Anthropic has told users to stop mistreating its chatbot. In an update to its terms of service, the San Francisco artificial intelligence firm announced that people can no longer direct sustained and needless abuse at its models, giving Claude the power to cut off conversations when interactions turn vicious.
The change arrived following the company’s regular policy review, first reported by tech publication The Verge. While Anthropic acknowledged it had already terminated sessions in rare instances, abusive behavior toward software now sits alongside severe violations like weapon development, automated election meddling, and the creation of non-consensual explicit images.
Where the Line Is Drawn
Anthropic insisted that regular users do not need to walk on eggshells. The ban does not police everyday frustration, nor does it penalize people who push back during an argument, stress-test the model’s safety boundaries, or explore dark creative fiction.
Instead, the rule targets extreme, deliberate hostility. The problem is that the company has not published an exact rubric for what crosses into genuine cruelty toward a text-generating tool. For many industry observers, that ambiguity is where the trouble begins.
The policy change quickly split tech communities online. A handful of safety researchers applauded the move as a nod toward model welfare, but skepticism dominated the broader reaction. Critics argue that treating a statistical calculation as an entity that can suffer devalues the concept of abuse entirely.
‘Cruel to Numbers and Maths’
Legal and technology experts were quick to challenge the premise of the update, warning against the creeping tendency to treat code like a living creature.
“You can’t be cruel to numbers and maths,” said Dr. Barry Scannell, a partner specializing in artificial intelligence at the Irish law firm William Fry. Writing on LinkedIn, Scannell argued that projecting human feelings onto software creates dangerous misconceptions about how these systems function. “This level of anthropomorphisation of AI is harmful. It leads people to believe that it’s something it’s not.”
Industry rivals appeared equally unimpressed. Mustafa Suleyman, head of consumer AI at Microsoft, reacted publicly with a melting face emoji. Suleyman had already criticized Anthropic in months past, cautioning the company against marketing conversational agents as pseudo-human companions.
Manners, Tokens, and ‘Clankers’
The controversy feeds directly into a long-running practical dilemma: does how we talk to software matter?
One camp argues that courtesy makes little engineering sense. Tacking “please” and “thank you” onto instructions wastes tokens—the basic fragments of text large language models process. Those extra tokens consume additional computing power and electricity for no functional gain. Yet other practitioners note that conversational niceties often nudge models toward better, more collaborative answers simply because they reflect polite exchanges in their training data.
OpenAI chief executive Sam Altman touched on the debate last year when asked about the power bills incurred by courteous prompts. Altman quipped that the expense amounted to “tens of millions of dollars well spent – you never know.”
At the same time, public sentiment toward automated software has developed a much harder edge. Annoyance with generative bots across social platforms like TikTok recently revived the word “clanker,” an old derogatory term for battle droids in the Star Wars franchise. The word saw such a sharp surge in everyday use that Collins Dictionary highlighted its comeback on its annual cultural shortlist.
A Question of Human Habit
Beyond questions of compute and computational logic, behavioral researchers suggest the policy might ultimately be less about protecting the machine than protecting the human behind the keyboard. Hurling unvarnished vitriol at a simulated conversational partner can easily turn into habit, subtly eroding how people interact with real people in workplace messaging apps or customer service queues.
Anthropic’s broader policy update also clamped down on real-world political dangers, explicitly prohibiting Claude from being used to run deceptive influence operations ahead of the US midterm elections. Yet none of those pragmatic guardrails attracted anywhere near the scrutiny sparked by telling humans to play nice with math.