After An AI Hallucination, Every Filing Becomes A Test
Not every bad cite is an AI hallucination. But once a firm gets hit with court admonishments - including one that costs real money - the presumption of competence evaporates. From that point on, every odd citation invites the same question: is this AI error or lax oversight? And courts are far less likely to extend grace.
The Gordon Rees Timeline
Last summer, an attorney at Gordon Rees filed a bankruptcy brief in Jackson Hosp. & Clinic Inc. with "pervasive inaccurate, misleading, and fabricated citations, quotations, and representations of legal authority," triggering an order to show cause. The firm reimbursed fees, apologized, and said it had updated policies with a "cite checking policy."
Months later, on December 3, 2025, the firm drew another reprimand in Villalovos-Gutierrez v. Pol. U.S. Magistrate Judge Carolyn Delaney wrote: "Counsel shall not file or otherwise present to the court any documents which contain AI-hallucinated citations or fictitious or non-existent legal citations. Counsel's failure to confirm the existence of, as well as the accuracy and veracity of a case or other legal citation created by an AI tool or taken from another indirect source, is a potential ground for sanctions."
A policy only works if people follow it. If these issues keep surfacing, the process is either being ignored or it isn't enough.
Fresh Allegations In Huynh v. Redis Labs
In a discovery fight now on its third motion to compel, the defense - represented by Gordon Rees - had not completed a deposition after two orders and monetary sanctions, citing a late medical emergency for a cancellation. The plaintiff, represented by Bach Mili LLP, renewed the motion; in reply, they allege the opposition brief leans on mischaracterized or fabricated authority.
According to the reply, the brief purportedly flips the holding of Doppes v. Bentley Motors, Inc., invokes Collisson & Kaplan v. Hartunian for a phrase that doesn't appear while that case affirmed striking an answer, attributes a non-existent definition to Biles v. Exxon Mobil Corp., cites language not in Rail Services of America v. State Comp. Ins. Fund, and suggests Huh v. Wang and Corns v. Miller excuse noncompliance based on medical emergencies. The reply calls these "reversals of holdings and quotations that do not exist."
The filing also cites local precedent that an attorney has a nondelegable duty to read and verify every cited case, and that reliance on hallucinated authority renders a filing frivolous and sanctionable. Whether any specific cite is AI-generated or simply wrong, the risk profile is the same.
Why This Matters Beyond One Case
Once a firm is linked to hallucinated citations, the reputational spillover is real. Opposing counsel will parse every cite with a loupe. Judges will remember. Even defensible characterizations start from a credibility deficit.
That credibility hit can influence sanctions, discovery relief, and how the court frames "willfulness" or "substantial justification." In close calls, benefit of the doubt goes to the party that looks careful and accountable - not the one with a history of citation problems.
A Practical Playbook To Keep AI Out Of Your Sanctions Orders
- Nondelegable cite check: Require the signing attorney to personally open, read, and pin-cite every authority. No exceptions. No summaries without source review.
- Two-source verification: Validate every case in a primary database (Westlaw/Lexis/Bloomberg or official court site). If AI helped, treat every output as untrusted until verified.
- Quotes with receipts: Match every quotation to the opinion PDF. Add pin cites and short parentheticals that reflect the actual holding, not dicta.
- AI usage log: If any AI tool touched research or drafting, keep a log of prompts, outputs, and human edits. You won't file it, but you'll be ready if questioned.
- Pre-filing red team: A second lawyer (or senior paralegal) independently re-checks cites and quotes, not just Bluebook form. Measure error rates and coach to trend down.
- Ban indirect sources for holdings: No relying on AI, headnotes, or secondary summaries to state a rule. Always return to the opinion text.
- Discovery discipline: Track compliance dates, court directives, and sanctions exposure on a single sheet. If a medical issue arises, paper it with declarations and alternatives (remote, stipulations).
- Use pin-cite templates: Standardize how your team writes case parentheticals for discovery sanctions (e.g., terminating sanctions, issue/evidence sanctions), with jurisdiction-specific authority.
- Escalation triggers: After one court warning, require partner sign-off on all briefs citing contested standards (Rule 37 equivalents, spoliation, terminating sanctions).
- Post-mortems: If a miscite slips through, hold a same-week review, update checklists, and audit the next three filings. Treat it like client risk, because it is.
Baseline Authorities Worth Re-Reading
Refresh your bearings on the obligations you're certifying with your signature and the discovery sanctions framework:
- Federal Rule of Civil Procedure 11 (representations to the court; reasonable inquiry)
- Federal Rule of Civil Procedure 37 (discovery failures; sanctions, including terminating sanctions)
Want To Uplevel Team Training
If your firm is tightening policy and workflows around AI-assisted research, start with practical, legal-specific training that reinforces verification habits and cite integrity: AI for Legal.
The Takeaway
Perception now matters as much as precision. If you've been burned once, you must over-index on verification and documentation. Treat every citation like an exhibit. Trust nothing you haven't personally opened and read.
Your membership also unlocks:
