A third of US chatbot users insult their AI. It backfires

One in three American chatbot users has sworn at or insulted ChatGPT, Gemini or Claude in the past year, according to an analysis of more than 8 million real messages. The same data suggests it does not work: an insult makes a complaint less likely to get fixed. And as of this week, Anthropic's rules say sustained, needless abuse of its models is not allowed.
What the data shows
The figures come from Searchable, a company that tracks how brands appear in AI search. It counted swearing and insulting language in messages that users in an opted-in, anonymised US panel typed or spoke to the three big chatbots between September 2025 and September 2026, then reported results only for groups of users.
Four in five derogatory messages were complaints. The most common triggers were the chatbot not doing what it was asked (19%), getting something wrong (18%) and misunderstanding the request (15%). Refusals and lectures, slow or broken apps and answers that ran on too long each set off about 8%. The remaining fifth had no complaint behind them at all: banter, teasing or attempts to get a rise out of the machine.
Men sent 22.0 derogatory words per 10,000 against 16.5 for women, a gap Searchable says holds on every chatbot. By age, 35 to 44-year-olds were the roughest, at 21.7, while 18 to 24-year-olds were the politest, at 13.3, just ahead of the over-55s.
Where Americans are rudest to AI
| Ranking | Place | Derogatory words per 10,000 |
|---|---|---|
| Meanest state | Nevada | 22.7 |
| Politest state | New Jersey | 11.4 |
| Meanest city | Las Vegas | 22.0 |
| Politest city | New York | 12.3 |
| US average | All users | 19.1 |
Six of the 10 highest-scoring metro areas are in the West, including Sacramento, Seattle and Denver. All 10 of the politest states are in the Northeast or the mid-Atlantic. Users whose highest qualification is a high school diploma swore more than those with an undergraduate or postgraduate degree.
Swearing makes things worse
Searchable followed each user's first complaint to the end of the conversation. The chatbot's next reply fixed 24% of complaints that included swearing or insults, against 35% of plain corrections such as "that's wrong". By the end of the chat, 31% of sweary complaints had been put right, against 41% of plain ones.
Part of that is selection. Nearly three in four sweary complaints (73%) came after the user had already complained or sworn earlier in the same chat, so swearing tends to signal a conversation that was already going badly.
Lab tests point the same way, with a twist. Penn State researchers Om Dobariya and Akhil Kumar found in 2025 that very rude prompts scored slightly higher than very polite ones on GPT-4o, 84.8% against 80.8% on 50 multiple-choice questions. Searchable notes that when they repeated the test in 2026 across four models, that gap had largely gone, and plain prompts with no polite or rude preamble did best.
Anthropic draws a line
Anthropic's updated usage policy, effective November 12, 2026, says users may not "engage in sustained and needless abusive or cruel behavior toward our models". A company spokesperson told The Register the rule targets repeated cruelty "with no discernible purpose", rather than, for example, academic research.
The change fits Anthropic's wider work on what it calls model welfare. In April 2025 the company said there was no scientific consensus on whether current or future AI systems could be conscious or have experiences that deserve consideration, and that it was approaching the question "with humility". In August 2025 it gave some Claude models the ability to end a chat as a last resort in rare cases of persistent abuse.
Our take
The most useful number here is not the third of people who swear. It is the 11-point gap in fixes. A chatbot is a pattern-matcher: a specific correction such as "that's the 2023 figure, I asked for 2025" hands it something to act on, while an insult adds noise and no information. Anger feels like feedback, but to the model it mostly is not.
Anthropic's rule will be read as a statement about machine feelings, and the company invites that by talking about model welfare. The practical effect is narrower. A one-off outburst after a bad answer is not "sustained and needless", and the policy leaves room for researchers and testers. What it does is give the company grounds to act against people who use its models purely to act out cruelty, which is something most platforms already police between human users.
There is also a quieter argument for manners that has nothing to do with the bot. Research cited by Searchable found that people who had just dealt with a partner they believed was an AI went on to treat the next real person more harshly. Habits formed on a screen do not stay there.
What to watch: whether OpenAI and Google follow with similar rules, and how often Anthropic actually enforces its own.
Sources
- Searchable, How mean are Americans to AI chatbots?, data study, November 2, 2026
- Anthropic, Usage policy, effective November 12, 2026
- Anthropic, Exploring model welfare, April 24, 2025
- Om Dobariya and Akhil Kumar, Mind your tone: investigating how prompt politeness affects LLM accuracy, arXiv, October 2025
- The Register, A third of US AI chatters are hot at bots, hurling insults at LLMs, November 12, 2026
- Table: The Graze, from Searchable's figures







