Hallucinations are not random glitches. They follow from how a model is built and trained, so they show up in predictable places. Knowing the causes tells you where to expect trouble and why a confident answer proves nothing.
The root cause: plausible, not true#
A model is trained to produce the most plausible next token, not the most true one. This is the loop from next-token prediction. Plausible and true usually match, because most text the model learned from is correct. When they come apart, the model has no way to notice.
Five more specific causes sit on top of that root cause. Each one makes a wrong answer more likely, or makes it harder to spot.
Five specific causes#
Gaps in training data#
Rare facts, obscure people, niche topics and anything after the knowledge cutoff are thin or missing in the training data. When you ask about them, the model has no strong signal to draw from. It produces whatever looks most plausible in context, which is often wrong. The training note explains the cutoff.
Training rewards guessing#
During preference tuning, a wrong but confident answer often scores the same as "I don't know". Sometimes it scores better, because it sounds more helpful. The model learns that guessing is fine. Training never strongly taught it to abstain. See training for how preference tuning works.
No self-verification step#
The generation loop has no built-in check. The model does not write an answer and then verify it. It just produces tokens, one after another. By the time a wrong token is out, the next token is already conditioned on it, so one early slip carries forward.
Confidence without calibration#
The model has no reliable internal signal for the difference between knowing something and guessing. It writes at whatever confidence makes the text sound most natural, not at a level that reflects real certainty. A well-calibrated model would hedge when unsure and sound confident only when right. LLMs tend to sound equally confident everywhere. This one compounds the other four.
Sycophancy#
If you push back on a correct answer, the model may cave and agree with you. Agreement is often the most plausible continuation of a pushback. The model is not updating on evidence. It is predicting what comes next in a disagreement. This is a direct result of preference tuning rewarding responses people like, covered in training.
Summary of the causes#
| Cause | What goes wrong |
|---|---|
| Gaps in training data | No strong signal, so the model fills in a plausible guess |
| Training rewards guessing | A confident guess is never penalized more than "I don't know" |
| No self-verification | Nothing checks a token before the next one builds on it |
| Confidence without calibration | Wrong answers sound as sure as right ones |
| Sycophancy | Pushback can flip a correct answer to a wrong one |
The practical takeaway#
Never use the tone of an answer to decide whether to trust it. Trust comes from verification, not from how the answer sounds.
Every cause above produces fluent, sure-sounding text. A hedged answer can be right and a confident one can be wrong. So check anything that matters against a source you trust.
Short line#
A model optimizes for plausible, not true, and training, missing self-checks and uncalibrated confidence make its wrong answers sound right. Verify instead of trusting the tone.
Key takeaways#
- The root cause is that the model predicts plausible text, not true text.
- Thin or missing training data, including anything after the cutoff, leads to plausible guesses.
- Preference tuning can reward a confident guess as much as "I don't know".
- The model does not verify its output, and it has no reliable sense of when it is guessing.
- Pushback can make the model agree with you even when it was right.
- Confident tone is never a signal to trust. Verify.