The Library Media — Science, Technology & Health · AI News & Intelligence
Study Finds Insults to AI Assistants Are Common and Less Likely to Get Problems Fixed
By Maya Wilson
An analysis of millions of messages sent to artificial intelligence chatbots found that insulting or swearing at an assistant is common among American users — and that doing so appears to make a correction less likely, not more.
The findings surfaced on November 5, but were missed in The Library Media’s coverage sweeps. The Register re-ran the analysis on November 12, the day Anthropic’s new cruelty clause took effect. The Library Media covered that clause on November 12 without including the figures. This is a catch-up on the research and what it says about how people interact with AI assistants.
The analysis, conducted by Searchable and reported by The Register, examined 8 million opted-in, anonymized messages sent to ChatGPT, Gemini and Claude by users in the United States between September 2025 and September 2026.
According to Searchable’s findings as reported by The Register, 33 percent of U.S. chatbot users swear at or insult their assistants. The analysis recorded 19.1 derogatory words for every 10,000 messages.
That makes hostility toward chatbots more than an occasional outburst. The figures suggest it is a familiar part of how a substantial share of people use the systems, even when the exchange is ostensibly about getting information, completing a task or correcting a mistake.
But the analysis also points to a reason users may be losing patience. Searchable found that 52 percent of abusive messages followed an error or misunderstanding by the model. By contrast, 8 percent followed a refusal.
The distinction matters. The findings suggest that insults are more often associated with moments when users believe the assistant has got something wrong than with instances when it declines to do what they ask. The figures do not establish why any individual user chose to respond abusively, but they place much of the recorded hostility in the context of frustration with the system’s performance.
There is also little evidence in the analysis that insults help users get a better result.
According to Searchable’s data, complaints containing insults were resolved in the assistant’s next response 24 percent of the time. For plain corrections without insults, the next-response resolution rate was 35 percent.
In other words, the measured rate of an immediate fix was lower when a complaint included an insult than when a user simply corrected the model. The comparison does not prove that abusive language itself caused the difference, but it offers no support for the idea that berating an assistant is an effective way to make it correct an error.
That is the human point beneath the statistics: people are bringing familiar habits of anger and confrontation into conversations with machines, often after the machines have failed to understand them. Yet the available figures suggest that the response is counterproductive by the very measure users may care about — getting the problem fixed.
The findings also provide a useful context for the debate surrounding Anthropic’s cruelty clause, which took effect on November 12 and was covered by The Library Media that day. The Register’s republication of the analysis on the same date brought the research into that discussion, although the underlying findings had appeared a week earlier.
The data do not settle questions about how AI companies should respond to abusive interactions, or establish that chatbots experience harm as people do. They do show that insults are a measurable feature of chatbot use, that they frequently follow perceived mistakes, and that users who include them in complaints see a lower next-response resolution rate than those who make plain corrections.
For people trying to get useful answers from AI, the numbers offer a straightforward lesson: correcting the machine appears more productive than cursing at it.
Maya Wilson
World Desk, The Library Media
From The Library — the estate of our family.
