Here is the single most useful reframe for understanding modern AI: a model never knows anything. It assigns likelihoods.
When a language model continues the sentence "The capital of France is…", it doesn't retrieve a fact. It computes a probability for every word it could say next: Paris 0.97, Lyon 0.01, the 0.005, and so on, then picks from that spread. The right answer being overwhelmingly likely is what correctness looks like from inside the machine. There is no separate compartment holding the truth.
Living with likelihoods
- Probabilities describe long runs, not single events. "70% chance of rain" doesn't mean it will mostly rain. It means: across many days like today, seven in ten get wet. One dry afternoon doesn't make the forecast wrong.
- Calibration is the virtue to ask about. A system is well-calibrated if, of all the times it says "80% sure," it's right about 80% of them. A model can be impressively accurate and badly calibrated — confidently wrong in exactly the cases that matter.
- Uncertainty is information. A model that says "I'm not sure" is handing you something valuable. Most chat interfaces sand this off, presenting the top guess in the same warm prose whether it was a 0.97 or a 0.4.
Why this explains so much
Once you see the likelihood machinery, several mysteries dissolve. Why do models sometimes give different answers to the same question? They're sampling from a distribution. Why do they make things up? Because some continuation always has the highest probability, even when no continuation is true. The machine is doing exactly what it was built to do — the surprise was ours, for mistaking fluency for knowledge.
What to read next
Likelihood tells you two things move together. It cannot tell you one causes the other — that trap gets its own essay: correlation vs causation.