Openai researchers in a new research article deals with one of the biggest issues of artificial ielligence; According to them, artificial ielligence models are hallucinated because standard education and evaluation methods encourage them to guess instead of confessing to the uncertaiy of the answers.
At first we need to know what illusion is; Hallucation occurs when the language model confidely expresses incorrect information as a reality. This problem can be seen even in the most advanced models, such as the Gpt-5 or Claude. Researchers at Openai have now described the cause.
Researchers say that in the curre evaluation system of artificial ielligence models, guessing a response is better even if it is wrong to confess to it. Because a luck guessing may get pois, but saying “I don’t know” has no pois. This problem plays all the leading models, from its Gpt-5 to the Claude ahropic, making users completely unable to trust the chats.
The cause of the illusion of artificial ielligence
The researchers liken the phenomenon of hallucinations to a multiple -choice test. If you don’t know the answer to the question, you may score with a chance, but with the blank the answer sheet will surely be zero. Similarly, when artificial ielligence models are only evaluated based on “accuracy” (ie the perceage of fully correct answers), they are encouraged to guess instead of “I don’t know”.
“Humans learn the value of expressing uncertaiy outside school and in the real world,” the researchers write in one article. But linguistic models are mainly evaluated using tests that penalize uncertaiy. “
To show this more precisely, Openai has compared its last two models:
| Criterion | gpt-5-threading-mini (newer model) | Openai O4-MINI (Older Model) |
| Reference rate (not responding) | 2 % | 2 % |
| Accuracy rate (correct answer) | 2 % | 2 % |
| Error rate (illusion) | 2 % | 2 % |
This table clearly shows that the older O4-MINI model, although lower, is much higher because it is almost always guessing. In corast, the newer model, although refusing to respond more, is much less illusory.
According to the researchers, there is a simple solution to this problem: redesigning evaluation criteria. They suggest that scoring systems should change in such a way as to fine high confidence errors than expressing uncertaiy.
But what is the source of illusions? Language models in the “Pre -Education” phase learn the next word in a huge volume of iernet texts. In this data, fixed patterns such as spelling or grammar are easily learned. But specific and low -key facts (such as the date of birth of a particular person) do not follow any predictable pattern. As a result, in the face of such questions, the model “conjectures” the most likely combination of words based on its data, instead of accessing a recorded truth, and here is the illusion.
