What Is an AI Hallucination?

The popular story is that an AI hallucination is a malfunction — the model "lied," "made a mistake," or "broke." That story is comforting and wrong, and believing it will get you burned.

An AI hallucination is a confident, plausible-sounding statement that is false or fabricated — and it happens not because the model failed, but because it is a text-prediction system with no built-in mechanism for knowing whether what it says is true. The same process that produces a correct answer produces a hallucinated one. There is no separate "lying mode" that switched on.

The consensus view (and what's right about it)

Let me steelman the mainstream take before dismantling it. People who call hallucinations a bug aren't being foolish. The output really is wrong. It really can cause harm — fake legal citations, invented statistics, a confidently mis-summarized document. And vendors really are working to reduce them, which implies they're undesirable. All of that is true. If "bug" just means "behavior we don't want," fine.

But the word smuggles in a second claim: that the model normally tracks truth and occasionally slips. That part is the dangerous fiction.

How it actually works

1. The model predicts text, not facts

A large language model — the technology behind tools like ChatGPT or Claude — works by predicting the next most likely chunk of text given everything before it. (For a fuller treatment, see [How Large Language Models Work](how-large-language-models-work.md).) It learned this from a vast amount of human writing. What it optimizes for is plausibility — text that looks like a good continuation — not accuracy. Most of the time plausible and true overlap, which is why these tools are useful. When they diverge, the model has no internal alarm. It cannot tell the two apart, because it was never measuring truth in the first place.

2. There's no "I don't know" by default

A human expert who hits the edge of their knowledge feels the gap and hedges. A base language model has no such signal. Asked for a citation that doesn't exist, it will generate one — author, title, year, page — because a citation-shaped string is the plausible continuation of "the source is…". Fluency is not evidence. The hallucination sounds exactly as confident as the truth because confidence, here, is just the smoothness of the text.

3. Hallucination is the same machinery as success

This is the part the "bug" framing hides. The mechanism that correctly tells you the capital of France is the mechanism that invents a court case. There's no clean line in the model separating the two. You cannot patch out hallucination without touching the very process that makes the model work — which is why the problem is reduced with techniques and guardrails, not eliminated.

A concrete example

Ask a model for "three peer-reviewed studies showing that standing desks improve memory," and you may get three crisp entries: plausible author names, a real-sounding journal, a year, a DOI. Check them, and the papers don't exist.

The model didn't decide to deceive you. You requested the shape of three citations, and it produced the most likely text of that shape. It had no database to check against and no instinct that the answer should be "I'm not aware of strong evidence for that." It gave you a convincing forgery because forgery and fact come off the same press.

Why this matters

If you believe hallucination is a rare glitch, you'll trust the model 95% of the time and get ambushed by the other 5% — precisely when the answer was fluent and you had no reason to doubt it. If you understand that the model is always predicting plausibility and never verifying truth, your posture flips: the output is a fast, useful draft to be checked, not an authority to be quoted. The honest caveat cuts both ways — newer systems that cite real sources or search the web genuinely hallucinate less, and on common, well-documented facts the risk is low. But "lower" is not "off," and you rarely know in the moment which kind of answer you got.

The practical implication of being right about this: treat verifiability, not fluency, as your signal. Trust AI output for things you can check or things that don't need to be exactly true (brainstorming, rephrasing, first drafts). Verify anything where being wrong has a cost — names, numbers, quotes, citations, anything load-bearing.

Try this

The next time a model hands you a fact that matters, do one thing before you use it: ask it to point to a source you can independently open and confirm. If it can't — or the source evaporates when you check — you've caught a hallucination in the act, and you've learned how this tool actually behaves.

And keep your own ground truth. The reason a [second brain](../ai-notetaking/how-to-build-a-second-brain.md) of notes you wrote and verified is so valuable in the AI era is precisely this: the model predicts what's plausible, but your own captured, checked knowledge is what's true. Jot what you confirm into JustJot.ai, and let the AI draft against a record that doesn't make things up.

The rule: a model never tells you it's unsure — so the certainty has to come from you.