# A federal appeals court fines a lawyer $2,500 for an AI-drafted brief, even though she used purpose-built legal-AI tools

> In Fletcher v. Experian Information Solutions, Inc. (No. 25-20086, filed Feb. 18, 2026), the U.S. Court of Appeals for the Fifth Circuit sanctioned plaintiff's appellate counsel Heather Hersh $2,500 under Fed. R. App. P. 46(c) and its inherent authority. Her reply brief contained '16 instances of fabricated quotations and 5 additional serious misrepresentations of law or fact'; the court found she used generative AI to draft 'a substantial portion, if not all,' of it, and identified two of her named sources, Casetext/CoCounsel and vLex, as purpose-built legal-AI products, not general chatbots.

- Verification status: verified
- Case type: deployment
- Provider: Generative AI, unattributed to a single product; the court identified Thomson Reuters CoCounsel (formerly Casetext) and vLex among the named sources as 'generative AI products'
- Client: U.S. Court of Appeals for the Fifth Circuit: Fletcher v. Experian Information Solutions, Inc., No. 25-20086, Courts / legal (AI-governance honest-negative) (named)
- Sector: legal-services / US / cross
- Verified on: 2026-08-20
- Canonical URL: https://theinternetninja.com/stories/fletcher-v-experian-5th-cir-hersh-2500-ai-fabricated-citations-frap-46c-inherent-authority-2026/
- Source: The Internet Ninja (theinternetninja.com), independent verified-proof platform

## Outcomes

| Metric | Before | After |
| --- | --- | --- |
| $2,500 monetary sanction on plaintiff's appellate counsel Heather Hersh, payable to the clerk of the Fifth Circuit within 30 days |  |  |
| Imposed under Fed. R. App. P. 46(c) ('conduct unbecoming a member of the bar') and the court's inherent power to sanction abuse of the judicial process, not Rule 11 |  |  |
| Show-cause order enumerated 16 fabricated quotations + 5 additional serious misrepresentations of law or fact; court found AI drafted 'a substantial portion, if not all,' of the reply brief and counsel 'was not forthcoming' |  |  |

## Verification method

Court-adjudicated. Every figure is quoted verbatim from the published Fifth Circuit opinion (No. 25-20086, filed Feb. 18, 2026), which is byte-tied to Wayback via the official ca5.uscourts.gov PDF, and corroborated firsthand by two mutually-independent newsrooms: The Texas Lawbook and Paul Mollica's Federal Courts Blog on the 16+5 enumeration, The Texas Lawbook and Legal Newsline on the $2,500 amount. All quotes re-bound verbatim against local captures this session.

## FAQ

**How much was the lawyer fined in Fletcher v. Experian?**

The Fifth Circuit ordered plaintiff's appellate counsel Heather Hersh to pay $2,500 in sanctions to the court within 30 days, resting the sanction on Fed. R. App. P. 46(c) ('conduct unbecoming a member of the bar') and the court's inherent power rather than Rule 11.

**Did purpose-built legal AI tools cause the fabricated citations?**

The court identified two of the sources Hersh named, Casetext/CoCounsel and vLex, as generative-AI products, but it did not pin the specific fabrications to any one of them, hedging that '[t]o the extent these products generated the inaccurate summaries.' The show-cause order enumerated 16 fabricated quotations and 5 additional serious misrepresentations of law or fact.

## Full case file

## The problem
Robert Fletcher sued Experian and a lender, Bridgecrest, under the Fair Credit Reporting Act, alleging an auto-finance account had been opened in his name by an identity thief; the case reached the U.S. Court of Appeals for the Fifth Circuit, where plaintiff's appellate counsel Heather Hersh, the only attorney to sign the reply brief, filed a brief the court found "containing numerous inaccurate citations, quotations, and statements of fact" [source](https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf). On December 18, 2025, the court issued an order to show cause "enumerating 16 instances of fabricated quotations and 5 additional serious misrepresentations of law or fact" [source](https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf). An independent bar publication read the same holding from the opinion, reporting that the Fifth Circuit "sanctioned a lawyer $2,500 for filing a reply brief with several 'hallucinated' case citations and then providing evasive responses to the court" [source](https://texaslawbook.net/fifth-circuit-sanctions-opinion-gives-practical-advice-for-ai-use/), and a second, separately-bylined legal blog recorded that "the lawyer allegedly filed a reply brief in the Fifth Circuit with sixteen 'instances of fabricated quotations and 5 additional serious misrepresentations of law or fact'" [source](https://paulmollicafederalcourtsblog.com/2026/02/19/fifth-circuit-gives-what-not-to-do-guidance-for-the-lawyer-who-learns-too-late-that-their-brief-was-filed-in-court-with-ai-hallucinations/).

## What was built
No system was built or deployed. This is a court-origin AI-governance honest-negative, a federal appeals court's response to unverified generative-AI output submitted as law, and the opinion (No. 25-20086, filed Feb. 18, 2026; Chief Judge Jennifer Walker Elrod writing for a panel with Judges Smith and Wilson) found that counsel used artificial intelligence to draft "a substantial portion, if not all," of the reply brief and then failed to verify the content it generated [source](https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf). The distinctive detail is that when pressed, Hersh named the sources she had relied on, and the court identified two of them as, in its words, "The remaining two sources, Casetext, which has been spun off into Thomson Reuters's CoCounsel … and vLex, are both generative AI products" [source](https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf), which are purpose-built legal-research tools, not a general chatbot, yet the brief still carried fabricated authority. The court did not pin the specific fabrications to any one product, hedging its attribution with "[t]o the extent these products generated the inaccurate summaries" [source](https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf).

## The outcome
**The sanction: <span class="kpi">$2,500</span>, payable to the clerk of the Fifth Circuit within <span class="kpi">30 days</span>, under Fed. R. App. P. 46(c) and the court's inherent power, not Rule 11.**

**What triggered it: <span class="kpi">16</span> fabricated quotations plus <span class="kpi">5</span> additional serious misrepresentations of law or fact in a single reply brief.**

The court sanctioned Hersh $2,500, ordering that "Heather Hersh shall pay $2,500 in sanctions to the United States Court of Appeals for the Fifth Circuit within 30 days of this order" [source](https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf). It rested the sanction on the court's inherent power and Fed. R. App. P. 46(c) rather than Rule 11. As The Texas Lawbook put it, the court "sanctioned the attorney $2,500, citing the court's inherent power and Fed. R. App. P. 46(c), which addresses 'conduct unbecoming a member of the bar'" [source](https://texaslawbook.net/fifth-circuit-sanctions-opinion-gives-practical-advice-for-ai-use/). The court was explicit that candor would have mattered: "Had Hersh accepted responsibility and been more forthcoming, it is likely that the court would have imposed lesser sanctions" [source](https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf). It framed the moment broadly, warning that AI hallucinations are "a problem that is getting worse—not better" [source](https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf). Legal Newsline reported the same result independently, noting the Fifth Circuit "ordered Dallas attorney Heather Hersh to pay $2,500 for submitting a brief filled with 'hallucinations,' or nonexistent court citations" [source](https://www.legalnewsline.com/southeast-texas-record/ai-hallucinations-in-dallas-lawyer-s-brief-lead-to-sanctions/article_27feefb2-bb9f-4ad8-8bb7-1a60333a2b63.html).

## The weakest load-bearing source, named where you meet it
Every critical figure here rests first on a **Tier 1** primary, the published Fifth Circuit opinion itself, an independent court-of-record origin, not a party's account, so the corroboration around it is confirmation, not the load-bearing weight. Of the corroborating sources, the weakest is **Legal Newsline** (Tier 2): it carries the $2,500 amount firsthand but paraphrases the "hallucinations" finding rather than reproducing the exact "16 instances … and 5 additional serious misrepresentations" enumeration, so it corroborates the sanction amount but not the count. The second corroborator for that count, **Paul Mollica's Federal Courts Blog**, is a bylined practising-attorney blog rather than a traditional newsroom, reputable and independent, but a single-author outlet; it is counted as an independent secondary, not as a primary.

## Conflation guards
- The **$2,500 is the only sanction in this order.** It is separate from, and much smaller than, the district court's earlier sanctions against a *different* lawyer, Shawn Jaffer (about $20,000 to Bridgecrest under Rule 11 and about $13,000 to Experian under 28 U.S.C. 1927), which the Fifth Circuit **vacated** in a separate Jan. 6, 2026 opinion; that is not this AI sanction.
- The order does **not** pin the fabrications to a single AI product. It names CoCounsel/Casetext and vLex as generative-AI products among the sources Hersh cited, but hedges attribution; secondary coverage that flatly says "ChatGPT" is not supported by the order.

## How this was verified
- **Method:** Court-adjudicated honest-negative. Every quoted figure and holding was read verbatim from sources captured this session and re-bound against local copies; no number is paraphrased, rounded, or recalled.
- **Primary anchor:** the published Fifth Circuit opinion, *Fletcher v. Experian Information Solutions, Inc.*, No. 25-20086 (filed Feb. 18, 2026): **Tier 1**, an independent court-of-record origin. It is **byte-tied**: the archived ca5.uscourts.gov PDF (Wayback `20260506114802`, `sources/fletcher-v-experian-5cir-opinion-25-20086.pdf`) matches the live CDX digest for that snapshot, and its extracted text carries the sanction order, the 16+5 enumeration, the "generative AI products" line, and the "getting worse—not better" passage verbatim.
- **Independence:** the $2,500 amount is carried firsthand by two mutually-independent newsrooms (The Texas Lawbook and Legal Newsline); the 16+5 enumeration is carried firsthand by two mutually-independent secondaries (The Texas Lawbook and Paul Mollica's Federal Courts Blog). The Fifth Circuit opinion is the origin and is never counted toward corroboration.
- **Open item (human, post-graduation):** whether Hersh paid the $2,500 within the 30-day window, any appeal or rehearing of the sanctions order, and any Texas State Bar or Fifth Circuit disciplinary follow-on (the opinion notes Hersh was previously sanctioned in *Powell v. Nelnet*). These are docket-only and asserted by no claim above; the sanction and the quotes are already verifiable from the primary.
- **Verified on:** 2026-08-04 (presentation rebuilt 2026-08-19; inline citations re-keyed to named-source links, all quotes re-bound verbatim against local captures).

## Related case files
- [United States v. Farris: a federal appeals court denies a lawyer his pay over AI-fabricated quotations from CoCounsel](/stories/us-v-farris-6th-cir-howe-cocounsel-ai-fabricated-quotations-cja-compensation-denied-removed-2026/): the closest sibling, another U.S. court of appeals confronting fabrications traced to a purpose-built legal-AI product (CoCounsel), but reaching for a different remedy, denied compensation and removal rather than a fine.
- [Torres Campos v. Munoz: a California appellate court sanctions a lawyer $5,000 for AI-hallucinated citations](/stories/torres-campos-v-munoz-cal-app-4th-buchanan-5000-ai-fabricated-citations-dog-custody-2026/): the money-sanction parallel at a state appellate level, showing how a published court frames a comparable fine for the same conduct.
- [Italy's Supreme Court fines an appellant €5,000 for AI-hallucinated case citations](/stories/italy-cassazione-23006-2026-5000-euro-ai-fabricated-citations-diligenza-digitale/): the foreign-court parallel, a supreme court in another jurisdiction reaching the same principle, that using generative AI does not exempt counsel from verifying its output.

## Sources
1. **[Tier 1]** U.S. Court of Appeals for the Fifth Circuit · *Fletcher v. Experian Information Solutions, Inc.*, No. 25-20086 (published opinion, Elrod, C.J.) · February 18, 2026 · https://www.ca5.uscourts.gov/opinions/pub/25/25-20086-CV0.pdf
2. **[Tier 2]** The Texas Lawbook · "Fifth Circuit Sanctions Opinion Gives Practical Advice for AI Use" · February 19, 2026 · https://texaslawbook.net/fifth-circuit-sanctions-opinion-gives-practical-advice-for-ai-use/
3. **[Tier 2]** Paul Mollica's Federal Courts Blog · "Fifth Circuit gives 'what not to do' guidance for the lawyer who learns too late that their brief was filed in court with AI hallucinations" · February 19, 2026 · https://paulmollicafederalcourtsblog.com/2026/02/19/fifth-circuit-gives-what-not-to-do-guidance-for-the-lawyer-who-learns-too-late-that-their-brief-was-filed-in-court-with-ai-hallucinations/
4. **[Tier 2]** Legal Newsline / Southeast Texas Record · "AI hallucinations in Dallas lawyer's brief lead to sanctions" · February 24, 2026 · https://www.legalnewsline.com/southeast-texas-record/ai-hallucinations-in-dallas-lawyer-s-brief-lead-to-sanctions/article_27feefb2-bb9f-4ad8-8bb7-1a60333a2b63.html