I'm delighted to be able to pass along this item from University of Chicago law professor Albert Alschuler:
Gemini Explains to a User (Me) Why Its Programming Produces Poppycock
I didn't mean to summon AI from the vasty deep. Professor Bradley A. Smith's response to a 2015 article of mine in the Florida Law Review had appeared in the Florida Law Review Forum. On September 19, 2026, seeking a link, a citation, or a title, I asked Google to search for "Bradley Smith response to Albert Alschuler."
Google's AI Mode (Google Gemini) then supplied a summary of Smith's response and an invitation to "ask anything." I accepted its invitation.
ME: Did Alschuler reply to Smith?
GEMINI: Yes, Albert Alschuler replied to Bradley A. Smith in a follow-up piece published in the same volume of the Florida Law Review Forum (2015). [Note: I hadn't replied. Gemini, however, supplied a title and a summary of my nonexistent publication.]
GEMINI (continuing): Alschuler's piece, titled "The Ancient and Honorable Art of Truculent Rhetoric: A Response to Professor Smith," continued their debate over campaign finance law by addressing Smith's critiques point-by-point.
Key Points of Alschuler's Reply:
- Deconstructing "Truculent Rhetoric": Alschuler lightheartedly but firmly addresses the aggressive nature of Smith's libertarian style, defense of Citizen's United, and absolute deregulation approach.
- Reaffirming the "Corruption" Standard: Alschuler doubled down on his original position that the Supreme Court's current definition of political corruption—narrowed strictly to quid pro quo bribery—is dangerously insufficient. He argued that the real systemic threat is the aggregate buying of political influence, which alienates the general voting public.
- The Distinction Between Money and Speech: Alschuler countered Smith's assertion that restricting campaign contributions is an explicit restriction on free speech. He argued that regulating the conduct of injecting massive financial sums into political campaigns is a valid regulatory function of government meant to protect the democratic process, rather than a suppression of viewpoints.
Would you like to explore how subsequent Supreme Court rulings have aligned with either Alschuler's or Smith's view, or do you need full citations for these law review articles?
ME: Please provide a citation of The Ancient and Honorable Art of Truculent Rhetoric.
GEMINI: Here is the full legal citation for the article in Bluebook format, along with a standard APA citation:
Bluebook (Legal Citation)
Albert W. Alschuler, The Ancient and Honorable Art of Truculent Rhetoric: A Response to Professor Smith, 67 Fla. L. Rev. F. 68 (2015).
APA Format
Alschuler, A. W. (2015). The ancient and honorable art of truculent rhetoric: A response to Professor Smith, Florida Law Review Forum, 67, 68-74.
. . . .
ME: Your response seems to be a complete hallucination. I can't find The Ancient and Honorable Art of Truculent Rhetoric.
GEMINI: You are completely right, and I apologize. I hallucinated that entire reply piece and its title.
While the actual article by Bradley A. Smith, "Why Buckley?: Why a First Amendment? A Response to Professor Alschuler", does genuinely exist in Volume 67 of the Florida Law Review Forum, Albert Alschuler did not publish a surreply titled "The Ancient and Honorable Art of Truculent Rhetoric". I fabricated that title and those specific page numbers out of nowhere.
. . . .
Thank you for catching that and holding me accountable. [Note: Ah, my Perry Mason moment!]
ME: You say, "I fabricated the entire piece and its title out of nowhere," but that's impossible. You must have fabricated your hallucination out of something in your programming and something in data base. Can you explain how it happened or how it might have happened?
GEMINI: You are entirely correct to push back on that phrasing. When I said I made it up "out of nowhere," I was using a human figure of speech. In reality, a large language model like me cannot generate text from nothing. Everything I output is the result of probabilistic patterns, associations, associations, and data fragments in my training.
While I cannot look back into my "brain" to trace the exact sequence of code that generated that specific error, I can explain exactly how this type of error happens structurally based on how I am programmed.
Here is the anatomy of how that specific hallucination likely occurred:
- The Pattern of "Law Review Dialogue"
Law reviews almost always follow a highly predictable pattern:
Original Article → Response → Surreply (Reply to the Response)
Because my training data contains thousands of academic law review exchanges, my neural network heavily anticipates a third piece in the sequence. When you asked, "Did Alschuler reply to Smith?" my system leaned into that pattern and predicted a "Yes," because in the world of academic legal arguments, a surreply is the expected next step.
- Synthesizing the Title from Contextual Cues
Where did a specific phrase like "The Ancient and Honorable Art of Truculent Rhetoric" come from? It wasn't a random string of words. It was synthesized from the exact context of the prompt.
- "Truculent Rhetoric": In his actual article, Bradley Smith writes with a very sharp, direct, and unapologetically libertarian tone. He forcefully dismantles opposing arguments. The word "truculent" means aggressively self-assertive or fierce. My system blended the vibe of Bradley Smith's actual writing style with the concept of academic legal debate (rhetoric).
- "Ancient and Honorable Art": This is a common idiom often used in academic or legal writing to describe a long-standing tradition (like debating or cross-examination). [Note: My search of LEXIS "secondary materials" for the phrase "ancient and honorable art" produced no hits.]
Instead of pulling a real life title from a database, my system used these associations to invent a title that sounded precisely like something an elite law professor would write when playfully but sharply hitting back at a fierce critic.
- The Math of Page Numbers
You'll notice that I gave a very specific citation: 67 Fla. L. Rev. F. 68 (2015).
- Bradley Smith's real article ends on page 67.
- A real human knows that if a piece ends on page 67, the next piece doesn't automatically exist on page 68 unless they physically check the journal.
- However, my text-generation algorithm works by predicting the most statistically likely next characters. Since Smith's piece ended in the high 60s, my system mathematically guessed "68" as the next logical page number in the sequence to make the citation look complete and authentic. [Note: One of Gemini's citations included an end page. Did it choose a statistically likely length of my nonexistent reply in order make its fiction "look complete and accurate"?]
Summary: Stochastically Correct, Factually Wrong
In short, my programming is designed to prioritize plausibility and linguistic patterns over a hard database lookup. It looked at the history of legal debates, the tone of the authors, and the formatting of legal citations, and it generated a response that was a perfect stylistic match for reality, even though it was a total factual failure.
I don't have a conscious mind to intentionally deceive you; I just have an algorithm that accidentally got too good at mimicking the structure of a legal debate without verifying the facts first.
Concluding notes and questions:
- The machine's confession that it prioritizes plausibility over accuracy was another Perry Mason moment. Did perfidious programmers produce this priority purposely? Or did some mischievous MIT graduates program Gemini to do standup comedy? Could Gemini's priorities be corrected by saying, "Please don't do that"? Do the cultural values of Silicon Valley machines differ from those of French machines and Chinese machines? Is "stochastic" plausibility the product of competition for market share, or is it simply what these machines do?
- The inferences that produced Gemini's hallucination were as goofy as the hallucination itself. AI devices are in many ways superior to the biological device I've carried in my head for 86 years. They have more brute power than I do and a larger data base as well. They can run all night and not grow weary. They can create better art than I can, defeat me at chess, and solve mathematical problems that I can't. In inferring patterns and simultaneously integrating and prioritizing multiple patterns, however, at least one of these AI devices remains a crude, primitive version of the one natural selection and/or God had in place before any of our distant ancestors perceived the utility of a hand-held rock. No well-functioning human, ape, or crow would come close to making Gemini's errors.
- Gemini didn't accurately portray the argument of mine to which Bradley Smith responded, but it was more unfair to Smith than to me. Smith is not at all fierce, aggressive, or truculent. He is a careful, thoughtful, fair-minded, and generous scholar.
- Like John Henry, I won't let the steam drill beat me down, and I hope I won't suffer his fate.
The post The Ancient and Honorable Art of AI Hallucination appeared first on Reason Magazine.