Why a large language model that tells you the truth still doesn't know anything, and why you don't either.
The Definition on the Box
For centuries, the standard definition of knowledge was simple and elegant: a Justified True Belief (JTB). To know something, you had to believe it, it had to be true, and you had to have a good reason—a justification—for believing it. It’s an intuitive model that seems to cover all the bases. If you believe it’s raining, you look outside and see rain falling, and it is in fact raining, you know it’s raining. Simple.
Then, in 1963, a three-page paper by a philosopher named Edmund Gettier dismantled 2,400 years of consensus. He did it by showing that you could have a justified, true belief that still, intuitively, wasn't knowledge.
The Man With Ten Coins
Gettier’s cases all share a common structure. Consider one of his classics: Smith and Jones are up for the same job. Smith has excellent evidence that Jones will get the job; the company president told him so. Smith also happens to know that Jones has ten coins in his pocket. From this, Smith forms a belief:
The man who will get the job has ten coins in his pocket.
As it turns out, the president was mistaken. Smith, not Jones, gets the job. And, unknown to himself, Smith also happens to have ten coins in his pocket. Now, let’s check the conditions. Is Smith’s belief true? Yes, the man who got the job (Smith) has ten coins in his pocket. Is it justified? Yes, his justification stemmed from the president’s testimony and his counting of Jones’s coins. He had good reason to believe what he did.
So Smith has a justified, true belief. But does he know it? Clearly not. His belief is true by sheer, dumb luck. The thing that justified his belief (the evidence about Jones) came completely apart from the thing that made his belief true (the facts about himself). The philosophical term for this, coined by Duncan Pritchard, is veritic luck.
The Gettier Problem is an Architectural Feature of LLMs
This brings us to large language models. A belief you acquire from a conversational AI has the Gettier structure built into it by default. When you ask an LLM a question and it gives you a factually correct answer, you have acquired a true belief. Your justification is that a powerful AI, trained on the sum of human text, gave you the answer. But the process is rotten with veritic luck.
An LLM’s mechanism is next-token prediction. It generates text by calculating the most statistically probable continuation of a prompt, based on patterns in its training data. It is not, and cannot be, a process that tracks the state of the world. Its architecture connects it to a vast corpus of text about the world, not to the world itself. When its output happens to be true, it's because the statistical patterns in the text it was trained on happened to align with reality.
The justification (the model's prediction) and the truth-maker (the actual state of affairs) are only contingently connected. The model isn’t reasoning or checking facts; it’s pattern-matching at an astronomical scale. Therefore, any true belief you get from it is Gettiered. It's true by accident.
This is why reliability is the only metric that matters. If the underlying process is not tethered to the world, any truth it produces is epistemically worthless. It isn't knowledge, for the model or for you. It’s a lucky guess, and building a strategy on lucky guesses is just a more sophisticated way to fail.
Method note. This piece was built in NeuraScene's Advertising Mode using an explicit psychological frame:
- Persuasion trigger: Social Proof
- Target affect: Cognitive Ease
- Gestalt principle: Law of Continuity
The article targets a technically literate audience by deconstructing a complex philosophical problem (the Gettier problem) and applying it to AI, creating Cognitive Ease through clarity. This display of expertise builds Social Proof for the agency. The Law of Continuity is used in the key art to create a visual metaphor for the argument: a path that seems correct but is disconnected from the true destination.
