Incident registry: Binnall Law — Claude Console phantom citations (2026) → — PSF-mapped entry alongside related court filing hallucinations.
This incident note documents a May 2026 filing error in AFGE v. Trump, N.D. Cal., No. 3:25-cv-03698. Facts below come from court declarations described by Bloomberg Law (May 15, 2026), the AI Incident Database (Incident 1499), and the firm's corrective filings cited in those reports.
Verified facts
| Field | Detail |
|---|---|
| Date surfaced | May 2026 (motion filed May 6; apologies reported mid-May) |
| Organisation | Binnall Law Group PLLC (counsel for witness Joseph Guy in federal layoffs litigation) |
| AI system | Anthropic Claude Console (enterprise, data-isolated per counsel declaration as reported) |
| Failure mode | "Phantom" quotations: language presented as case quotes that does not exist in the cited opinions |
| Court | U.S. District Court, Northern District of California (Judge Susan Illston) |
| PSF domains | D2 D6 D5 |
What happened
- Partner Jason Greaves used Claude Console to produce an initial draft of a motion to quash a subpoena in AFGE v. Trump, citing tight deadlines (per Bloomberg Law and AIID summaries of his declaration).
- Greaves sent the draft to an associate for review. He later stated he checked cases for general applicability but did not verify that quoted language actually appeared in those opinions.
- Judge Illston flagged quotations that could not be found in the cited cases. Greaves, founding partner Jesse Binnall, and partner Lindsay McKasson filed apologies and described firm policy gaps.
- The firm committed to supplemental training and tighter use of established legal research platforms (Westlaw or LexisNexis) for citation work, as reported by Bloomberg Law.
PSF mapping
D2 Output validation
Generative drafts can read as authoritative even when quotations are fabricated. For any workflow where model text becomes a filing, treat every citation as untrusted until matched against a live reporter database. See the D2 output validation guide and the registry pattern on court filing hallucinations.
Practitioner action: Block filing export until each quote passes an automated string match against retrieved opinion text, not just a human skim of case summaries.
D6 Human oversight
Associate review did not catch phantom quotes before submission. Supervising counsel remained the signer and accepted responsibility in declarations summarized by Bloomberg Law. High-consequence outputs need a named accountable reviewer with a checklist tied to citation integrity, not general topical review.
Practitioner action: Require partner sign-off only after a citation verification log is attached to the filing packet.
D5 Deployment safety
The firm reportedly had AI policies but still routed a subpoena motion through Claude Console under deadline pressure. Production paths for legal AI should define which motion types may use generative drafts, which require traditional research tools only, and what rollback procedure applies when a judge questions citations.
Practitioner action: Add a deployment gate: no generative draft may enter e-filing without a completed citation attestation record stored with the matter file.
Prevention controls
- Ground drafting tools on retrieved opinion segments; never ask the model to invent quotations from memory.
- Run post-generation citation extraction and match each quote to source PDF or reporter text before human review.
- Separate "brainstorm" and "fileable" environments so deadline pressure cannot bypass the fileable path.
- Log model version, prompt hash, and reviewer identity for every court-facing draft (supports sanctions defense and internal audit).
Context on production legal AI
Judges have sanctioned or admonished counsel in more than 100 cases linked to careless generative AI use since 2022, per reporting summarized in AIID Incident 1499. The Binnall episode is distinct because it names a specific enterprise console (Claude Console) in active federal employment litigation, not a consumer chatbot experiment.
For failure taxonomy and monitoring expectations, see seven failure modes of production AI and AI behaviour contracts.
Sources
- Bloomberg Law: Lawyers apologize for fake AI quotes in Trump mass layoffs case (May 15, 2026)
- AI Incident Database: Incident 1499 (indexed May 18, 2026)
- Reuters: Lawyer apologizes for phantom AI quotes (May 18, 2026, cited via AIID)
- Docket: AFGE v. Trump, No. 3:25-cv-03698 (N.D. Cal.)
- PAI incident registry: Court filing hallucinations
This record is maintained by PAI and free to cite. If something is wrong or missing, tell us. Corrections and source suggestions keep the record honest.