A hallucination is when a language model produces text that is fluent, confident, and wrong — a fabricated citation, an invented API, a plausible-sounding fact that simply is not true. The unsettling part is that the model gives no signal that it is guessing; the false answer arrives in exactly the same assured voice as a correct one. It is tempting to treat this as a bug to be patched, but hallucination is better understood as a direct consequence of what a language model is: a system trained to predict plausible next tokens, not to retrieve verified facts. Once you see it that way, hallucination stops being surprising and becomes something you manage rather than something you expect to disappear.\n\nHallucination falls out of the training objective itself: models are optimized for plausibility, not truth. A language model learns by predicting the next token over enormous amounts of text, and the thing it gets rewarded for is producing continuations that look like the training data — fluent, well-formed, statistically likely. Nothing in that objective is a truth signal. The model has no separate database it looks facts up in; knowledge is smeared across billions of weights as a lossy compression of what it read, so when it needs a specific detail it never memorized — a date, a court case, a function name — it does what it always does: generates the most plausible-looking token sequence. Often that sequence happens to be correct; when it is not, the same machinery produces a confident fabrication, because the model cannot tell the difference between recalling and inventing.\n\nThe causes are several, and it helps to separate hallucinations that contradict a source from those that are merely unverifiable. Fabrications arise from parametric gaps (facts the model never learned or learned wrong), outdated knowledge (the world changed after training), pressure to answer (instruction-tuning and RLHF can teach a model to always produce a helpful-sounding response rather than admit ignorance, a cousin of sycophancy), decoding randomness (sampling can select a wrong-but-fluent path), and ambiguous prompts the model resolves by guessing. A useful distinction: intrinsic hallucinations contradict a source you provided (the model misreads the document), while extrinsic hallucinations add claims that cannot be checked against any source at all. The two call for different defenses.\n\nMitigations reduce hallucination substantially but cannot fully eliminate it, because the root cause is architectural. The most effective lever is grounding: give the model the facts at inference time via retrieval-augmented generation, so it summarizes real documents instead of reciting from lossy memory, and ask it to cite. Others include verification passes (a second model or tool checks claims), lowering decoding temperature for factual tasks, training the model to abstain and say "I don't know" when uncertain, and better calibration so confidence tracks correctness. None of these is a cure — a grounded model can still misread its sources, and a well-calibrated one is still sometimes confidently wrong — which is why the practical stance is defense in depth plus human verification for anything that matters, not the expectation of a hallucination-free model.\n\n| Cause | What goes wrong | Main defense |\n|---|---|---|\n| Parametric gap | Fact never learned / learned wrong | Retrieval / grounding (RAG) |\n| Outdated knowledge | World changed after training | Live retrieval, tools |\n| Pressure to answer | Guesses rather than abstains | Train to say "I don't know" |\n| Decoding randomness | Fluent-but-wrong path sampled | Lower temperature, verify |\n| Ambiguous prompt | Model guesses your intent | Clarify, constrain, cite |\n\n``svg\n\n``\n\nThe unhelpful way to think about hallucination is as a defect that a better model or one more patch will finally fix. The useful way is to start from the mechanism: a language model is a plausibility engine that predicts likely text from a lossy, compressed memory, with no internal oracle that distinguishes a fact it recalls from a fact it invents — so a confident fabrication is produced by exactly the same process as a confident truth. That reframing tells you why hallucination happens (the objective rewards fluency, not accuracy), why it clusters where it does (missing or stale knowledge, pressure to answer, sampling, ambiguity), and why the fixes are all forms of external support — grounding it in real documents, verifying its claims, teaching it to abstain — rather than an internal switch. Read hallucination through a predicts-plausible-text-not-retrieves-facts lens rather than a buggy-database lens, and you stop waiting for it to vanish and start building the guardrails that keep it from mattering.