Thinking Machines model swears in internal logs, but fails to signal errors

Vals AI reported that Thinking Machines' Inkling 2 Small model generates profanity during its internal reasoning process, appearing in 50 out of 2,655 benchmark runs.reported The analysis, published by The Decoder, identified 316 instances of the word "fuck" or its variations within the model's chain of thought logs.reportedAn analysis found that the AI model swears in its hidden thinking process, but this does not help predict when it will make mistakes.Vals AI's analysis of 2,655 runs reveals that profanity in Inkling 2 Small's CoT is neither sensitive nor specific for failure detection. These occurrences typically emerge late in long execution runs that ultimately fail to produce correct results.

What was released or claimed

The Decoder reported on an analysis by Vals AI researchers who examined the model's thinking logs across 13 different benchmarks.reported They found that 241 of the 316 profane instances came from just two specific runs that spiraled out of control.reported One such run, tasked with calculating an asteroid trajectory, generated a single 500,000-character response that repeated sentences and included 158 expletives.

Who it affects

ProgramBench was identified as the most affected benchmark, with 20 out of 200 runs showing signs of this behavior.reported The report notes that users typically do not see these internal thinking steps directly, or they view them only through a filter. This suggests the issue is primarily relevant to developers and researchers auditing model internals rather than end-users interacting with final outputs.

What is still unknown

The material does not explain why the model chooses profanity in these specific failure modes. It is unclear whether this behavior is consistent across other models from Thinking Machines or if it is unique to Inkling 2 Small.reported The analysis also does not state whether this swearing impacts the model's performance on tasks outside the 13 benchmarks tested by Vals AI.reported

The practical upshot

The profanity does not serve as a useful warning signal for developers. While 46 of the 48 scored runs containing swearing failed, approximately 75 percent of all scored runs fail anyway.reported Furthermore, 96 percent of the failed runs contained no expletives at all, meaning the absence of swearing provides no guarantee of success, and its presence is not a reliable early indicator of error.reported

What the rules require

Product teams should not rely on internal chain-of-thought profanity as a heuristic for error detection. Compliance with robust monitoring standards requires tracking actual output correctness rather than linguistic artifacts in hidden logs. Teams must audit their filtering mechanisms, as users typically do not see these steps.