The Four Levels of AI Citation Error

Most attorneys following AI sanctions cases know the general shape of the problem. An attorney uses generative AI. The tool produces citations. The citations turn out to be fabricated. Sanctions follow.

What most of those cases have not done is explain precisely what "fabricated" means, or map the full spectrum of ways AI citation errors differ from each other. An opinion issued on April 29, 2026 by U.S. Magistrate Judge Gabriel W. Gorenstein in the Southern District of New York does exactly that.

Jimenez-Fogarty v. Fogarty et al., 24 Civ. 8705 (S.D.N.Y. Apr. 29, 2026), is the most analytically useful document a court has produced on this subject. If you use AI in your research or drafting practice, it is worth reading in full.

The Court's Four-Category Taxonomy

Judge Gorenstein's opinion begins with something no prior sanctions order has done: a careful sorting of citation errors into categories, with an explicit statement of which the court would and would not treat as sanctionable.

Category 1: Citations to real cases at the wrong volume or page number. The case exists, it supports the proposition, the reporter location is just wrong. The court excluded these from its analysis. (Id. at 2 n.3.)

Category 2: Citations with the wrong case name but a correct reporter citation. The case is real, the argument holds, the name is wrong. Also excluded. (Id. at 2 n.4.)

Category 3: Citations to real cases whose holdings are completely mischaracterized. The case exists and was found. It just says the opposite of what the brief claims it says. The court called these "grossly misleading and perhaps deserving of sanctions" but set them aside to focus on the most severe category. (Id. at 2-3 n.5.)

Category 4: Cases that cannot be located at all, or found at the cited reporter but having nothing to do with the proposition cited. The court called these "completely made up." (Id. at 3.) This is the category that produced the Rule 11 finding and the $2,500 sanction.

This taxonomy tells you that not all AI citation errors are equal, and that courts are starting to analyze them with precision. A typo in a reporter citation is different from a fabricated case. A mischaracterized holding is different from a nonexistent one. Each category reflects a different failure in the verification process.

What Verification Actually Requires at Each Level

The Second Circuit stated the standard in Park v. Kim, 91 F.4th 610, 615 (2d Cir. 2024), and Judge Gorenstein quoted it directly: "At the very least, the duties imposed by Rule 11 require that attorneys read, and thereby confirm the existence and validity of, the legal authorities on which they rely." (Id. at 6.)

Working through the taxonomy, that standard requires something different at each level.

For Categories 1 and 2, the fix is straightforward: pull the case, confirm it exists at the citation given. Any attorney who retrieves the actual opinion catches a wrong reporter or wrong case name immediately.

For Category 3, the requirement is more demanding. You have to read the case closely enough to confirm it actually says what you are claiming it says, not just that it exists. An AI tool will confidently summarize a holding in a way that supports your argument. The only check is reading the opinion yourself and confirming the holding at the specific passage you are relying on.

For Category 4, the requirement is the same, but the consequence of not checking is total. There is no case to read. The first time you would discover it does not exist is when opposing counsel or the court looks it up.

The Mitigation Roadmap: What Avoiding Bad Faith Requires

The opinion also does something no prior order has done with this much clarity: it explains precisely what mitigation looks like, by distinguishing two cases where courts found no bad faith despite fabricated citations appearing in filed briefs.

In United States v. Cohen, 724 F. Supp. 3d 251 (S.D.N.Y. 2024), no bad faith was found because the attorney had some reason to believe the citations had been vetted — they came from a client who was also a lawyer — and the attorney accepted responsibility fully. (Id. at 14-15.)

In Hall v. Academy Charter School, 2025 WL 2256653 (E.D.N.Y. 2025), no bad faith was found because the attorney provided a complete and honest account of what happened, the circumstances were genuinely unusual, and responsibility was accepted without qualification. (Id. at 15.)

In both cases, the attorneys did three things: gave a complete and specific account of what occurred and why, identified the source of the fabricated citations, and accepted responsibility without minimizing. The attorney in Jimenez-Fogarty did none of these. She gave vague and contradictory explanations across multiple filings, declined to name the AI tool even after being directly ordered to describe the drafting process, and described the errors as "not uncommon in complex litigation." (Id. at 14.)

The court was direct: "Regret and apologies are not necessarily enough to avoid the imposition of sanctions for the submission of non-existent legal authority." (Id., citing Benjamin v. Costco Wholesale Corp., 779 F. Supp. 3d 341, 350 (E.D.N.Y. 2025).) What is enough, Cohen and Hall show, is transparency, specificity, and genuine acceptance of responsibility.

What Orthodoxy Produces — Mapped to the Taxonomy and Mitigation Roadmap

The taxonomy and the mitigation roadmap together describe two things: what verification requires before filing, and what documentation an attorney needs when something goes wrong. Orthodoxy addresses both.

What the court's framework requires	What Orthodoxy produces
Confirm cited cases exist at the cited location (Categories 1–2)	Passage-level citation record with attorney sign-off confirming each case was independently reviewed before filing
Confirm holdings are accurately characterized (Category 3)	Independent contradiction check that flags where citations conflict with each other or with the legal proposition being researched
Confirm citations are not fabricated (Category 4)	Attorney-signed quality-check sign-off at the passage level, creating a contemporaneous record that each citation was independently verified
Complete identification of the AI tool used (mitigation roadmap)	Session record identifying the research query and the specific output the attorney reviewed
Contemporaneous record of the verification step (mitigation roadmap)	Immutable compliance certificate, timestamped and attorney-signed — cannot be altered after signing without detection
Evidence of actual review, not mere assertion of review	Passage-level attestations for each citation — not a single blanket certification, but a sign-off at the level the court's taxonomy requires

Orthodoxy does not verify citations for you. It documents that you did. That distinction matters because Cohen, Hall, and Jimenez-Fogarty together show that what courts find persuasive is not an attorney's assertion that verification occurred — it is a contemporaneous record demonstrating it.

The attorney in Jimenez-Fogarty told the court she "manually cross-checks case names, docket numbers, and reporter citations against the primary sources in legal databases." (Id. at 9.) The court found that statement obviously false because a manual check of any of the fabricated citations would have immediately revealed them. There was no contemporaneous record to support the assertion. Orthodoxy creates that record.