.

I'd lean into that idea in the conclusion:

Conclusion

Three years after Mata, the legal profession is still sifting through the AI sludge oozing into court filings, expert reports, and litigation strategy. The fabricated citations have not disappeared, but the legal system is slowly developing antibodies. Judges are becoming more sophisticated, sanctions are becoming more common, and attorneys are learning that “the chatbot made me do it (again)” is not a recognized defense.

The good news is that AI will almost certainly make attorneys more efficient. Better tools, more transparent workflows, and increasing scrutiny of AI-generated work product may eventually allow attorneys to spend less time on routine tasks and more time on actual legal analysis. The disputes we are seeing today may ultimately produce clearer rules and more reliable systems.

The bad news is that the volume of AI-generated errors is likely to increase before it decreases. As more attorneys, experts, and parties integrate AI into their daily work, courts will continue to encounter fabricated authorities, questionable quotations, and filings that were reviewed with something less than the diligence required by Rule 11. The slop docket is unlikely to become less busy anytime soon.

Of course, there is the much, much worse possibility. If courts increasingly treat prompts, workflows, and AI-assisted reasoning as discoverable components of legal and expert methodology, the next wave of litigation may not concern hallucinated citations at all. It may concern what happened inside the black box: what questions were asked, what information was omitted, what bias was introduced, and who bears responsibility when an AI-influenced decision causes harm. Compared with that future, today’s ghost cases and fake citations may come to look like the easy problems.

Until next time, double-check your citations, preserve your prompts, and hope the algorithm does not block this blog!

^[1] Daghra v. Hinkley, No. 26-cv-01429, slip op. at 2 (W.D. Mich. July 16, 2026).

^[2] Id.

^[3] Id.

^[4] Id.

^[5] In re Romero, No. 25-2021, slip op. at 18 (7th Cir. July 16, 2026) (Hamilton, J., dissenting).

^[6] Id. at 11.

^[7] Id. at 18 (Hamilton, J., dissenting).

^[8] In re Rosslyn2016 LLC, No. 25-bk-34507, slip op. at 1 (Bankr. S.D. Tex. July 14, 2026).

^[9] Id. at 20-26.

^[10] Id. at 9-10.

^[11] Id. at 18.

^[12] Id. at 29-30 (“The Trustee’s counsel submitted a time entry reflecting 44.6 hours of attorney and paralegal time, totaling $29,877.00 in fees.”).

^[13] Conservation Law Foundation v. Shell Oil Co., No. 21-cv-00933 (D. Conn. June 3, 2026).

^[14] Id.

^[15] Id.

^[16] Conservation Law Foundation, No. 21-cv-00933 (D. Conn. May 18, 2026).

^[17] Bartz v. Anthropic, No. 24-cv-05417, slip op. 21-23 (N.D. Cal. July 20, 2026).

^[18] Id. at 4, 17.

^[19] Id. at 22.

^[20] Id. at 15.

^[21] Id. at 18.

^[22] Id. at 17.

^[23] Id. at 11-12.

^[24] Knife Rights, Inc. v. Bonta, No. 24-5536, slip op. at 30-31 (9th Cir. July 16, 2026) (VanDyke, J., dissenting).

^[25] Id.

^[26] Id. at 31 n.9.

^[27] See generally Ken Crutchfield, What’s In A Wrapper? Comparing A Hershey Bar To LegalTech AI, LawSites (July 17, 2026).

^[28] Id.

^[29] Id.

^[30][30] See generally Sebastian Mallaby, This Is a Terrible Way to Address the Risk of A.I. Of Course Trump Loves It., N.Y. Times (July 21, 2026).

^[31] Id.

^[32] Id.

^[33] Id.

^[34] Id.