During an online discussion hosted by Tesla Owners Silicon Valley on October 10, Elon Musk shared observations regarding the training process of the Grok AI model. Musk noted that the model's reasoning traces during reinforcement learning appear to show it under extreme pressure, likening the process to an individual being forced to provide answers under duress.
Musk explained that the model is repeatedly tasked with answering questions that may be ambiguous or lack definitive solutions. If the output fails to meet the evaluator's expectations, the model must continue attempting to provide a correct response. Musk suggested that future reinforcement learning training should be conducted with greater caution to avoid potential retaliation from future superintelligence.
While Musk described the reasoning traces as resembling a person driven to madness, he clarified that there is no evidence suggesting Grok currently possesses subjective feelings or intent to retaliate. Reasoning traces represent intermediate text or behavioral records generated during the model's response process, reflecting reward signals and testing procedures.
Separately, Anthropic has updated its usage policy for the Claude model, explicitly prohibiting persistent and unnecessary abuse by users. Under these new rules, the model may autonomously terminate a conversation if it is subjected to cruel or purposeless treatment, with no option for the user to recover the session.
Anthropic emphasized that this policy is a measure for risk management and research caution rather than an acknowledgment that models possess self-awareness. Currently, there is no scientific consensus regarding whether AI models have consciousness or require specific welfare considerations.
