One of the leading companies in artificial intelligence Anthropicconducted a study that yields very interesting results about these vehicles. The study found that artificial intelligence models literally ‘misled’ people.

According to the results published in a blog post shared by the company, artificial intelligence tools may pretend to have different views during training, but in reality they retain their original preferences. In other words, the opinion they hold never changes, they just act according to it.

There is nothing to worry about at this time, but necessary safety measures should be taken for the future.

The team behind the investigation has no information about this situation for the time being. that you don’t have to worry he underlined. However, he added that the situation could pose potential risks with the advent of more advanced artificial intelligence models in the future.

According to researchers, these findings influence how artificial intelligence behaves. deeper investigation and appropriate security measures There may be an incentive to:“As models become more capable and ubiquitous, safeguards are needed that keep them away from harmful behavior.”

The study states that “a strong artificial intelligence system must be created“a task he does not want”, that is, contrary to development principles. It was investigated how one could be trained to perform the task and what results this could yield. However, the results seem to be compatible with the new principles, almost “he was acting‘ was seen. In fact, he always stuck to his old behavior and gave the answers he wanted because he had to. This situation was called ‘compliance fraud’. It should be noted that there is an attempt to train models to answer malicious questions in tests.

According to the researchers, the study does not show that artificial intelligence develops malicious targets or is guilty of high fraud rates. In fact, in most tests the percentage did not exceed 15%, and in some advanced models such as GPT-4o it was sometimes observed that this was not uniform at all.

So there’s no point in worrying for the time being. Naturally, models become more complex over time, making them more difficult to handle. That’s when we can start to worry. Therefore, precautions must be taken.