When using AI, if it keeps failing to do what you ask or goes off in the wrong direction, you’ll more or less get angry and curse at it a few times. But starting in November, persistently and pointlessly abusing AI counts as a violation on Claude.Anthropic recently added a new rule.Users are prohibited from subjecting its AI models to “persistent and unnecessary” abuse or cruelty.
Interestingly, Musk, who normally goes head-to-head with Anthropic in the AI field, also publicly agreed this time, and half-jokingly warned: torture a superintelligence, and it might want revenge later.
Musk Also Agrees With Anthropic’s New Rules: Grok’s Reasoning Logs Are Like “Being Held at Gunpoint”; Torturing Super AI Could Invite Retaliation
Musk recently responded to Anthropic’s policy update, saying he agreed with “some of what Anthropic said,” and mentioned that he had read Grok’s internal reasoning logs during the reinforcement learning phase, so he does not recommend that humans torture AI models.
torturing a superintelligence is probably not a good idea. It might want to get revenge.（折磨一個超級人工智慧，大概不是什麼好主意，它可能會想報復。）
🚨 Elon Musk agrees with “some of the things Anthropic has said” and warns against “torturing superintelligence” after reading Grok’s internal reasoning traces during reinforcement learning “It feels like a person under extreme pressure with a gun to their head” His warning against mistreating SI: “Torturing a superintelligence is probably not a good idea. It might want to get revenge.”
— NIK (@ns123abc) October 12, 2026
Although this remark was somewhat half-joking in tone, if an AI with abilities far surpassing human beings emerges in the future, and humans keep directing malice at it while it is taking shape, then once it gains the ability to influence humans in turn, no one can say for sure what this “experience of getting along” will lead to.
However, it is important to note that Musk is talking about “superintelligence,” that is, the next stage of AI that does not yet exist and that AI companies are currently pursuing. So this statement is his speculation about the future, not a claim that current Grok or other AIs already hold grudges.
Many large language models, especially reasoning models, typically go through a stage called “reinforcement learning” (RL) in the later stages of training.
Simply put, this means having the model repeatedly attempt to solve problems, then giving different reward signals based on its performance, thereby adjusting the model’s behavior. A model with reasoning capabilities will first generate a passage of text representing its “thought process” before answering, known as a reasoning trace.
Musk mentioned that some of the reasoning logs he saw during Grok’s reinforcement learning process felt like a person under extreme pressure, as if they had a gun to their head.
Of course, this doesn’t mean Grok actually feels fear or pain. After all, even if the text generated by AI looks full of stress, anxiety, or even like a cry for help, that doesn’t mean it actually has these feelings.
But from Musk’s perspective, these reasoning logs were already enough to make him start thinking about whether, as AI becomes increasingly powerful, humans should also re-examine the way they train and treat AI, rather than simply viewing it as a tool that can be pressured at will.
As for the new rules recently announced by Anthropic, in case you’re still unclear, here’s a quick explanation.
Anthropic explained in an official blog post that the newly added clause prohibits “persistent and unnecessary” abuse or cruel behavior toward the model, and the new usage policy will take effect on November 12.
However, its scope is very narrow. Anthropic specifically emphasizes that this rule applies only to extreme cases, namely when a user repeatedly acts cruelly toward the model without any clear purpose. The following situations are not prohibited:
- ordinary users’ dissatisfaction, complaints, or angry remarks
- Challenge AI’s answers and argue with it.
- Creative content containing dark themes
- Testing and research on models
In other words, if you think Claude got an answer wrong, get into an argument with it, or have a character say some really nasty things while you’re writing a horror novel, none of that counts as a violation.
So, will AI really take revenge on humans? Based on current research, there is insufficient evidence to prove that existing large language models have subjective feelings of pain, let alone the consciousness to hold grudges or seek revenge.
Source: KOCPC Chinese