Will AI Ever Stop Making Things Up?

Ask a chatbot a question and it will almost always answer, fluently and with apparent confidence, even when the answer is wrong. These confident falsehoods have a name, hallucinations, and a popular assumption rides along with the word: that they are a temporary glitch, the kind of thing the next model will iron out. It is a comforting idea. The evidence suggests it is not quite true.

The most useful account comes from OpenAI's own researchers, in a 2025 paper plainly titled Why Language Models Hallucinate. Their argument is not that models are broken. It is that the way they are built and graded makes a certain amount of making things up almost inevitable.

Two places the problem comes from

Start with how a model learns. During pre-training it absorbs enormous amounts of text to predict what word comes next, but nothing in that text is labelled true or false. The model picks up the shape and rhythm of language brilliantly. What it does not pick up is a reliable sense of which specific facts are real. For arbitrary details that follow no pattern, such as a particular person's birthday, there is nothing to generalise from, so the model is prone to producing a plausible guess that happens to be wrong.

The second cause is subtler, and it is about incentives. After pre-training, models are refined and tested against benchmarks. Those tests tend to reward a confident answer and give no credit for saying "I don't know." A model that guesses will sometimes be right and score points. A model that admits uncertainty scores nothing. Over many rounds of this, the training quietly teaches the system that guessing beats honesty. The hallucination is not a malfunction. It is the behaviour we accidentally asked for.

Better, but not cured

This does not mean nothing improves. OpenAI reports that its GPT-5 generation hallucinates significantly less, especially when it is allowed to reason through a problem before answering. Models that can check themselves, cite sources or defer to a search tool are more trustworthy than their predecessors. The direction of travel is real.

But the company is blunt that the problem is not solved, calling hallucinations a fundamental challenge for all large language models. That framing matters. If the errors came from a shortage of data or computing power, more of both would eventually fix them. If they come from the basic design, as the paper suggests, then they can be reduced and managed but probably not eliminated.

So the honest answer to whether AI will stop making things up is: less often, yes, and completely, probably not. The practical lesson is the same one that applies to a high benchmark score: a confident tone is not evidence of a correct answer. Treat these systems as fast, fluent and occasionally sure of things that are not so, and check anything that matters. That is not a knock on the technology. It is just how to use it well.