Try a simple example
Why can AI sound certain and still be wrong?
This answer sounds ready to use. Check three details against the actual note.
Check this answer against the note.
The confident-sounding answer
Definitely! The walk leaves at 09:00, meets at the north gate, and dogs are welcome.
Author-written answer and invented note. This is a controlled example, not a measured AI failure or a live fact-checker.
The worked example
The confident-sounding answer says 09:00, north gate and dogs welcome. The invented organiser’s note says “Meet at the north gate” on line 2 and “Departure: 09:30” on line 3. It says nothing about dogs. So the meeting point is supported here, the time is contradicted and the dogs claim is unconfirmed. A bounded answer uses 09:30 and the north gate, and leaves the dogs policy unknown.
“Definitely” did not make the details true.
The answer mixes a supported meeting point, a wrong departure time and an unstated dogs policy. Those are different evidence problems, even though every sentence sounds equally confident.
Check the claim that would change your decision
Find the source for an important time, number or promise. Separate what it supports from what is contradicted or still unknown. If the evidence is missing, ask rather than filling the gap.
Go deeper
What does “hallucination” mean here?
The term is often used for false or unsupported generated content. The departure time here is contradicted by the supplied note. The dogs policy is unsupported, which is not the same as proving that dogs are forbidden.
Why can fluent wording be misleading?
The TruthfulQA study documented models producing false answers that echoed common misconceptions. That supports checking content, not treating fluency as evidence. Our invented example does not measure any model’s present-day error rate.
Does a check mark prove the whole answer?
No. “Supported by this note” is a narrow comparison with the supplied text. The note itself could be wrong or outdated, and other claims need their own evidence. These annotations were written by people; the demo cannot fact-check your own text.
Explore the original sources
Primary references. These explain the mechanisms; the examples here are authored teaching simulations.
What would you like to try next?
Choose another small example, or return to this one whenever you need it.
See all the examples