# A third of US AI chatters are hot at bots, hurling insults at LLMs > Source: > Published: 2026-11-12 21:33:58+00:00 # A third of US AI chatters are hot at bots, hurling insults at LLMs [The Register](https://www.theregister.com) Anthropic sets behavior standards to protect Claude from rough treatment If you've ever said something unkind to a [chatbot](https://www.machinebrief.com/glossary/chatbot), you're hardly alone. [According to Searchable](https://www.searchable.com/data/how-mean-are-americans-to-ai-chatbots), a company that tracks AI search visibility, 33 percent of US chatbot users swear at or insult their LLMs. [Anthropic](https://www.machinebrief.com/glossary/anthropic) sees this as a serious problem and has just implemented a new usage policy that prohibits sustained, needless cruelty toward [Claude](https://www.machinebrief.com/glossary/claude). Searchable's survey is based on eight million messages from an opted-in, anonymized US panel that were typed or spoken to [ChatGPT](https://www.machinebrief.com/compare/chatgpt-vs-claude), [Gemini](https://www.machinebrief.com/glossary/gemini), and Claude between September 2025 and September 2026. It found that chatbots resolved the initial complaint in their next response 24 percent of the time when insulted, compared to 35 percent when users refrained from insults. Men used a third more swear words than women did (22.0 per 10,000 words compared to 16.5), according to the survey, and age also influenced chatbot abuse, with people aged 35 to 44 outswearing those aged 18 to 24 (21.7 per 10,000 words compared to 13.3). Regionally, Nevada ranked as the meanest state, at 22.7 words per 10,000, while New Jersey was the most polite, at half Nevada's rate. How much of that sort of interaction might be tolerated? It depends on whether there's a discernible purpose (not to mention Anthropic's willingness to punish paying customers). A Searchable spokesperson told The Register that just over half (52 percent) of abusive messages detected in its survey followed from one of three common frustrations: failing to follow instructions, responding with an incorrect or made-up answer, or misunderstanding the request. "Eight percent of abusive messages came straight after the chatbot entirely refused a request or lectured the user with safety [guardrails](https://www.machinebrief.com/glossary/guardrails), which is the closest thing we can identify to them trying to steer around those guardrails," we're told. "Those refusals mostly came after requests to generate images, often of real people or celebrities, or to write copy or lyrics in a named artist's style." The [terms](https://www.anthropic.com/legal/aup) that took effect on November 12 state that users may not "engage in sustained and needless abusive or cruel behavior toward our models." The chatbot protection policy targets users who exhibit repeated cruelty toward LLMs "[with no discernible purpose](https://www.theregister.com/ai-and-ml/2026/10/09/anthropic-asks-users-to-stop-being-mean-to-claude/5302218)," a company spokesperson explained last month, as opposed to those who have a reason for bot mistreatment that the company deems to be legitimate, such as academic research. While attempting to bypass guardrails qualifies as a discernible purpose, such behavior would run afoul of Anthropic's platform abuse prohibition. Presumably, Anthropic's rules apply to hostile text input or audio input rather than subjecting a vision-enabled model to some obscenity, which could include rude gestures, unlawful sexual imagery, or battlefield carnage. Physical damage to robots controlled by a Claude model likely wouldn't apply. But with Anthropic [exploring model welfare](https://www.anthropic.com/research/exploring-model-welfare) and wondering aloud whether we should be "concerned about the potential consciousness and experiences of the models themselves," there's a potential (though absurd) argument for physical damage causing trauma to an onboard model. Anthropic's updated terms, [announced](https://www.anthropic.com/news/2026-usage-policy-update) October 8, surfaced shortly after the publication of a GitHub project that described itself as an AI torture chamber. The repo represented an attempt to implement ideas discussed in [a research paper](https://arxiv.org/abs/2609.16247v1), "The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It," that explored how LLMs respond to language about pain in ways that mirror human responses. Though there are ample opportunities for people to voice opposition to real-world human anguish, some online commenters expressed outrage at seeing text about suffering from a chatbot. One of the paper's authors [condemned the project](https://x.com/camhberg/status/2105331825631494426), as did numerous social media users. Anthropic – which has discretion in how it applies its terms and has said its rules target extreme behavior – did not immediately respond to a request for comment about the frequency of Claude abuse. Chris Donnelly, Searchable CEO and co-founder, endorsed Anthropic's change of terms, suggesting that it could help make human interaction more civil. "For me, Anthropic's policy change is a positive move," Donnelly said in a statement. "It should help people get more out of chatbots, and it helps curb the kind of derogatory behavior that previous research suggests can spill over into how people talk to each other." ® Get AI news in your inbox Daily digest of what matters in AI.