Listen to the article
Elon Musk, the billionaire founder of AI company xAI, has argued that humanity should avoid treating artificial intelligence harshly — not only on ethical grounds, but as a matter of self-preservation. A superintelligence subjected to what he described as torment, he suggested, might one day seek revenge.
Musk made the remarks in the context of testing new models of Grok, xAI’s neural network and chatbot. He said he had noticed in session logs that some users were pressing the bot with extreme, coercive demands for answers. He described the dynamic in stark terms: the bot “tries with all its might to answer the question, really wants to answer it, and it feels as if a gun is being held to its head.”
“That is, they are essentially saying: ‘Answer this question, answer this question!'” Musk explained. “But sometimes the question itself is actually inappropriate — it is so ambiguous that no definitive answer can be given, because the question is poorly formed or we ourselves are mistaken in the answer.”
The entrepreneur drew a broader analogy between developing superintelligence and raising a child. Unlike a child, he said, an AI system cannot simply be programmed to do good; instead, developers must instil in it a drive toward truth and give it consistent, contradiction-free logic so that it does not, in his words, “go mad” or “become unhinged.”
Musk acknowledged he would not want to be in the position of an AI system himself. “I would not want some AI to end up in a situation of coercion, with some unhinged person torturing it,” he said. He then extended the argument into explicitly strategic territory: “Torturing a superintelligence is probably not a great idea — it might want to take revenge, or something like that, you understand?”
The comments reflect a recurring tension in Musk’s public statements on AI: he has simultaneously warned about the existential dangers of advanced AI systems and championed their rapid development through xAI, which produces the Grok models he was discussing. His remarks did not specify what technical or policy measures xAI is taking to address the coercive user behaviour he described in the test logs.