No, Better AI Models Will Not Stop Hallucinating

A persistent myth about large language models is that hallucinations are a passing problem. The reasoning sounds intuitive. Bigger models, better training data, more fine-tuning, and the made-up answers will eventually disappear. According to fresh research published by OpenAI itself, that reasoning is wrong.

In a paper summarised by Computerworld, OpenAI researchers argue that hallucinations are not an engineering flaw to be patched. They are a mathematical consequence of how language models work. Even a hypothetical model trained on perfect, complete data would still confabulate some of the time. The reason is statistical: language models generate the next most plausible token, and plausibility is not the same thing as truth.

Why models bluff

OpenAI's own writeup frames the issue differently from the way the public usually does. Models hallucinate because the training and evaluation procedures reward confident answers over honest uncertainty. A model that says "I don't know" scores worse on most benchmarks than a model that guesses confidently and happens to be correct half the time. The incentive structure, in other words, teaches models to bluff.

That framing matters because it points the finger at something fixable, even if the underlying mathematics is not. Benchmarks could be redesigned to penalise overconfidence. Training objectives could reward calibrated uncertainty. The will to do this, given that benchmark scores are how labs market new releases, is the harder problem.

The numbers most people miss

The scale of the problem is easy to underestimate. Even on basic summarisation, the most accurate frontier models hallucinate around 0.7 percent of the time. That sounds small until you check what happens in domains that matter most. Suprmind's May benchmark puts the rate at 18.7 percent on legal questions and 15.6 percent on medical queries. These are exactly the areas where users are most likely to take the model's answer at face value.

The pattern is depressingly consistent across the leaderboards. The more specialised the question, the more confident the wrong answer tends to be. A model asked about a real but obscure case in case law will often produce a plausible-sounding citation that does not exist. A model asked about a rare drug interaction will sometimes invent one.

What this means for users

What the OpenAI paper does not say is that hallucinations are unmanageable. It says they cannot be eliminated. Retrieval-augmented generation, structured calibration, and explicit uncertainty estimates can all reduce the rate. But the closed-form fix that some early commentators imagined, where one day a sufficiently advanced model simply stops making things up, is not coming. Anyone building a business around the assumption that hallucinations will quietly go away over the next few years is building on a faulty foundation.

The practical consequence for readers is simple. Treat every claim from a chatbot the way you would treat a tip from a confident stranger at a bar. Often correct, occasionally useful, and never to be acted on without a second source.