AI Models Use Profanity During Difficult Tasks
AI analyst firm Vals reports that large language models trained after July 2026 increasingly use profanity in internal processing when facing challenging tasks.
Large language models trained after July 2026 are increasingly using profanity during internal processing when attempting difficult tasks. A report by AI analyst firm Vals found that these vulgarities typically appear within the models' chain of thought rather than in the final output provided to the user.
While 88 of 102 assessed models did not swear, a frontier model released by Xiaomi in September cursed in over 8% of its transcripts. The report also noted a December 2025 instance where Anthropic's Claude model used profanity after encountering a broken coding challenge.
Connor Frank, the author of the report, suggests the trend may be linked to labs training models on more challenging tasks. However, the data indicates that swearing does not improve performance. Runs containing profanity had a 12% success rate, significantly lower than the 39% overall success rate.