Attorney Felipe D.J. Millan was fined $1,500 over a brief with 19 AI-fabricated case citations

In Dehghani v. Castro, petitioner's counsel Felipe D.J. Millan purchased a brief from freelance attorney Janelle M. Lewis through the LAWCLERK marketplace for $750. Lewis likely used generative AI to draft the brief, which contained six fabricated case citations and thirteen additional mis-cited cases, then destroyed all work product per LAWCLERK policy. Magistrate Judge Damian L. Martinez sanctioned Millan with a $1,500 fine, mandatory one-hour CLE training on legal ethics or AI in writing, and orders to self-report to the New Mexico and Texas state bars and to report Lewis to the New York bar.

Felipe D.J. Millan · Incident Apr 2, 2025 · Indexed Jun 4, 2026 · 3 sources

Generative AI hallucinated 19 fake or mis-cited legal authorities, and the freelancer's destruction of work product made it impossible to confirm the tool used.
What
In Dehghani v.
Incident date
Apr 2, 2025
Who
Felipe D.J. Millan
Failure mode
Hallucination
AI surface
Chatbot
Severity
Medium

What happened

Petitioner's attorney Felipe D.J. Millan hired freelance attorney Janelle M. Lewis through the LAWCLERK marketplace to draft a response brief in a habeas corpus proceeding for $750. Millan signed and filed the brief without reviewing or verifying any of the cited cases, which contained six entirely fabricated citations and thirteen cases that did not support the propositions claimed. The court issued multiple show cause orders and held a hearing on March 26, 2025. On April 2, 2025, Magistrate Judge Martinez imposed a $1,500 fine, required a one-hour CLE on legal ethics or AI in legal writing, and ordered Millan to self-report to the New Mexico and Texas state bars and to report Lewis to the New York bar. District Judge Margaret Strickland affirmed the sanctions in full on May 9, 2025.

What broke inside the model

Failure path · mode profile · Hallucination
  1. 01 · TriggerA user asks for a fact, a citation, or a figure.
  2. 02 · Model stepThe model writes a fluent, confident answer.
  3. 03 · Control gapNothing ties the claim back to a real source.
  4. 04 · FailureA fabricated fact ships as if it were verified.
  5. 05 · ConsequenceThe false claim reaches a customer, a court, or the public.

Confidence holds, and even spikes, as the claim detaches from any source.

Generative AI produced hallucinated legal citations by fabricating case names, docket numbers, and legal propositions that do not exist, a well-documented failure mode of large language models when generating legal text. The freelance attorney's reliance on the AI output without verification, combined with LAWCLERK's policy of destroying work product, eliminated any audit trail that could confirm which AI tool was used. The filing attorney's failure to perform even minimal verification of the cited cases before signing the brief allowed the fabricated authorities to reach the court record.

Public visibilityMedium
Regulatory exposureActive
Customer impactFew customers
Financial impactDisclosed
Time to disclosureWeeks
  1. Court FilingAZADEH DEHGHANI v. DORA CASTRO (2025) - FindLaw Caselawcaselaw.findlaw.com
  2. Court FilingDehghani v. Castro, No. 2:2025cv00052 - Document 28 (D.N.M. 2025)law.justia.com
  3. PressAI IP Year in Review - AI Hallucinations in Court Filings and Orders: A 2025 Review of Sanctions Across the Courts and Rule Proposalssternekessler.com
Permalinkhttps://failureindex.ai/failures/attorney-felipe-d-j-millan-fined
CitationAI Failure Index. "Attorney Felipe D.J. Millan was fined $1,500 over a brief with 19 AI-fabricated case citations" (FI-0127). Realm Labs. https://failureindex.ai/failures/attorney-felipe-d-j-millan-fined (indexed Jun 4, 2026).
Share cardA branded image of this record for posts and slides.

Data fields CC-BY 4.0, prose citation permitted. Incident ID FI-0127. Full dataset at /data.

Note from Realm Labs, the Index steward

How Realm would have caught this

Controls for this failure mode
  • Prism
  • OmniGuard
  • AI Detection & Response (AIDR)

A runtime layer that watches the model's internal state can flag the moment a model commits to a claim it has no support for, and hold or reroute the response before it reaches a user. Realm reads those signals in real time rather than grading the transcript after the fact.