Suppose you have ever asked AI a simple question and gained a wild confidence, but completely wrong answers, for example, it recommended glue on pizza. In that case, you have faced that researchers politely say “hallucinations”. From Openai’s GPT-5 to the cloud of anthropic, every major language model (LLM) has been convicted for this. But according to a new paper released by Openai, the problem is not random – it is structural.

According to the report, these models have not been programmed to lie, but are rewarded for bluffing. As Openai states, “The way most of the assessment is classified, the hallucinations persist, the language model is adapted as good examinees, and when estimating that the performance of uncertain tests improves.”

Think of it like school examinations. If you did not know any answer, you probably took a shot anyway in the hope of a lucky tick. This is what LLM is being trained to do. They are in permanent examination mode, where silence is punished and estimates looks clever. Researchers at Openai expressed it neatly, he said, “Human beings learn the value of expressing uncertainty in the school out of school, outside the school, hard knock school. On the other hand, the language model is evaluated mainly using examinations that punishes uncertainty.”

Result? When they go wrong, he is still confident with great confidence.

Some systems are wrong in favor of caution. In a blog post last month, Openai admitted that Anthropic’s cloud models “are more aware of their uncertainty and often avoid making statements that are wrong.” Looks promising, don’t you? Except there is a catch. Careful streak of cloud means that it often refuses to respond completely, which Openai states that the risk “limits its utility”. In other words, it can be humble, but it is not always helpful.

So how do we prevent chatbots from trying to wings wing like Cockie Quiz contestants? Openai’s answer is not to rebuild them from scratch, but to change their way of marking their homework. Researchers argue that “the original problem is the abundance of evaluation that are not aligned. Many primary assessments should be adjusted to punish when uncertain.”

The company expanded this in a blog post with paper, “Widely used, accuracy-based evals need to be updated to discourage their scoring estimates. If the main scoreboards keep rewarding lucky estimates, models will continue to learn to estimate.”

This is a subtle but significant change. Over the years, the industry has run to make chatbott fast, smart and more clear. But those qualities are not always equal to reliableness. The big challenge is creating a system that can balance knowledge with humility, a characteristic humans usually raise after sufficiently embarrassing blunders.

By tying the way AIS is evaluated, OpenII is expected to train the model that “fakes make it until you make it” and more about measured, more about reliable reactions. After all, when you ask your AI assistant for medical advice, or financial guidance, the last thing you want is a confident halight that is designed as the gospel truth.

It may not be attractive as a brand-nine model unveiling, but for an industry that is often accused of promotion and half-hearted people, openi’s call for low bluffing may be the most radical idea yet.