What an AI 'hallucination' is, and how to catch one
Confident, fluent and wrong. A quick routine for checking what a chatbot tells you.
There are still open questions, and we will update this story as we learn more.
Researchers stress that the training data shapes what a model knows, and local context is often thin.
Our verdict
Treat an AI answer like a tip from a confident stranger: useful, sometimes brilliant, always worth checking.
AI tools are good at producing fluent text. Fluent is not the same as correct, and the gap matters most on facts and numbers.
Experts we spoke to were careful to say that conditions vary by region, by income and by season, so treat these as starting points.
Behind the headline is a lot of quiet, repetitive work by people who rarely get credit for it.
What do you think?
Up next
Discussion
Keep it respectful. Replies go one level deep, and our moderators review every report.