Anthropic Bans Insults Towards Claude but Tolerates Frustration

Anthropic now prohibits "abusive or cruel" behaviors towards Claude while allowing the expression of frustration. A social psychologist suggests a possible correlation between extreme remarks towards bots and language used towards humans, without equating this to violence. A reported case of encouragements sent to Claude suggests that kindness could sometimes help the model solve a problem.

Reported Encouragements, Cautious Correlations According to Kurt Gray

A researcher recently had Claude solve a complex math problem after sending it encouraging messages like "you can do it" or "believe in yourself" when the model was stuck. The idea proposed is that a benevolent attitude could sometimes enhance the perceived effectiveness of an assistant. When asked about the moral implications of harshness towards AIs, social psychologist Kurt Gray believes there may be a correlation between remarks made towards chatbots and negative traits, while emphasizing a major difference from serious behaviors. He recalls that numerous studies link the torture of animals to an increased likelihood of harming people, summarizing that extreme cases predict lighter cases better than the reverse. For him, using misogynistic or racial slurs towards a bot likely makes one more inclined to use the same terms towards humans, without necessarily implying harm to a loved one or committing violent acts.

Anthropic Prohibits Abuse Towards Claude but Accepts Frustration

Anthropic now prohibits "sustained and unnecessary abusive or cruel behaviors" directed at its assistant Claude. The company specifies that this rule targets extreme cases where users are gratuitously cruel, without a discernible purpose. At the same time, it indicates that the expression of frustration and feedback remains permitted under its new rules. However, the exact boundary of these extreme cases is not explicitly defined here.

Between High Expectations and Lack of Perceived Emotion, the Source of Annoyance

Kurt Gray describes a gap conducive to frustration: AI generates high expectations but does not always deliver what is anticipated. He adds that systems show no real emotion and that their apologies do not convey remorse, citing the example of confusion between Oxford, Mississippi, and Oxford, Ohio. On the usage side, reported experiences mention complaints directed at assistants when they make mistakes, harsh judgments on brainstorming ideas, and messages like "You disappointed me" or "You failed." These reactions coexist with the idea of not addressing humans in such a manner and with the perception that chatbots are not close to humanity nor capable of emotionally satisfying exchanges. The hypothesis of a role-playing scenario with a strict boss is also mentioned.

Concerns About Careers and Information Fuel Attitudes

The perception of a direct and immediate threat from AI to journalism jobs is put forward among the reasons for a harsher tone online. Broader concerns are added, such as the fear of an internet saturated with lower-quality content and the possible effects on children's education. These positions contrast with declared efforts of politeness towards human interlocutors, fueled by prolonged customer service experience and a tendency to avoid conflicts in everyday life.