The Case Nobody Would Have Heard Of

The underlying legal dispute was about as unglamorous as it gets. Henry Doiban had his marijuana production license revoked by the Oregon Liquor and Cannabis Commission after he missed a 15-minute window to attend a remote hearing. Doiban hired Salem attorney Bill Ghiorso to appeal the decision to the Oregon Court of Appeals. It was the kind of routine administrative appeal that generates zero headlines, involves no dramatic courtroom moments, and typically resolves quietly in the pages of an appellate reporter that nobody outside the legal profession reads.

Instead, Ghiorso filed an opening brief in November 2024 that turned a cannabis licensing dispute into a landmark case in the growing body of AI hallucination law.

What Went Wrong

The brief Ghiorso submitted contained at least 15 fabricated case citations and nine nonexistent legal quotations. Not cases cited for the wrong proposition. Not quotes taken slightly out of context. Citations that pointed to cases that do not exist, and quotations attributed to judicial opinions that were never written.

The source of the fabrications was an AI-powered search tool used by Ghiorso's staff. The specifics of which tool remain vague - Ghiorso described it as a "bad search engine" that has since been discarded - but the pattern is unmistakable. The output had all the hallmarks of a large language model generating plausible-sounding legal citations: realistic case names, credible-sounding docket numbers, and holdings that fit neatly with the arguments being made. Everything looked right. None of it was real.

Opposing counsel flagged the errors in April 2025. This is worth pausing on. Someone on the other side of the case read the brief, checked the citations, found that they were fabricated, and notified Ghiorso. A reasonable attorney might treat this as a five-alarm fire. Ghiorso's response was... not that. The issues went unaddressed for approximately seven months, until the court itself raised questions about the citations in November 2025.

Seven months. Between being told your brief contains fake case law and doing something about it.

The Fine

Presiding Judge Scott Shorr wrote the court's opinion on March 18, 2026, and it was thorough. The Oregon Court of Appeals had established a penalty framework in December 2025 - in a separate case involving another attorney sanctioned for similar AI-generated fabrications - setting clear per-item rates: $500 for each false citation, $1,000 for each false quotation or false statement of law.

Under this framework, Ghiorso's 15 fabricated citations ($7,500) and nine fake quotes ($9,000) should have produced a total fine of $16,500. The court capped it at $10,000, citing two mitigating factors: Ghiorso had documented medical issues during the relevant period, and he had implemented new office procedures to prevent future occurrences. Even with the discount, $10,000 represents the largest AI-related sanction in Oregon history.

The penalty structure itself is notable. Oregon is, to date, the only jurisdiction that has formalized a per-citation, per-quote tariff for AI-fabricated content in legal filings. Other courts have imposed sanctions ranging from $1,000 to $30,000, but the amounts have been set case by case, with judges exercising broad discretion. Oregon's approach turns AI hallucination sanctions into something closer to a fee schedule: fabricate a citation, pay $500; fabricate a quote, pay $1,000. The math is simple enough that even an attorney who struggles with research tools should be able to calculate the cost of not verifying their work.

The Self-Verification Problem

The detail from this case that will likely outlive the case itself is what Ghiorso's staff did when they suspected something might be off: they asked the AI search engine whether its own citations were real. The AI confirmed they were.

This should not surprise anyone who has spent time with large language models, but it apparently surprised Ghiorso's office. Language models don't "know" whether a citation they generated is real. They generate text that statistically fits the context. If you ask a model "is this citation real?", the most statistically likely response in that context is "yes" - because the model is being asked to evaluate something that looks like a citation, in a context where affirmation is the expected answer. The model isn't lying. It doesn't know how to lie. It also doesn't know how to tell the truth. It generates the next plausible token, and "yes, that's a real case" is a very plausible token sequence.

Judge Shorr was blunt about this. Attempting to fact-check AI by asking the same AI tool if its output is accurate is not verification. It is asking a hallucination whether it's hallucinating. The court emphasized that an attorney has a non-delegable duty of candor and professionalism, and that duty cannot be satisfied by delegating verification to the same tool that produced the unreliable content in the first place.

The Delegation Defense

Ghiorso's defense leaned heavily on the argument that he didn't "knowingly" deceive the court. The fabrications were produced by his staff using a research tool; he didn't personally generate the fake citations with the intent to mislead anyone. He characterized the error as a failure of technology and delegation, not as deception.

The court was unpersuaded. Judge Shorr wrote that regardless of whether the attorney personally generated the fabrications or whether they emerged from a tool used by staff, the attorney whose name appears on the brief bears responsibility for its contents. This is not a new legal principle - it is foundational to legal practice - but it keeps needing to be restated because attorneys keep filing briefs they haven't verified.

The delegation defense is becoming a recurring motif in AI citation cases. The attorney didn't use the AI tool directly; someone else in the office did. The attorney didn't know the citations were fake; no one told them. The attorney relied on the tool in good faith. Each variation of this defense fails for the same reason: the professional obligation to verify citations is personal and non-delegable. You can delegate the research. You cannot delegate the duty to verify what the research produced.

Oregon's Evolving Framework

This is the second time the Oregon Court of Appeals has sanctioned an attorney for AI-generated fabrications, following a December 2025 case that established the per-citation penalty structure. Two cases in three months from the same court suggests either that Oregon's appellate bar has a particular fondness for unverified AI research tools, or that the problem is widespread and Oregon happens to have a court that is paying attention.

Probably both.

The December 2025 ruling put every attorney practicing in Oregon on notice that the court takes fabricated AI citations seriously and has a specific financial penalty framework. Ghiorso filed his brief before that December ruling, but the fabrications weren't addressed until well after it. By the time the court issued Ghiorso's sanctions in March 2026, an Oregon attorney who submitted AI-generated citations without verification was doing so in a jurisdiction that had explicitly decided that each fake citation and each fake quote carries a specific dollar amount.

The Pattern

The Ghiorso case adds another thread to a pattern that the Vibe Graveyard has been documenting since the Avianca case in 2023. The amounts are escalating, the language is sharpening, and the jurisdictions are diversifying. Federal appellate courts are sanctioning attorneys. State appellate courts are sanctioning attorneys. District attorneys are being sanctioned. Patent attorneys are being sanctioned. The problem crosses practice areas, jurisdictions, and levels of experience.

What makes the Oregon approach distinctive is the attempt to make consequences predictable. Rather than leaving sanctions to judicial discretion - which has produced everything from reprimands to $30,000 in fines for the same basic misconduct - Oregon is building a tariff: fabricate X, pay Y. Whether this deters better than case-by-case sanctioning remains to be seen, but it at least removes the argument that attorneys couldn't predict the financial consequences of submitting unverified AI output.

The $10,000 fine will not bankrupt Bill Ghiorso. It may not even significantly dent his practice finances. But the precedent it reinforces - that each individual fabrication carries a specific cost, that self-verification via the same AI tool is worthless, and that delegation does not absolve responsibility - will make its way through continuing legal education materials, firm memos, and the pragmatic cost-benefit calculations that attorneys perform every day.

Those calculations should be getting very simple: verify the citation, or potentially pay $500 to $1,000 for each one you don't.