Credit: © Unidentified / Getty Images via Canva.com Back in 1986, a Princeton philosopher named Harry Frankfurt published a short essay with a title you don’t expect to see in an academic journal, “On Bullshit.” It stayed largely inside academic circles for the better part of twenty years. Then in 2005 Princeton University Press decided to reprint it as a tiny hardcover, 67 pages, and basically small enough to fit in your back pocket. Within a month, it had sold more than 110,000 copies. A philosophy professor’s little book with a rude word title had suddenly become a runaway bestseller.
On Bullshit’s whole argument rests on a distinction most of us have never bothered to make, even though we run into both kinds of people each and every single day… A liar and a bullshitter aren’t the same animal. The liar knows the truth. That’s the key thing about them. They know exactly what’s true, and they work hard to steer you away from it.
Which means, oddly enough, the liar still respects the truth. They have to. So they keep one eye on it at all times so they know what to hide. The liar and the honest person are playing the same game, just on opposite sides.
The bullshitter isn’t playing that game at all. They don’t know what’s true and they don’t care. They’ll tell you the sky is green if green is what gets him through the conversation, and they’ll tell you it’s blue thirty seconds later if that works better. Truth and falsehood are the same to them because they aren’t aiming at either one.
They’re aiming at sounding good. That, Frankfurt argued, is what makes the bullshitter the more dangerous of the two. The liar at least admits the truth is out there somewhere. The bullshitter has simply decided that it doesn’t matter.
Now. Hold that thought, and let’s talk about the thing on your phone. Your LLM AI isn’t lying to you There’s a comforting story people tell about LLM AI hallucinations. The machine “made a mistake.” It “got confused.” It “lied.” And somehow, we’ve been conditioned to let it slide.But none of those are quite right, and the difference isn’t as easy as splitting hairs.
Here’s the very plain and simple truth of it all. A large language model doesn’t know what’s true. It was never built to. Underneath all the flash and polish it’s doing one thing, which is guessing the next word that ought to come next based on patterns that it saw in its training.
It’s about as attached to the truth as a weathervane is to the wind. It just points wherever the breeze of probability sends it, and it does it with the same smooth certainty whether it’s right or reciting complete and utter fiction. In other words, your LLM AI isn’t a liar. It’s a bullshitter.
And this isn’t some hyped up philosophy seminar. It shows up in the numbers, and the numbers are ugly. In a 2024 Stanford study, the models that were tested hallucinated in 58 to 88 percent of answers to specific, verifiable questions about randomly selected federal court cases. That’s the result from legal research questions, where convincing sounding answers could end up in a brief with somebody’s name on it.
The lawyers in Mata versus Avianca found this out the hard way. One of them used ChatGPT for legal research, and the brief that was filed with the court cited cases that didn’t exist. To make matters even worse, when the court asked for copies, the lawyers submitted purported opinions that were also fabricated. ChatGPT even went so far as to assure the lawyer that one case it invented existed in major legal databases.
It didn’t. The judge sanctioned the lawyers and their firm. That’s not lying. A liar would’ve at least known the cases were fake.
This was just pure bullshit in Frankfurt’s exact sense. The machine had no idea whether any of it was true, and more to the point, it had no way or reason to care. What it takes to actually care Here’s the thing nobody wants to say out loud. You can’t fix bullshit by making it more fluent.
A slicker bullshitter is just a bullshitter you believe for longer. And the whole LLM industry just keeps right on building bigger, smoother, more articulate versions of the same machine and acting surprised when it lies to them in ever more convincing ways. And the fix isn’t to just make a better liar, though amazingly a few have actually tried. It’s to demand an answer that shows its relationship to the evidence.
What supports it? Where does that support run out? Where has the system started reaching? Those are the things a person in any profession needs to be able to see before trusting what sounds right.
That’s the problem Vertus addresses through something it calls Metacognitive Trajectory Analysis. Vertus says it examines the course of its reasoning and identifies where support is strong and where it grows thin. The entire point is to expose where that boundary rests instead of dressing up an unsupported claim like a pig with lipstick and sending it out the door. That doesn’t make an AI infallible.
But it does give the person using it a clearer view of what deserves checking. That’s an answer to Frankfurt’s warning. The dangerous part isn’t merely getting something wrong. It’s sounding equally sure when the solid ground beneath an answer has disappeared.
And if your name is going on the brief, you need to know where that ground that you’re standing on ends. Drop it into reality We’ve spent years being wowed by how well these LLMs talk and have spent almost zero time asking the one question that really counts, and that’s whether they have the faintest idea if what they’re saying is actually true. The truth is, they mostly don’t. And an answer that’s a perfect fit because it cites a case that never existed isn’t an assistant you can ever trust to take at its word.
It’s a very expensive bullshitter with a great vocabulary and there’s no telling the damage it might do to any number of lives before anyone thinks to check. Frankfurt saw the whole thing coming, decades before any of this, in sixty-seven pocket-sized pages. The real danger was never the confident liar who knows the truth and hides it. It’s the smooth talker who stopped caring whether there was even a truth to be concerned about at all.
So next time an LLM hands you an answer that sounds just right, drop it into reality and ask the only question worth asking. Does this thing actually care whether that’s true, or does it just believe that it sounds good? Vertus was built to show where that difference matters. And that turns out to be the whole ballgame.
Sources • Princeton Alumni Weekly, “Notebook,” April 20, 2005. https://www.princeton.edu/~paw/archive_new/PAW04-05/13-0420/notebook.html • Stanford RegLab, “Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models,” Journal of Legal Analysis, 2024. https://doi.org/10.1093/jla/laae003 • U.S. District Court for the Southern District of New York, Mata v. Avianca, Opinion and Order on Sanctions, June 22, 2023. https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1%3A2022cv01461/575368/54/ Contributed article. Not produced by the TNW newsroom and does not reflect the editorial stance of TNW.
Leave a Reply