Ten citations. Zero held.
Fifteen thousand dollars
for the difference.
A citation and an invention read identically to the system that produces them — same confidence, same formatting, same tone. On July 28, an Illinois appellate court spent weeks, and fifteen thousand dollars, building the discriminator no answer engine ships with.
Start with the arithmetic a court had to do by hand. Ten citations, filed across a set of appellate briefs, had no basis in the law at all — four fabricated statutory quotations, one case that has never existed, and five cases that exist but do not say what they were cited for. Not typographical slips. Ten separate claims of support the law never gave: some invented outright, some real authority bent to say what it never said, every one dressed precisely as the law dresses, down to the font and the reporter format, and delivered with the confidence of settled authority.
The attorney’s own account is on the record. A paid subscription to a leading AI system drafted the briefs. When the fabrications surfaced, the explanation offered was inexperience — a claim the court rejected outright, noting that citation verification does not change by jurisdiction. What the record does not dispute is simpler, and more consequential: the system produced the citations, in a register indistinguishable from the real ones, and nothing inside that output ever flagged which words were load-bearing and which were invented.
Nothing in the output registered hesitation. No hedge, no qualifier, no flag of uncertainty on a single false citation: each arrived with the same typographic and rhetorical confidence as the citations that were real, because confidence, in a generative system, is a property of the sentence, not a measurement of the fact. An invention and an authority take the same amount of effort to produce, and read the same on the page. Nothing in the output tells the two apart, because the system was never asked to know the difference — only to sound like something that did.
So the court became the discriminator. Line by line, across every citation in the filing, a panel of judges did the work no part of the drafting process had done — checked every citation against the only things that settle one: the reporter, the statute book, the source itself. The court states this plainly in its opinion — it had to “painstakingly distinguish fact from fiction in every citation,” at real cost to itself and to the opposing party, because the confidence of the output gave it no reason to trust any one line over another.
The confidence tax
between a source and an invention
The condition has a name. We file it as Authority Misclassification — the first clinical characteristic of Digital Derangement Syndrome™: the failure to confidently associate a claim of authority with its true standing. A generative system, asked for an authority, classifies the real and the false identically at the point of output; both arrive dressed as authority, and nothing downstream forces the distinction before a reader trusts either.
Money does not fix this. The court’s finding was explicit: paying for a “premier” or “corporate” subscription changes nothing. A subscription tier buys fluency, nothing more. The gap between a real authority and an invented one closes only where something external to the model checks it.
Note what this scales to. A courtroom is the rare place false authority gets caught, priced, and published. When an answer engine recommends a name instead of a citation, the identical mechanism runs with nobody keeping score: the name it surfaces on deep evidence and the name it surfaces on almost none arrive in the same confident register, and no panel of judges follows.
What the correction requires
The corrective is not a better prompt, and it is not a more expensive subscription — the court has already ruled on the latter. The corrective is what Answer Engine Authority™ calls third-party corroboration: structuring an entity’s authority so that its truth does not depend on the confidence of whichever system is asked to state it. A citation with no independent verification is a sentence wearing the shape of proof. A credential that exists nowhere but a single generative output is exactly the same thing: a claim of authority trusted on the strength of its own delivery. Nobody downstream is positioned to check it before it gets acted on.
The court had thirty pages and a panel of judges to perform that discrimination once, after the fact, at cost. Most authority never gets checked at all — it simply gets believed, or it doesn’t, on the strength of how confidently the system said it.
Primary source: Scott v. Illinois Human Rights Comm’n, 2026 IL App (1st) 251462 (opinion filed July 28, 2026), read in full. ¶56: the court orders attorney Mason Cole, of Cole Sadkin, LLC, to pay $15,000 to the clerk of the Appellate Court, First District — “a $1,500 fine for each false citation and quotation,” ten in total. ¶57: “no matter how much one pays for ‘premier’ or ‘corporate’ versions of AI products, it does not negate an attorney’s obligation to verify all citations of authority.” ¶59: the court rejects the attorney’s inexperience explanation (“attorneys verify citations the same way in every court”) and separately orders the clerk to forward the opinion to the Illinois Attorney Registration and Disciplinary Commission. Full opinion, corroborated by ABA Journal.