Surely you have ever heard about the «hallucinations» of artificial intelligence. Even without being an enthusiast of this technology or tools such as Chatgpt. The truth is that errors in AI results are very common. According to experts, more and more. A new study, without going any further, argues that Chatgpt-5 confuses one in four times.
Actually, it is a proportion of very high failures. So the question that almost everyone is asked, starting with companies as OpenAi, is the following: Why does AI put the leg so much? The truth is that everything seems to indicate that there is a logical reason to explain the reason. Although of course, one thing is to understand why it happens and another much more complicated to remedy it.
Chatgpt is not very trustful, says a new study
Image generated with AI
Of all the artificial intelligences that exist at the moment, the truth is that Chatgpt is the most popular of all. Hence many studies focus especially on it. Sometimes, without leaving Openai and Sam Altman chatbot in what is said a good place. An investigation shared by the Middle Tom’s Guide has analyzed in this case its error rate.
It is what many know as hallucinations of AI, but that in reality are nothing other than mistakes. Chatgpt responds as if he were convinced that what he is sharing is very successful, when he really is not right. That is usual it was known. What surprises is that, according to a study, it tends to pass one in four times, nothing more and nothing less.
To find an explanation to this trend, experts say, it is necessary to address the way in which artificial intelligence is trained. At the moment, Evaluation systems penalize models to admit uncertainty. Or what is the same in the background: it drives them to generate confident responses even when they are not very sure of something.
In the same way, the study also emphasizes that models focused on reasoning, such as OpenAi’s O3 and O4-mini, tend to hallucinate more than previous models. Exactly for the same thing: since they are trained to affirm more and more, this also means that they fail more frequently.
Will there be an end to AI elucinations?
As a solution to this problem, researchers propose to change the way IA is evaluated. Without going any further, rewarding nuanced responses that indicate uncertainty or that refer to other sources, instead of promoting excessively confident responses. But of course, that perhaps would mean an inconvenience for the interests of artificial intelligence itself.
Because without his «reasoning», he might like seeking lifelong, without more. And therefore would lose, perhaps, much of its appeal for some. A complicated situation that suggests that, at least for the moment, the hallucinations of AI will continue to be there, in one way or another.
Know How we work in NoticiasVE.
Tags: Artificial intelligence