Why does AI make things up?

Say you ask a chatbot for five good books about gardening in the shade. It gives you a tidy list, with authors and titles. Four are real. The fifth has a real author's name on it and a title that author never wrote. Nothing marks it out. It sits in the list looking exactly like the others. This happens often enough to have a name, and once you see why, it gets much easier to handle.

Most people call it a hallucination: an artificial intelligence (AI) tool stating something false as if it were a fact. The National Institute of Standards and Technology (NIST), a United States government agency, prefers the word confabulation. It describes confidently stated but false content that can mislead the person reading it. Same thing, two names. The answer sounds sure, and it's wrong.³¹

A chatbot writes by predicting. It looks at everything so far, picks a likely next word, then the next, over and over. Each word fits the ones before it. But nothing in that process stops to ask whether the whole sentence is true. Most of the time the likeliest answer is also the right one, because the right answer turned up often in what the model learned from. When the facts are thin, it still produces something shaped like an answer. So a book title gets built the way a real one would be: a plausible author and a plausible-sounding title.³

In 2025, researchers at OpenAI, the company behind ChatGPT, compared it to a multiple-choice exam. If an empty answer earns nothing and a guess might earn a point, guessing is the smart move. Many of the tests used to train and grade AI models are scored the same way. A right answer counts, "I don't know" earns zero, and so models learn to guess. The researchers suggest scoring a confident wrong answer below an honest "I'm not sure."²

Made-up answers cluster in a few places, so you know where to look.

• Rare facts, like a small company's founding date or a local rule.
• Names, numbers, and dates.
• Quotes, citations, and links to sources.

Those are details that appear only a handful of times in what a model learned from, so it has little pattern to go on. Common facts are safer. Ask about something people have written about endlessly, and it's usually right.²⁴

It's tempting to say the AI lied. But lying means knowing the truth and choosing to say something else, and there's no intent like that here. The chatbot produced the likeliest-looking answer, and a likely answer can still be wrong. It's a guess, not a lie. That difference changes what you do about it. You don't need to catch it out. You need to check it.

Many chatbots can now search the web and show where an answer came from. That helps. Even so, OpenAI's own researchers say hallucinations still happen in the newest systems. So keep the check for the claims that matter: the date you'll plan around, the quote you'll repeat, the book you'll go looking for. Next time a list comes back, pick the one item you'd act on, and look it up somewhere that would know.²⁴