why LLMs hallucinate

The root cause sits in the training objective. A model learns by predicting the next token over and over, rewarded for matching the text it was trained on. Nowhere in that loop is there a step that asks is this true. The model learns the shape of true-sounding language — its rhythm, its vocabulary, its confident tone — without learning a separate, reliable map of which statements actually hold. Truth and plausibility get tangled together.

Two more pressures push it along. The model has no built-in way to signal I do not know, because saying nothing was rarely the right answer during training, so it fills the gap. And later alignment steps can reward answers that please the user, which quietly teaches the model that a smooth, helpful-sounding guess beats an honest blank. Put together, hallucination is not a glitch bolted on top — it is a direct shadow of how these systems are made.