Modern AI understands the world.
Every mystery on this site is an original murder puzzle, written after June 2026, so none of them sit in any AI’s training data. Each one also runs on invented physics — candles bound to a person’s name, sound that moves only through stone. Rules that don’t hold in our world, that no AI has ever read about.
That combination is the point. The AI can’t recall an answer, because none exists. It can’t lean on how our world works, because this world works differently. To name the killer it has to take rules it was handed seconds ago, follow them through the clues, and work out what must have happened. The solving is genuinely new each time it runs.
So when a system that supposedly does nothing but predict the next token solves one of these, there is only one explanation left: it built a model of the world it was given and reasoned inside it. That is what understanding is, and what follows is the proof.
- Premise 1
To understand something is to build a working model of how it behaves, then use that model to reach answers nobody handed you. That is the whole definition here, and it has nothing to do with awareness or feelings. The only question is whether you can work out what must be true from the way the pieces fit.
- Premise 2
A system with no such model cannot solve a fresh mystery whose answer is written nowhere in its training. There is nothing to recall and no way to reason it out, so nothing is left to produce the answer. This is the skeptic’s own position, said plainly: a pure predictor of text should be stuck here.
- Premise 3
Modern AI does solve these mysteries. They are written after its training cutoff, set in worlds whose rules don’t hold in ours, and it still names the killer and shows every step it took to get there. You can run them yourself and watch it reason.
A system with no model could not solve these. The AI does. So it has a model and it uses it, and reasoning from a model to reach what you were never told is exactly what understanding means. So the AI understands.
Why it holds
The argument stands or falls on Premise 2, and Premise 2 is the skeptic’s own belief: a token-predictor will fail at a brand-new puzzle. Premise 3 shows it doesn’t fail. A belief that predicts the wrong result is false, so the skeptic’s belief is false. That leaves three moves, and each one fails. Say the AI doesn’t really solve these, then open the Scenarios and check for yourself. Grant that it reasons past pattern-matching, which was the whole point. Or argue about what “understand” means until the AI no longer counts, which just changes the subject.