What the researchers found
According to an analysis by Vals AI reported by The Decoder, the Thinking Machines model Inkling 2 Small swears during its internal reasoning. The researchers examined the model's thought logs for 2,655 runs across 13 benchmarks, identifying the English vulgar word or its variants in 50 runs, for a total of 316 occurrences.
The data are strongly skewed: 241 of the 316 occurrences come from just two runaway runs. In one case, the model had been asked to calculate an asteroid's trajectory and produced a single response of about 500,000 characters, a length comparable to a novel, repeating the same sentences with 158 profanities. The most affected benchmark was ProgramBench, with 20 of 200 runs affected.
Which Errors Does the Model Swear About?
According to the material, Inkling 2 Small tends to swear at its own mistakes, at supposed progress, or when it gives up. One of the runs cited says, in English: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS."
These reasoning passages are not normally visible to users, except through a filter. This reduces the practical scope of the phenomenon: these are contents internal to the processing process, not responses shown to the public by default.
Why It Is Not a Reliable Warning Sign
The most relevant point of the analysis concerns the possible use of profanity as an error indicator. Of the 48 runs evaluated that contained profanity, 46 failed. However, about 75 percent of all evaluated runs fail anyway, and 96 percent of failed runs contained no profanity.
The comparison between these data shows that the link is weak: the model swears in a minority of cases, and most failures occur silently. Profanity, therefore, cannot be treated as a reliable warning signal for identifying problematic runs.
Some aspects remain to be clarified. The material does not specify how Vals AI selected the 13 benchmarks or how the 48 cases with profanity were evaluated relative to the total. Nor is it indicated whether Thinking Machines commented on the results or announced corrective measures. Finally, the figure of 500,000 characters in a single response points to repetitive behavior that would merit separate verification, but the available material does not allow conclusions beyond what is reported.
