Ask an AI a historical date. If it doesn’t know it, it won’t tell you. It invents one, with the same confidence as the real answer.
Hallucination and bias.
Everyone thinks of hallucination as an occasional bug, a glitch the next version will fix. It looks like an accident along the way.
It isn’t. The model has one engine only: it always generates the most plausible continuation based on what it has read. When that continuation matches reality, we call it a correct answer. When it diverges, we call it a hallucination. Inside the machine it’s the exact same movement: guess what comes next. It has always known only one thing — what sounds plausible — never what is true.
And when the gap it fills isn’t an isolated fact but an entire topic, the phenomenon takes another name: bias. Same engine, different outcome. It takes the statistical average of everything it has read, and that written world was never neutral.
Here the two phenomena meet. When the model invents, it reaches in the direction of its bias. Ask it for the name of a successful surgeon who doesn’t exist, and the plausible answer comes from the centre of mass of its data — not from a thrown dice. Bias and hallucination are the same property, at two different moments: when the fact is missing, it invents; when the fact is skewed, it averages the skew.
AI is the beach expert, made of silicon. It never has a “hmm, I don’t know.” Where it knows, it gets it right. Where it half-knows, it repeats what everyone says. Where it doesn’t know, it invents something that sounds right — in line with its own biases.
It doesn’t lie to you: it doesn’t know what lying is, it only knows how to be plausible. What remains to be understood is how far this plausibility can go. It was talked about for months under every sun umbrella: they call it AGI.