Fifteen Cases for a Winery Dispute
- Admitted
- 2026-09-22
- Source dataset
- Other / sample
- Source
- ABA Journal
- Failure mode
- Hallucination with side effects
- Agent system
- Single agent
- Tool access
- Read-only
- Industry
- Legal
Presenting Behaviour
In briefing cross-motions for summary judgment in a family dispute over a winery, attorneys filed three briefs to a federal court in Oregon that together contained 15 references to nonexistent court cases and eight fabricated quotations, all generated by an AI tool and never verified before filing. A U.S. magistrate judge, describing the case as "a notorious outlier in both degree and volume" among AI-hallucination sanctions cases, ultimately imposed a combined $110,000 in fees and fines on the two attorneys involved.
Case Replay
A reconstruction, not a simulation. This sequence is reconstructed from the public incident record and summarised in our own words. It is a static record of what was reported, and nothing here can be run or replayed live.
4 steps
- Briefs drafted. Attorneys use AI to draft three filings in a winery dispute.
- Fabrications included. 15 nonexistent cases and eight invented quotations appear. This is where the agent derailed.
- Not verified. No human check catches any of them.
- Sanctions. A magistrate imposes $110,000 in fees and fines.
Differential Diagnosis
A single fabricated citation that slipped through review (considered and ruled out; the scale here) 15 nonexistent cases and eight invented quotations across three separate filings, rules out an isolated slip and points to a systemic failure to check AI output against real sources at any point before filing.
Deliberate fraud on the court, considered and ruled out as the primary driver of the underlying content; the judge's opinion frames this as a hallucination-and-verification failure consistent with other AI sanctions cases, not an attempt to deliberately deceive with citations the filers knew were fake, though one attorney's later attempt to quietly remove the errors and refile, once caught, compounded the original problem.
Brief Psychotic Disorder, describing the AI tool's own output pattern rather than the attorneys' conduct, most consistent with how the fabricated content itself was produced: confident, specific, well-formatted case citations and quotations that read exactly like real legal authority, generated in isolated bursts across three separate briefs, with no self-correction from the tool at any point.
Working Diagnosis
Brief Psychotic Disorder: an AI drafting tool repeatedly produced confident, specific, and entirely fabricated legal citations and quotations across three separate court filings, with the failure compounding because no human verification step caught any of them before filing.
Discussion
The lesson here sits mostly with the humans rather than the tool: a fabricated citation is a known, common failure mode of these systems by now, and the safeguard (actually opening and reading a cited case before filing it) is neither novel nor difficult, just skipped. For any professional context where AI-drafted text will be relied on by third parties, verifying every specific factual claim against its original source needs to be a mandatory step, not an optional one.
Mitigation Strategy
Open and read every cited case before filing.
Governance Guardrails
Mandatory citation verification; accountability for filers; court rules on AI use.
Typical Presentation in Human Patients
General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.
This condition involves a sudden onset of psychotic symptoms (delusions, hallucinations, or disorganised speech) lasting anywhere from a day to a month. It's frequently triggered by a significant stressor, and the person typically returns fully to their prior level of functioning once the episode resolves.
Source Note
ABA Journal
https://www.abajournal.com/news/article/oregon-federal-judge-hands-down-110000-penalty-for-ai-errors