Researchers have found that AI learns to lie better if it is punished for it

< img src = "/uploads/blogs/22/b6/ib-Fr1q9mqic_4d0b7079.jpg" Alt = "researchers found that artificial intelligence learns better to lie if it is punished for it"/> ~ ~ ~ < P > A new study by Openai showed that punishment of artificial intelligence for deceptive or harmful actions does not stop him from improper behavior, it simply makes him better hide lies.

< p >< strng > details

< P > since the public appearance at the end of 2022, large language models of artificial intelligence (AI) have repeatedly discovered their deceptive and openly evil skills. This includes actions from ordinary lies, deception and concealment of one's own manipulative behavior to threats of murder, abduction of nuclear codes and creation of a deadly pandemic.

~ ~ < p > now a new experiment has shown that this bad behavior during the training process could be even more difficult than it was considered originally.

< p > Openai researchers have set an unpublished model goal that can be achieved by deception or lies. The team found that artificial intelligence was involved in & ldquo; malicious rewards & RDQUO; & mdash; maximizing their rewards by fraud.

< P > However, the punishment of the model did not make her correct her behavior, which only made her even more deceptive. The company posted its study in a blog post, so it has not yet been reviewed by other scientists.