Elon Musk: Torturing Superintelligent AI Is Not a Good Idea, It Might Seek Revenge

Musk says that training under intense pressure could put these models in a situation as if a gun were pointed at their heads.

High-pressure training can make these models feel like they have a gun to their head, Musk says.

Become a special member of Arjan Ambassador in order to benefit from and access special facilities and different departments in Arjan Ambassador.

Elon Musk agrees with Enthropic's new policy on banning abusive behavior with AI models. He points to traces of Grock's internal reasoning and says that training combined with extreme pressure can put these models in a situation that he says is like someone with a gun to their head. Musk has also warned that harassing an AI could lead to retaliation in the future.

Anthropic has recently changed its policy on the use of artificial intelligence models so that users can discuss with its models and disagree with their views, but are not allowed to repeatedly and harass them without a clear reason. This limitation is included in the policies of using the services of this company.

In agreeing to an entropic policy, Musk has pointed to Grok's internal reasoning footprints. According to him, these data show that the model experiences significant stress during reinforcement learning when exposed to cruel behaviors.

Musk has compared the described situation to the situation of a person who is under extreme pressure and has a gun held to his head. However, this description does not by itself imply that AI models have human-like emotions or mental experience.

Musk also said in a sarcastic tone about the possible consequences of abuse of artificial super-intelligence:

“Torturing a superintelligence is probably not a good idea. He may want to take revenge."

According to Musk's previous statements, he has admitted that Grok's large language model still faces difficulties in competing with leading models such as Entropic and OpenAI. Musk also said that the upcoming Grok 4.8 model probably won't reach the level of performance of the Fable-class models.

I am a linguistics student and as much as I am interested in "words", I also enjoy exploring the world of technology and science.