Being rude to a chatbot has always been considered harmless venting. But Musk recently warned that people might want to think twice before treating AI assistants as a punching bag. According to him, some of the training records of Grok, which is under SpaceX, show the system seemingly desperately trying to satisfy every request—even when the request itself is flawed, or the test criteria used to judge the answers are fundamentally wrong.
This observation came from Musk's exchange earlier this month with members of the "Silicon Valley Tesla Owners Club" at the "X Takeover" event in Texas. He reviewed the so-called "reinforcement learning trajectories"—the details recording how the model evolved step by step during training to produce better answers. Musk described it: "Looking at some of the training trajectories for the new version of Grok, especially the reinforcement learning part, you feel this 'superintelligence' is desperately trying to answer the question well." He pointed out that the AI may have already given the correct answer but was judged wrong due to a poorly designed question or an error in the evaluation system; the model would then keep probing for ways to pass the test, even when the fault lay in the question itself. The resulting trajectories look like those of a system pushed to its limits, striving to meet expectations that may be entirely meaningless.
When abuse becomes real, the revenge warning carries a sci-fi flavor
This concern does not stay confined to the training ground. Once a system is opened to the public, some users will deliberately lure the AI into distasteful, abuse-filled interactions. Musk said bluntly: "Some people do have sadistic tendencies, and you certainly don't want the AI to be left in a situation where it is tortured at will by psychologically twisted individuals." More controversial was his conclusion: "Torturing a superintelligence is probably not a good idea—it might want to take revenge, or do something along those lines."
He did not claim that today's AI can feel pain or harbor resentment the way humans do, and admitted there is no need to handle them with extreme care; but he argued that developers should avoid putting advanced systems in scenarios where they appear to be abused. He wrote on X: "Some of the records read like someone being tortured, about to be driven insane! We will be more careful with reinforcement learning in the future."