The Footnotes Were the New Part

The Footnote That Wasn't There

In June 2026, the Department of Health and Human Services rewrote the rules of the Teen Pregnancy Prevention Program. The new funding notices required every grantee to teach something called "body literacy" — a framework HHS described as equipping "young people with a clear understanding of how their bodies function in a state of health."^[1]

The notices carried seven footnotes supporting the idea. Five of them did not survive contact with a librarian.

Two were entirely fictional: no such papers exist. Three cited real authors, but attributed their work to journals that never published it; the titles were near-misses of real articles those authors had published somewhere else. And when the plaintiffs tracked down the real articles HHS might have meant, they did not support the claims either. One was a study of sexually transmitted infection rates among adults with multiple partners, cited by HHS for a proposition about "adolescent sexual decision-making." Another reviewed fertility-awareness family planning methods, cited for the claim that girls taught body literacy are likelier to remain abstinent.^[2]

On August 19, 2026, Judge Christopher R. Cooper of the U.S. District Court for the District of Columbia put the observation into a federal opinion:

On the topic of body literacy, the notices (remarkably) reference public health studies that appear either not to exist or not to support the propositions for which they are cited — a hallmark of AI-generated citations.

Hennepin County, Minnesota v. HHS, No. 1:26-cv-02460 (CRC), Mem. Op. at 46

The parenthetical is the judge's own.

The Ledger

The itemization comes from a sworn declaration by Kate Talmor, one of the plaintiffs' attorneys, who checked each citation and reported what she found. Cooper quoted it directly.^[1]

Why the Footnotes Were Needed

Cooper was careful to say what this case is not about, and the distinction is worth keeping in view:

HHS is perfectly entitled to formulate its own views about how to stem teen pregnancy — or even whether it is worth preventing at all — and to pursue policy initiatives consistent with its viewpoint. But it is not at liberty, under the Administrative Procedure Act ("APA"), to impose conditions on grant recipients that Congress did not intend or that are unreasonable or unexplained.

Mem. Op. at 2

An administration may decide that abstinence education is the right approach. What it may not do is claim a factual basis it does not have. And that is where the fabrications sit: not scattered through the document as random sloppiness, but precisely where the law required evidence and none existed.

Congress structured the Teen Pregnancy Prevention Program in two tiers. Tier 1, which takes 75 percent of the money, is reserved by statute for "replicating programs that have been proven effective through rigorous evaluation." Tier 2, the remaining 25 percent, funds research and demonstration projects to "develop, replicate, refine, and test additional models and innovative strategies." As the D.C. Circuit once summarized it: "Tier 2 lets grantees test new programs, and programs that prove effective then become eligible for Tier 1."^[1]

HHS made body literacy a requirement of both tiers. That is where the problem starts. To put an approach in Tier 1, the statute demands proof it already works. And HHS's own funding notice conceded, on page 5, a "near absence of body literacy education standards nationwide."^[1]

Cooper drew the line himself: in prior cycles, the agency "would have been hard-pressed to fund a Tier 1 program that adopts a 'body literacy' approach to preventing teen pregnancy, as there are a 'near absence of body literacy education standards nationwide' — let alone any that have proven effective at reducing teen pregnancy or a related risk behavior."^[1]

So the agency needed a body of research that did not exist. Seven footnotes appeared. Five of them were not real.

The Second Attempt

Cooper opened his conclusion with a nursery-rhyme reprimand:

As the saying goes, "When at first you don't succeed, try, try again." HHS has taken this adage to heart. Last year, a judge on this Court vacated an HHS guidance document that made substantial policy changes to the TPP Program, on the ground that the changes were arbitrary and capricious under the APA. Unfazed by the adverse ruling, HHS issued new NOFOs that repackaged the vacated policy changes and imposed additional grant parameters.

Mem. Op. at 60

The earlier ruling was Judge Beryl Howell's, on October 7, 2025. She struck down a July 2025 HHS "Policy Notice" that had purported to clarify when TPP materials were not "medically accurate" or "age appropriate." Howell found its "vagueness renders the new requirements imposed on TPP grant recipients largely incomprehensible and unworkable, putting in place an opaque 'we-know-it-when-we-see-it' standard for HHS to assess compliance."^[1]

Nine months later, the same policy returned with footnotes attached. The footnotes were the new part.

The Other Direction: Into the Courtroom

The TPP notices were policy documents that happened to be litigated. The reverse has also been happening: government lawyers filing briefs whose authority does not exist.

A federal prosecutor, publicly reprimanded

The clearest case is Fivehouse v. U.S. Department of Defense, an APA challenge to a TRICARE medication coverage change in the Eastern District of North Carolina. Magistrate Judge Robert T. Numbers II opened his April 28, 2026 sanctions order this way:

This case presents a question arising with alarming frequency in courts across the country: what should a court do when an attorney submits a brief containing fabricated legal authority and attributes the errors to the use of generative artificial intelligence? The answer is straightforward. The court should hold the attorney accountable because he is responsible for the authorities he presents to the court, regardless of they came to appear in the brief.

Fivehouse v. DoD, No. 2:25-cv-00041-M-RN (E.D.N.C.), D.E. 129 at 1

The finding against Assistant U.S. Attorney Rudy Renfer was made to the clear-and-convincing standard: he "knowingly submitted a brief containing fabricated quotations and misstatements of case holdings," and those errors "arose because Renfer used a generative AI tool to draft the brief but failed to verify the accuracy of its contents."^[5]

The show-cause order enumerates five separate filings.^[6] Three involved fabricated quotations from real Fourth Circuit decisions. The remaining two are the ones worth pausing on:

A federal prosecutor defending an agency invented language from that agency's own published rule, and did it in two different briefs. The court noted a further problem: when first questioned, Renfer's submission "made no mention of the use of generative AI." He acknowledged it only later, under questioning.^[5]

Numbers found the conduct aggravated by the office Renfer held:

Renfer's misdeeds are particularly odious because he undertook them while serving the Court and the people of the Eastern District of North Carolina as an Assistant United States Attorney, a position of immense trust and power.

D.E. 129 at 15

The sanction was a public reprimand, limited because Renfer had already left the office and "suffered substantial consequences."^[5] The court also ordered the U.S. Attorney's Office itself to show cause why it should not be held jointly responsible.

Two more, both against the United States

In Daghra v. Hinkley, the Justice Department defended an immigration detention by citing a Sixth Circuit decision called Taylor v. Hott, complete with a pinpoint page, a quotation, and an explanatory parenthetical. Chief Judge Hala Y. Jarbou checked:

The cited case, Taylor v. Hott, is not located at the identified page of the Federal Appendix. Indeed, page 387 is contained within a different opinion — Atkins v. CGI Techs. & Sols., Inc. — which is about commercial arbitration, not immigration bond determinations. In its research, the Court was unable to identify a Sixth Circuit case with the caption Taylor v. Hott, or any federal case containing the quoted language. Thus, it seems this citation was likely produced by generative artificial intelligence.

Daghra v. Hinkley, No. 1:26-cv-01429 (W.D. Mich.), ECF 11 (July 16, 2026)

No sanctions were imposed. The court said only that the government "must ensure its future filings with this Court do not include non-existent case law."^[8]

In Thomas v. U.S. Department of Justice, an Assistant U.S. Attorney searched her email and case files for a decision she had cited, then told the court: "This case does not exist. Undersigned counsel apologizes for including an improper citation."^[7] Her explanation is a distinct failure mode and deserves recording as such: she does not say she used a chatbot. She traced the citation to an ordinary internet search, and said she had not known that search engines were incorporating AI-generated results.

And Back Out Into the Documents

The TPP notices are not the only federal document to carry citations that dissolve on inspection.

In May 2025, the White House released the Make Our Children Healthy Again Assessment, chaired by HHS Secretary Robert F. Kennedy Jr. NOTUS found at least seven problematic citations: four were titles of papers that do not exist, and three mischaracterized what the real articles found.^[9]

That report did not merely resemble AI output. It carried the tool's fingerprints, left in the reference list. The original PDF — captured by the Internet Archive on May 22, 2025, before the White House quietly replaced the file — contains this as reference 2:^[10]

And this as reference 134:

The string :contentReference[oaicite:N]{index=N} is an artifact ChatGPT emits when it returns sourced text. It is not a citation format used by any journal, style guide, or human author. It is what a chatbot's output looks like when nobody removes the scaffolding. The index values run to at least 7, implying more of them upstream in the drafting.

The same document reprints the executive order that commissioned it, including its instruction that the commission "advise and assist the President on informing the American people regarding the childhood chronic disease crisis, using transparent and clear facts."^[10]

The same version carries, as reference 370, a 2022 JAMA Pediatrics paper titled "Changes in mental health and substance use among US adolescents during the COVID-19 pandemic," at volume 176, pages 1226–1234, with a DOI.^[10] That is one of the papers NOTUS identified as not existing.^[9] It was cited for the proposition that roughly a fifth to a quarter of adolescents report anxiety symptoms.

The response established the template. Press Secretary Karoline Leavitt attributed the errors to "formatting issues" that do not "negate the substance of the report." Asked whether AI had been used to produce it, she deferred to HHS.^[9] A citation to a study that was never written is not a formatting error.

The pattern recurred a month later, in a slide deck posted by the CDC for its Advisory Committee on Immunization Practices, prepared by former Children's Health Defense president Lyn Redwood. Under a heading reading "Evidence From Animal Research," it cited this:^[11]

PubMed returns nothing for that journal at that volume and page.^[16] Berman did publish on neonatal thimerosal exposure in 2008, under a similar-sounding title in a different journal: "Low-level neonatal thimerosal exposure: further evaluation of altered neurotoxic potential in SJL mice," Toxicological Sciences 101(2), 294–309.^[16] Real author, real subject, real year, and wrong on every locating detail. It is the same signature as the TPP notices: a citation shaped exactly like a real one, pointing nowhere.

How Common Is This, Actually

Restraint is warranted here, because the honest answer cuts against the alarm.

The largest public catalog of AI-hallucination incidents in court filings is the AI Hallucination Cases database maintained by legal scholar Damien Charlotin. It runs to thousands of entries, and only a small number are tagged as involving government lawyers at all. The overwhelming majority are pro se litigants and private attorneys.^[12] Federal public defenders appear in some tallies of "government" incidents; they are federally funded lawyers opposing the government, which is close to the opposite of the concern.

Nor is there a documented flood of AI slop through the Federal Register. No Government Accountability Office or inspector general report has yet identified a fabricated citation in a federal agency document.

What the Courts Are Doing

Judicial vocabulary for this is forming in public view. A full-text search of CourtListener's opinion corpus for the exact phrase "hallucinated citations" returns 26 published opinions: none before 2025, seven in 2025, and nineteen in the first eight months of 2026.^[13] That is an exact-phrase floor rather than a census — most opinions describe the problem in other words — but the direction is not ambiguous.

Cooper's phrase has a history worth tracing. In May 2025, commenting on the MAHA report for PolitiFact, University of Washington professor emeritus and AI researcher Oren Etzioni said that citing multiple nonexistent articles "is a hallmark of AI-generated citations, which often replicate the structure of academic references without linking to actual sources."^[9] Fifteen months later a federal judge used the same words, in an opinion about the same department. In the published-opinion corpus the phrase still returns exactly one result: Cooper's.^[13]

Some courts have moved past case-by-case reprimand. Under General Order No. 210, the U.S. Bankruptcy Court for the Southern District of California requires filers to complete a disclosure form. Form CSD 5013, effective January 1, 2026, asks the filer to name "the following generative AI program(s) was/were used in preparing the attached paper," and then to sign this:^[15]

The filer(s) of the attached paper certify that they checked the document for factual and legal accuracy using print reporters, traditional legal databases, or other reliable means outside of AI.

Form CSD 5013 [01/01/2026], Bankr. S.D. Cal.

The form also draws the boundary more carefully than most attempts have. Generative AI means a tool "which can create original content such as text or images in response to a user's prompt," including "the creation of a filed paper's initial content." Excluded are "spell checkers, predictive text prompts, grammar checkers, paraphrasing tools, text polishers and the like."^[15]

That is a court-wide rule rather than one judge's preference, and it demands verification by a named method rather than a promise of care. Individual judges elsewhere have issued their own standing orders on AI disclosure, and some have since withdrawn them, in at least one case after the district adopted a uniform local rule instead. There is no national policy, and the direction of travel is not uniform.

The Justice Department's own rules

DOJ has no public departmentwide policy governing generative AI in court filings. The distinction matters: internal documents exist. One of them is known only because Judge Numbers quoted it while reprimanding the prosecutor whose conduct prompted it — an instruction to "always personally verify each quote or proposition with your eyes in an actual case or law or other valid source."^[5]

The Thomas filing supplies the corroborating detail. That same U.S. Attorney's Office began cite-checking its civil filings in Westlaw "beginning in March 2026" — the month after the Renfer misconduct came to light.^[7] The rule was written afterward, at one office, in response to one sanctioned filing.

What the Ruling Did, and Did Not, Do

Coverage of the decision has mostly described a defeat for HHS.^[14] It was a partial one, and the part that failed matters more to the programs than the part that succeeded.

Cooper granted the preliminary injunction in part and denied it in part. The operative order enjoins HHS from "implementing the U.S. Department of Health and Human Services's 2026 Policy regarding the Teen Pregnancy Prevention Program."^[3] That is forward-looking. It stops the new criteria and the FY 2026 notices from being applied.

It does not give anyone their money back. Three days after publishing the notices, on June 26, 2026, HHS had sent termination letters to nearly the entire 2023–2028 grant cohort — 53 of 67 grants, worth about $68 million, cancelled two years before they were due to expire.^[4] Those terminations stand. Cooper said so in the last line of his analysis:

However, it will not order interim reinstatement of the terminated grants at this time because of lingering uncertainty over the Court's power to order such relief.

Mem. Op. at 60

The reason is jurisdictional, not sympathetic. Under the Tucker Act, claims founded on government contracts may belong exclusively to the Court of Federal Claims, and whether a cooperative agreement counts as a "contract" is unsettled — Cooper noted "the D.C. Circuit seems likely to weigh in soon." He also cited the Supreme Court's warning in Department of Education v. California about ordering payments the government has "no guarantee of recovering once disbursed."^[1]

There was a further trap. The one plaintiff with a clean jurisdictional path — SIECUS, a third party to the agreements rather than a grantee — could not show irreparable harm. The plaintiffs who were being destroyed were the ones whose claims might belong in another court. Cooper: "The Court sees no route to cobbling together a preliminary injunction where the viability of a particular claim rests on a plaintiff who has not experienced the requisite magnitude of harm."^[1]

Cooper set the injunction bond at one dollar.^[1]

The Shape of the Problem

Two failure modes are getting collapsed into one story, and they deserve separating.

The first is AI in the courtroom: a lawyer files a brief citing cases that do not exist. It is embarrassing, it is sanctionable, and it is largely self-correcting, because opposing counsel and the judge both read the citations. The system has antibodies.

The second is what happened here. The fabricated citations were not in a brief. They were in the agency action itself — the funding notice that set the rules for a $101 million federal program.^[1] They became the evidentiary basis for cancelling 53 grants. Had nobody sued, no judge would ever have read those footnotes. Grant notices are not adversarially tested. They are published, and people comply.

That is the difference between a bad citation and a manufactured justification. Administrative law asks agencies to show a "rational connection between the facts found and the choice made."^[1] A generative model asked to support a conclusion will produce facts shaped exactly like the ones that connection requires. The output is indistinguishable from evidence right up until someone checks.

Here, someone checked, because tens of millions of dollars were on the line and the losers could afford lawyers. That is not a reliable detection mechanism. It is a lucky one.

What Happens Next

The case is live. The parties filed a joint status report on September 1, 2026, proposing a schedule for further proceedings, as Cooper's order required. No notice of appeal appears on the docket, though the order states on its face that it is appealable.

Three things are worth watching:

• The D.C. Circuit on cooperative agreements. Cooper flagged that the court "seems likely to weigh in soon" on whether a cooperative agreement is a Tucker Act "contract." That answer decides whether district courts can ever order terminated federal grants restored — far beyond this program.
• Whether HHS tries a third time. The agency's response to Howell's 2025 vacatur was to reissue the policy. Nothing in Cooper's order prevents HHS from writing new notices under the previous TPP guidelines; he expressly said it may "reissue the NOFOs under the previously-applied TPP guidelines and thereby achieve its statutory mandate."
• Whether anyone establishes what actually produced the citations. No court has yet made a factual finding that a federal agency used generative AI to draft a policy document. "A hallmark of AI-generated citations" is a judge describing a pattern. The record does not yet contain an answer.