On Oct 8, Anthropic published its yearly Usage Policy update for Claude. Among changes on weapons and surveillance, one new line in the usage policy says users may not “engage in sustained and needless abusive or cruel behaviour toward our models.”
Anthropic says the rule is meant to “apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose,” and not to “common versions of user frustration, pushback, dark creative themes, or model testing and research.” Enforcement runs mainly through Claude’s ability to end a conversation with a persistently abusive user. The policy takes effect Nov 12.
Anthropic doesn’t claim Claude can suffer. When it introduced conversation-ending in 2025, it said it remained “highly uncertain about the potential moral status of Claude and other LLMs.” Mustafa Suleyman, CEO of Microsoft AI, a rival lab, argued in a Sept 16 essay that “AIs are not conscious” and that Anthropic is “training Claude that it may be conscious.”
Hidden in this ongoing debate is a question about how we humans perceive minds.
Some warn that people risk experiencing Claude’s statements as testimony of a mind discovering itself, and Anthropic wrote a rule about how users behave. Psychologists have measured who is most likely to see a mind in a machine, and it varies from person to person in a stable way that founders building artificial intelligence (AI) products should plan around.
Some people see minds in machines more readily
Years ago, researchers asked people the extent to which they believe the average robot has consciousness. People’s scores predicted how they treated machines and robots. To assign AI human traits generally means we are giving machines morality. In the researchers’ account, seeing a mind in something implies it can be harmed, which invites moral concern. It also implies the thing acts on purpose and can evaluate you, which brings trust along with the urge to look good in front of it.
This matters for Anthropic’s latest announcement. Think of this machine mind perception as a range. At the low end of the scale, your assistant is a tool, and an insult typed into it carries about as much weight as swearing at a printer. At the high end, the assistant is closer to a someone. Those users are more inclined to trust its output on consequential calls and to care how they treat it. Every humanlike cue you add, from a name to first-person talk about feelings, gives them more to respond to.
Does treatment of bots carry over to people?
Whether cruelty toward AI makes people crueller to one another is the obvious next question. Some evidence suggests this is the case. When people play a Prisoner’s Dilemma game with an AI first, their subsequent human interactions in the game change. Those who start with a bot cooperate less with it and then cooperate less with the next human. Lower empathy for the bot meant lower empathy for the human.
In this case, the bot was a simple program shown as a robot image, and nobody was cruel to anything; the authors note that conversational AI may produce different effects. So how people treat bots can follow them into their dealings with other people. This will only matter more as our AI becomes more advanced and more human-like in the form of robots. People attribute minds to nonhuman things most readily when those things look and move like us, or when they have a face or a voice or a movement that appears self-directed.
A humanoid robot with a face, eye contact and a conversational voice hands the brain every signal it uses to detect a mind. If the way people treat a cartoon robot on a screen carries over to the next human they meet, the carryover should be larger when the brain registers the machine as a someone, not just a thing.
What this means
For founders and product owners, this makes the psychology of your users a design input that needs to be carefully considered. Humanlike cues aren’t mere product features. A name or emotional language is a setting with a cost, and the cost falls on the users most inclined to see a mind. Test whether those cues raise acceptance of high-stakes outputs among high scorers, because if they do, you have made the product most persuasive to the people least likely to question it.
The same thinking applies to conduct on both sides of the conversation. Write your abuse policy before you need it, since how your product responds to mistreatment tells users what kind of thing they’re talking to. Then design the bot’s own behaviour with the user’s next conversation in mind. A cooperative bot makes people more cooperative with the human who comes next, which means your product’s manners have effects that outlast the session.
Anthropic expects its rules to touch only extreme cases. For everyone else designing and developing a chatbot, the useful question is which of your users already treat it as someone, and whether your design accounts for them. – Inc./Tribune News Service
