Hallucination Hunter.

Why AI makes things up — and how to catch it

An AI hallucination is a statement that is confidently worded, grammatically perfect, plausible on its face — and completely false. It is not a bug that will be patched away, and it is not the model lying. It is a direct consequence of how language models work, which is why learning to spot hallucinations is more useful than waiting for them to stop.

Where hallucinations come from

A large language model does not store facts in a database and look them up. It predicts the next chunk of text, one piece at a time, based on statistical patterns absorbed from an enormous amount of training data. Knowledge ends up distributed across billions of numeric weights rather than filed in retrievable records.

That design produces fluent, flexible language. It also means the model has no internal sense of the difference between recalling something and inventing something. A fabricated citation is generated by exactly the same process as a real one: both are simply text that fits the pattern. The model is not choosing to deceive you; it has no mechanism for knowing it is wrong.

This is why hallucinations are so dangerous in practice. A human who is unsure usually sounds unsure. A language model that is completely wrong sounds exactly like a language model that is completely right.

Five red flags worth memorising

• Suspiciously round or dramatic numbers. “Reached a million users in 24 hours” is the kind of statistic that gets repeated because it sounds good. The real figure was about five days. Near-miss numbers are the hardest hallucinations to catch because they are almost right.
• Absolutes. Words like completely bans, always, never, and the first are rare in reality and common in generated text. Regulations restrict and carve out exceptions; they seldom ban outright.
• Named sources you cannot immediately verify. Invented paper titles, journal names, case numbers, and author lists are the single most documented category of AI hallucination. If a citation matters, open it.
• Famous names attached to old inventions. Models drift toward well-known entities. The first neural network machine, the perceptron, was built by Frank Rosenblatt in 1958 — decades before Google existed — but “built by Google” is the kind of attribution that sounds unremarkable until you check.
• Confident answers about very recent events. Every model has a knowledge cutoff. Beyond it, the model does not know that it does not know.

The myth this game exists to kill

Many people assume hallucinations only happen at the edges — obscure topics, or events after the training cutoff. They do not. Hallucinations occur on mainstream subjects that are thoroughly represented in the training data, because the failure is in how text is generated, not in what the model has seen. Treating “well-known topic” as a safety guarantee is exactly how confident errors slip into finished work.

What the twelve rounds cover

Each round presents three statements about one topic. Two are accurate; one is a realistic hallucination of the kind models actually produce. The topics span acronyms, AI history, transformer architecture, model training, how LLMs work internally, hallucination itself, image generation, statistics, AI ethics and regulation, tokens, fine-tuning, and AI agents. After every round you are shown which statement was false and precisely why — so a wrong answer teaches you as much as a right one.

Why this skill is worth building

Verification is becoming the differentiating professional skill around AI. Anyone can prompt a model; far fewer people reliably notice when the answer is wrong. The habit is simple and learnable: check numbers, distrust absolutes, open citations, and be most careful when the output is most fluent. Confidence in generated text carries no information about accuracy.

For the full treatment — including bias, overconfidence, and the other ways AI systems fail — work through Module 06: When AI Gets It Wrong, or look up any unfamiliar term in the AI glossary.