TwiddleLab / Artificial Intelligence / AI and Us: Using AI Well
Hallucinations and Fact-Checking
What you will learn
- A hallucination is a confident answer that isn't true. A chatbot picks likely next words; it doesn't look facts up unless a tool does that for it.
- Made-up sources look real: a name, a university, a year. Search for them before you trust them.
- Fact-checking: split the answer into claims, search each claim on its own, and find a primary source.
- Rare errors add up: 1 wrong claim in 20 means about 3 wrong in an essay of 60.
How the lesson goes
You ask a chatbot for help with a history report. It gives you 4 claims. How many are true?
They all sound sure of themselves. Make a bet with the marker, then Check.
Check each claim against what you found.
Give every claim a verdict, using the search results on the right.
A study with a name, a university and a year sounds solid. But when a search finds no paper, no author page and no news at all, the safest verdict is: no source found, probably made up.
How does a chatbot invent a study? Watch it write one.
Press Play, or step one word at a time. The bars show the likeliest next words.
What you just saw has a name.
A hallucination: a fluent, confident answer that isn't true. The model predicts likely words, it doesn't check them.
Fact-checking: split an answer into claims and test each one on its own.
A primary source is where a fact comes from: the study itself, the official record, the museum's own page. Go there, not to another summary.
Not everything needs checking. Which pieces of a chatbot answer do?
Drag each piece into a lane.
Use chatbots freely for ideas and wording. Anything that claims to be a fact, a number, a quote or a source gets checked.
Find the made-up claim. You have 3 searches.
Search a claim to see what comes up. Flag the invented one, then Check. Fewer searches, more stars.
Search the riskiest claim first: the one with a specific name, year or number. General facts that every source repeats are rarely the invented ones.
Build a fact-check routine you can use on any answer.
Put the 4 stations in order.
Searching the whole answer just finds more copies of it. Searching one claim at a time, and going to the original source, is what catches errors.
Math: a chatbot gets about 1 claim in 20 wrong. Your essay uses 60 of its claims.
Drag the marker to the number of wrong claims you'd expect, then Check.
"Almost always right" is not enough for an essay: 60 x 1/20 = 3 errors you can't see. That's why you check the claims you hand in.
Last one. Tap every line you must check before this goes into your report.
Then Check to stamp it.
The date is right and the quote is famous, but you still check them: you'll put your name on them. The "report" is the classic trap: scientists estimate the Moon shrank about 50 metres over hundreds of millions of years, not every year.
Previous lesson: Bias: Whose Data? · Next lesson: Deepfakes and Media Literacy · Part of the course: AI and Us: Using AI Well
Lessons are interactive and need JavaScript. Turn it on to play them.
