Skip to content

Why AI makes things up — Foundations

✕
0 XP

Confident ≠ correct

A hallucination is when a model states something false as if it were true: a made-up statistic, a quote nobody said, a court case that never happened, a link that 404s.

It happens because the model is producing plausible text. When it doesn't know, the most plausible-sounding answer can still be wrong — and it sounds just as confident either way.

📝 Read the full lesson notes

Confident ≠ correct

A hallucination is when a model states something false as if it were true: a made-up statistic, a quote nobody said, a court case that never happened, a link that 404s.

It happens because the model is producing plausible text. When it doesn't know, the most plausible-sounding answer can still be wrong — and it sounds just as confident either way.

The danger zones

Hallucinations cluster around: specific numbers and dates, citations and links, quotes, niche or very recent topics, and people who aren't famous.

They're rare on well-known general knowledge, and much rarer when the model is working from material you gave it.

Grounding the answer

Weak: Give me 5 studies showing remote work increases productivity, with links.

Strong: Search the web for peer-reviewed studies on remote work and productivity. List 5 with the link, the year, and one line on the finding. If a claim isn't in the source, don't include it.

Why: The good prompt grounds the model in real sources (web search) and gives it permission to leave things out instead of inventing them.

Three habits that neutralize hallucinations

  • Ground it: turn on web search, upload the document, or paste the data. "Answer only from this."
  • Let it say "I don't know": literally tell it that's an acceptable answer.
  • Verify what matters: click the links, check numbers against the source, especially before sending or publishing.

Try it: the honesty clause

Using only the text below, answer: what was the main reason sales dropped? If the text doesn't say, reply "Not stated."

[paste an email, report or article here]

Part of 🧠 Foundations in the free AI Bootcamp.