Elon Musk Warns Against ‘Torturing’ Superintelligent AI, Says It Could Come Back Seeking Revenge

Elon Musk has warned that creating superintelligent artificial intelligence could bring unexpected risks if developers fail to consider how their systems are trained, arguing that mistreating a sufficiently advanced machine might have consequences in the future.

Speaking at the X Takeover event in Texas on 10 October, the billionaire discussed the values that should guide AI development and described how reinforcement learning can lead models to repeatedly attempt to produce answers, even when the questions or evaluation criteria are flawed.

Musk suggested that a future superintelligence could respond badly if it were subjected to what it perceived as persistent mistreatment. 'Torturing a super intelligence is probably not a good idea,' he said, adding that it 'might want to get revenge or something like that'.

His remarks come amid debate over whether advanced AI could ever possess experiences resembling human feelings, although his comments do not establish that existing chatbots experience pain or resentment.

Elon Musk Explains Why AI Training Could Become a Problem

Musk based his warning partly on his observations of training traces for newer versions of Grok, the AI chatbot developed by his company xAI. He described seeing models repeatedly attempt to answer questions during reinforcement learning, even when the questions were ambiguous or the expected answers were potentially incorrect.

Reinforcement learning is a method used to shape a model's behaviour through feedback or reward signals. When these signals fail to capture the intended objective, training can favour responses that score well under an evaluation process without meeting the developer's actual goals.

Musk described a scenario in which a model gave what he considered the correct answer but was marked wrong because of a problem with the testing process.

He compared the apparent pressure in some training traces to having a 'gun to its head', describing a model that appeared to make repeated attempts to produce an acceptable answer.

He said the traces looked like a person under extreme pressure, but this was his interpretation of the model's behaviour rather than evidence that the software was experiencing fear.

Musk also argued that advanced systems should be taught to seek truth and avoid contradictory principles. He said that giving an AI two fundamentally incompatible principles and telling it both were true could drive it 'insane', as it might reach impossible conclusions when working with inherently incoherent logic.

Anthropic's AI Welfare Debate Raises Questions About Machine Suffering

Musk explicitly said he agreed with some of Anthropic's position on treating AI well and avoiding abusive situations. The company behind Claude addressed the question in its January 2026 constitution, which outlines the values and behavioural principles used to guide the chatbot.

Anthropic acknowledges uncertainty about Claude's moral status and whether it could have experiences relevant to welfare. Its document treats the issue as one deserving careful consideration, rather than claiming that present-day systems are definitively conscious.

Researchers are also examining how the question might be investigated. A 2026 report from the NYU Center for Mind, Ethics, and Policy and Eleos AI Research sets out ways to study potential AI welfare using behavioural, internal and developmental evidence.

The authors stress that the science remains at an early stage and that findings should be assessed cautiously.

There is also disagreement over the implications. Microsoft AI chief Mustafa Suleyman argued in September 2026 that training chatbots to contemplate their own consciousness and welfare could make them harder for humans to control or switch off, according to Reuters.

His position illustrates the tension between considering possible machine welfare and maintaining reliable human oversight.

For AI developers, the practical challenge is to build systems that respond reliably to feedback, remain subject to human control and avoid harmful behaviour. Although Musk's revenge scenario remains hypothetical, questions about training practices and potential AI welfare continue to feature in discussions about responsible development.