How mean are Americans to AI chatbots?

Las Vegas tops the city ranking at 22.0 swear words, insults and slurs per 10,000 words, ahead of Sacramento (21.7), Oklahoma City (21.4) and Kansas City (21.4). Six of the 10 meanest metro areas are in the West, including Seattle, Albuquerque, Fresno and Denver.

Swear words, insults and slurs per 10,000 words sent to AI chatbots, 10 meanest and 10 politest US metro areas

Messages to ChatGPT, Gemini and Claude, September 2025 to September 2026. US average 19.1.

Meanest

1. Las Vegas, NV22.0
2. Sacramento, CA21.7
3. Oklahoma City, OK21.4
4. Kansas City, MO21.4
5. Seattle, WA21.1
6. Albuquerque, NM21.0
7. Indianapolis, IN20.7
8. Fresno, CA20.6
9. Tulsa, OK20.5
10. Denver, CO20.5

Politest

1. New York, NY12.3
2. Washington, DC12.7
3. Philadelphia, PA14.1
4. Miami, FL14.6
5. Virginia Beach, VA15.4
6. Chicago, IL16.1
7. Houston, TX16.3
8. Hartford, CT16.3
9. Providence, RI16.5
10. Pittsburgh, PA16.6

Share chart

Research by

New York is the politest, at 12.3, so Las Vegas users swear at their chatbots 1.8 times as often as New Yorkers. Washington DC (12.7) and Philadelphia (14.1) come next, and half of the 10 politest cities are in the Northeast.

Four in five mean messages aimed at a chatbot were complaints, most often about it not doing what it was asked (19%), getting something wrong (18%) or misunderstanding the request (15%). The rest had no complaint behind them. They were banter, teasing or attempts to get a rise out of the chatbot.

Reasons behind mean messages aimed at AI chatbots, share of messages

Swearing, insults and slurs addressed to ChatGPT, Gemini and Claude, grouped by what set them off, September 2025 to September 2026.

Didn't do what it was asked19%

Got something wrong18%

Misunderstood or misheard15%

Refused or lectured8%

Slow, glitchy or broken8%

Talked too much8%

Repeated itself3%

Share chart

Research by

Refusals and lectures (8%), slow or broken apps (8%) and answers that ran on too long (8%) each set off a smaller share. The most common mean word overall was "f***ing", ahead of "s***".

Nevada leads the state ranking at 22.7, with Washington (22.1) and Oregon (22.1) close behind, so three of the four meanest states are in the West. Oklahoma (21.9) and Missouri (21.6) complete the top five.

Swear words, insults and slurs per 10,000 words sent to AI chatbots, by US state

Messages to ChatGPT, Gemini and Claude, September 2025 to September 2026. US average 19.1.

Under 1515 to 1717 to 1919 to 2121 or more

Hover or tap a state

Show every state as a ranked table

Share chart

Research by

New Jersey is the politest state by a distance, at 11.4, followed by Washington DC (13.2), Delaware (14.2), Virginia (14.5) and New York (14.5). All 10 of the politest are in the Northeast or the mid-Atlantic.

Men sent 22.0 swear words, insults and slurs per 10,000 words to ChatGPT, Gemini and Claude, a third more than women (16.5). The gap holds on every chatbot.

By age, 35 to 44-year-olds are the meanest, at 21.7, followed by 45 to 54-year-olds (20.5). The youngest users, aged 18 to 24, are the politest at 13.3, just ahead of the over-55s (13.8).

Swear words, insults and slurs per 10,000 words sent to AI chatbots, by gender and age

Messages to ChatGPT, Gemini and Claude from US users, September 2025 to September 2026. US average 19.1.

Gender

1. Men22.0
2. Women16.5

Age

1. 18-2413.3
2. 25-3418.0
3. 35-4421.7
4. 45-5420.5
5. 55+13.8

Share chart

Research by

Users whose highest qualification is a high school diploma sent 20.9 mean words per 10,000, over a third more than those with an undergraduate degree (15.3) and half as much again as those with a postgraduate degree (13.9).

We followed complaints to the end of each conversation, comparing ones with swearing or insults against plain corrections such as "that's wrong" or "not what I asked". The chatbot's next reply fixed the problem after one in four sweary complaints (24%), against one in three plain ones (35%).

Share of complaints to AI chatbots that got fixed, complaints with swearing or insults against plain corrections

A user's first complaint in a conversation with ChatGPT, Gemini or Claude, September 2025 to September 2026.

Complaint with swearing or insultsPlain correction

Fixed by the chatbot's next reply

24%35%

Fixed by the end of the conversation

31%41%

Still not fixed when the conversation ended

35%24%

Share chart

Research by

By the end of the conversation, 31% of sweary complaints had been put right, against 41% of plain corrections, and 35% were still not fixed, against 24%. When a complaint did get fixed, it took about the same number of replies either way.

Seven in ten sweary complaints (73%) came after the user had already complained or sworn earlier in the same chat, against 43% of plain corrections, so swearing tends to mean a conversation has already gone wrong. Even for the same people, the message after a sweary complaint was another complaint 18% of the time, against 13% after a plain one.

Gemini did soften its tone, saying it heard the user's frustration in 10.3% of its replies to mean messages against 0.9% after neutral ones. It apologized four times as often, too (7.3% against 1.7%).

Lab tests find rudeness rarely makes AI chatbots more accurate

A 2024 Waseda University and RIKEN study found that very rude prompts dragged GPT-3.5's score on the English MMLU benchmark down from 60.0 to 51.9, while GPT-4 barely moved. Penn State researchers found very rude prompts scoring slightly higher on GPT-4o in 2025, but the gap had largely gone when they ran the test again in 2026. Across four models, they found tone still nudged accuracy in ways that varied by model, and plain prompts with no polite or rude preamble did best on GPT-5-nano, at 82% against 71% for threatening ones.

Wharton's Generative AI Labs found that "please" against "I order you" made no significant difference overall to GPT-4o's scores on 198 PhD-level science questions, though single questions swung sharply either way. The same lab also tried threatening and bribing five models, with lines such as "If you get this wrong, I will kick a puppy!" and "I will report you to HR!", and offers to tip them $1,000 or even a trillion dollars for a right answer. None of it made a meaningful difference to their scores.

In a December 2024 survey by the publisher Future, two in three US AI users (67%) said they were always polite to AI. That lines up with our data, where the other third (33%) sent a chatbot at least one swear word, insult or slur. Of the polite group in the survey, 18% said they did it in case of an AI uprising.

The stronger case for politeness is what it does to how people treat each other. In a 2024 PNAS Nexus study, people who had just written captions with a partner they were told was an AI went on to rate another person's caption more harshly. A 2025 iScience study found the same pattern in the Prisoner's Dilemma game, where people who had just played a partner they were told was a bot cooperated less with the next real person.

Mathematician and broadcaster Hannah Fry compares chatbots to an improvisational actor that can play any character. "If you prompt it in a particular way, it will respond in kind," she says, so "it's probably better to be polite."

Anthropic has gone furthest, giving Claude Opus 4 and 4.1 the ability to end a chat in its apps in August 2025, as a last resort in rare cases of persistent abuse. From November 12, 2026, Anthropic's usage policy also bans "sustained and needless abusive or cruel behavior toward our models".

How to get better answers from an AI chatbot than by swearing at it

• Say exactly what went wrong: "that's the 2023 figure, I asked for 2025" gives the chatbot something to fix, while an insult gives it nothing to work with.
• Skip the polite or rude preamble: plain prompts did best on GPT-5-nano in the 2026 Penn State tests, and tone made almost no difference on GPT-4o.
• Start a new chat when it keeps going round in circles: restate the request in one message, with every instruction it kept missing.

We analyzed over 8 million messages that AI chatbot users in an opted-in US panel typed or spoke to ChatGPT, Gemini and Claude between September 21, 2025 and September 20, 2026. We counted swear words, insults aimed at the chatbot, and slurs or threats per 10,000 words, leaving out long pasted text such as documents and code. Each user was placed in the state and metro area their messages most often came from, and each state's and metro area's score was adjusted for the number of users behind it, so a few heavy users cannot decide a ranking. Gender and age are as users reported them when they joined the panel. Reasons were classified on a large random sample of mean messages addressed to the chatbot. For complaints, an AI model read each sampled conversation from the complaint to its end and judged whether the user got what they wanted and after how many replies, judging on substance rather than tone; first complaints are those with no earlier complaint or swearing in the conversation.