Fifth Circuit sanctions lawyer over unchecked AI citations

You must verify each individual claim against a primary source before relying on AI-generated output in any professional context where the decision carries legal, medical or financial consequence. Record that verification by name, date and outcome, and do not treat a general disclaimer as evidence that verification occurred.

The prompt is an analysis of 22 October 2026 of claim-level verification in AI-assisted professional work, which argues that language models generate plausible-looking but fabricated or incorrect references at scale, and that existing professional responsibility rules apply in full regardless of the tool used. A lawyer was sanctioned by the Fifth Circuit for filing an appeal brief containing fabricated citations and incorrect legal statements that had not been checked. In our assessment, the ruling establishes that you cannot transfer responsibility for accuracy to the tool or to a disclaimer; instead, you must demonstrate a concrete verification step before any output enters a decision-making process.

What does the Fifth Circuit ruling actually require?

The court imposed a financial sanction on a lawyer who used generative AI to draft substantial portions of an appeal brief without verification, resulting in fabricated references and misstatements of law being filed with the court. The ruling did not prohibit AI use; it held that the signatory's duty to verify accuracy does not change because of the tool employed. Several citations and legal assertions in the brief were not supported by the underlying sources, despite their apparently authoritative form. The court stressed that existing professional and procedural rules apply in full.

Why do language models generate false citations at such scale?

Language models predict probable language sequences; they do not retrieve or verify truth. They can generate a reference that reads as authoritative—complete with author names, publication titles, dates and page numbers—while being entirely fabricated or substantively wrong. A peer-reviewed study of three language models generating citations in neurocritical care found that 165 of 300 references contained an inaccuracy and 85 were entirely fabricated. That study examined zero-shot, retrieval-free web interfaces without linked sources; the figures should not be read as a universal error rate across all models or workflows. The risk deserves particular attention in specialised subjects with thin or fragmented source coverage.

What controls actually work to catch these errors?

A general warning or disclaimer does not ensure that verification actually happens. Academic research on organisation-backed AI advisors distinguishes between scepticism, factual verification and their influence on trust, and concludes that general warnings have weak or mixed effects in existing studies. More targeted cues about typical failure points show promise, but remain limited in scope. In our assessment, this means that a checkbox policy or a lone warning in tooling is not a defensible control measure; it does not prove that anyone tested the output.

You must establish the following:

1. Assign verification to a named competent professional — identify who is responsible for checking each claim before the output is used or filed, and record that assignment in writing.
2. Check each claim against a primary source passage — do not rely on the model's citation; retrieve the original source and confirm that the cited passage genuinely supports the specific assertion.
3. Record the verification outcome by name, date and result — document who checked what, when, and whether the claim was supported, unsupported or required correction.
4. Escalate high-consequence decisions to independent review — for legal, medical or financial decisions, consider a second independent assessment before action, in line with your risk policy.
5. Treat one incorrect reference as a flag, not a rejection — one fabricated citation does not invalidate the entire piece, but must be corrected and the correction recorded before anyone builds on it.

How do you separate accuracy from source support?

Three things quickly become conflated: whether a claim happens to be factually correct, whether it is supported by the sources cited, and who bears legal responsibility for the decision. A claim may be correct and yet lack source support; conversely, a decision that matches a human reference may still be insufficiently supported by sources. You assess not the accuracy of the model itself, but the source support for each individual claim. That distinction determines what you must demonstrate to a regulator, a court or an auditor.

Tooling can help you organise and record this work—linking claims to sources, flagging unverified passages, tracking who checked what and when. What tooling cannot do is replace your own professional judgement about whether a source genuinely supports a claim, or whether the claim matters enough to escalate to independent review. That remains your responsibility.