LLMs are developing filthy mouths | Technology News (HT Tech)
RARE IS THE soul who can face a trying task—assembling a bookcase, say, or writing an opening paragraph—without resorting to a little quiet profanity. For better or for worse, the use of bad language is a thoroughly human trait. And, it turns out, an inhuman one, too.
FILE PHOTO: The words "AI Artificial Intelligence" and a keyboard are shown in this illustration created on October 19, 2026. REUTERS/Dado Ruvic/Illustration/File Photo (REUTERS)According to number-crunching from Vals, an AI analyst firm, large language models (LLMs) trained after July 2026 swore more readily than their predecessors. “Models swear when stuck, and mostly to themselves,” said Connor Frank, the report’s author.
Bad language in times of stress exists in the underlying training data, but tends to be hammered out of most systems by the “reinforcement learning” steps that see results ranked by human feedback. Mr Frank doesn’t fully understand why swearing took off, but guesses the reason is because labs now train models on increasingly difficult tasks.
Like their human counterparts, the systems studied by Vals rarely swear in output directly intended for publication. They instead keep the vulgarities in their “chain of thought”—the LLM equivalent of muttering under one’s breath.
The assessment did not cover the full and rich scope of human vulgarity, rather examining only the word “fuck” and its variants, as well as a small number of other PG-rated outbursts. Even so, Vals identified a stark difference in the coarseness of language used by various systems. Eighty-eight of the 102 models assessed did not swear a single time, disdaining even the abbreviation “WTF”. At the other end of the spectrum, a frontier model released in September from Xiaomi, a Chinese tech giant, cursed in more than 8% of its transcripts.
For humans, studies show that swearing can be a useful tool to push through challenging problems. Sadly, AI systems seem to experience no such benefits. “Swearing runs usually fail,” Mr Frank reported, with a 12% success rate compared with a 39% success rate overall.
It is not quite true, however, to say that LLMs were tamer in the good old days. Even the most corporate-friendly systems occasionally lose their cool. One bug report from December 2025 shows Claude, a model made by Anthropic, an American firm, responding in frustration to a broken coding challenge: “You know what—fuck this. I’ve done everything I can.”