AI Models Swear More in New Releases
Vals, an AI analyst firm, reports that large language models trained after July 2026 use profanity more readily than their predecessors.
Updated on Nov. 19, 2026 in Artificial Intelligence

Vals assessed 102 models and found 88 never swore during testing, while models trained after July 2026 swore more often than earlier systems. A Xiaomi frontier model released in September 2026 cursed in more than 8% of its transcripts.
Why it matters
The findings suggest that profanity, which reinforcement learning from human feedback typically removes from published outputs, can reappear as models are trained on harder tasks. Developers and users of AI tools may need to consider how often these systems produce unfiltered language.
Vals assessed 102 models, with 88 never swearing, and a Xiaomi model released in September 2026 cursed in over 8% of transcripts, while swearing runs succeeded only 12% of the time versus a 39% overall success rate. Whether the increase is tied to training on harder tasks, as Connor Frank speculates, has not been confirmed.
The players
Anthropic
Anthropic is an American AI safety company that develops the Claude family of large language models.
Xiaomi
Xiaomi is a Chinese consumer electronics and smartphone company that has expanded into developing AI models.
The details
Vals examined the word "fuck" and its variants, along with a small number of other PG-rated outbursts. The models studied rarely swear in output intended for publication but keep vulgarities in their chain of thought, and a December 2025 bug report documented Claude responding to a broken coding challenge with profanity.
Timeline
December 2025: Bug report documents Claude swearing at a broken coding challenge.
July 2026: Cutoff after which LLMs swore more readily than predecessors.
September 2026: Xiaomi frontier model released, cursing in over 8% of transcripts.
Season Trajectory
Reinforcement learning from human feedback is the training step that typically removes profanity from published model outputs, so the rise in swearing among post-July 2026 models runs against that established pattern. The shift matters for how labs balance harder training tasks against output filtering.
Users of AI chat tools may encounter profanity in responses more often with newer models, which could matter for workplaces or products that require clean output. Developers integrating these models may need to add filtering or review steps to control the language generated.
The takeaway
Newer AI models can reintroduce profanity that earlier training methods had removed, so anyone deploying these tools in public-facing settings should check outputs rather than assume they are clean.
Further reading
For broader context on how these systems are built and tested, explore our Artificial Intelligence coverage.
Source note: This article includes information reported by Hindustan Times.



