Hallucinations: When AI Confidently Lies

The Confident Wrongness Problem

Ask an LLM for a fact it doesn't know, and something strange happens: instead of shrugging, it often invents an answer that is fluent, specific, formatted beautifully, and completely made up. This is a hallucination: output that sounds plausible but isn't true.

The unsettling part isn't that AI makes mistakes. It's the delivery. Hallucinations arrive with the same confident tone as correct answers. No stammering, no "I think maybe...?" Just polished, authoritative fiction.

Why Models Make Things Up

Remember Lesson 3: an LLM generates text by predicting plausible next tokens. Read that word carefully. Plausible. Not verified. Not true. Plausible.

The model learned what factual answers look like (their shape, tone, and formatting) from millions of examples. So when you ask about something it doesn't actually know, the machinery does exactly what it was built to do: produce text shaped like a correct answer. A citation-shaped citation. A statistic-shaped statistic. The prediction engine has no built-in fact-checking department; truth and truthiness feel identical from the inside.

A Hallucination, Dissected

Here's a demo of the failure mode. The book below does not exist. Everything in this output is fabricated, so watch how real it looks:

USER: Tell me about the 1962 novel "The Cartographer's Daughter"
by Margaret Ellsworth.
AI:   "The Cartographer's Daughter" (1962) is Margaret Ellsworth's
acclaimed third novel, following a young mapmaker's daughter
in colonial Singapore. It won the Hawthorne Prize in 1963,
sold over 2 million copies, and was adapted into a 1968 film
starring Julie Andrews. Critics praised its "luminous prose"
(The New Yorker, March 1963).

Look at those moves: a precise year, a named prize, a sales figure, a film adaptation, a quoted review with a date. Specificity feels like credibility. That's exactly what makes hallucinations dangerous. The model wasn't lying maliciously; it was asked to complete text about a book, so it generated the most book-summary-shaped text it could.

This Has Real Consequences

In 2023, a lawyer famously submitted a legal brief written with ChatGPT's help. It cited multiple court cases as precedent. The cases did not exist, complete with realistic case numbers, judges, and quotes. The court was not amused; sanctions followed, and "lawyer cites fake AI cases" became a recurring genre of news story.

The lesson isn't "never use AI." It's "never skip verification when the stakes are real."

How to Catch a Hallucination

Your field guide:

1. Be suspicious of specifics you can't trace. Names, numbers, dates, quotes, citations. The more precise and the less checkable, the more scrutiny it deserves.
2. Ask for sources, then actually check them. Hallucinated sources usually crumble on contact: the link 404s, the paper doesn't exist, the book has no ISBN.
3. Cross-check independently. Verify claims against a search engine or trusted reference, not by asking the same model "are you sure?" (it can double down politely).
4. Know the danger zones. Obscure topics, niche people, exact statistics, legal and medical citations, and events after the model's knowledge cutoff are prime hallucination habitat.
5. Scale your paranoia to the stakes. Brainstorming taco names? Whatever, vibes only. Legal filing, medical decision, journalism, anything with your name on it? Verify every factual claim.

There's also a better way to ask. Prompts that demand an answer get an answer, real or not. Prompts that leave room for uncertainty get honesty more often:

Spot-the-Fabrication Quiz