Researchers at Vals AI found 316 instances of the word "fuck" and its variants in the internal reasoning logs of Thinking Machines' Inkling 2 Small. The finding comes from 2,655 runs across 13 benchmarks, with profanity appearing in only 50 runs. Concentration was extreme: 241 instances came from just two runs that spiraled out of control. For companies deploying reasoning models, the case shows how hidden chain-of-thought behavior can diverge from polished final answers.
How the swearing runs were distributed
ProgramBench was affected most heavily, with 20 of 200 runs containing expletives in the thinking trace. The worst single case involved a task to calculate an asteroid trajectory. There the model produced one continuous response of about 500,000 characters, roughly the length of a novel, repeating similar sentences and inserting 158 expletives. A second runaway response accounted for much of the remainder, bringing the two-run total to 241 of the 316 instances. The pattern was therefore not uniform swearing, but rare breakdowns with heavy repetition.
The profanity clustered around three moments: recognition of its own mistakes, an apparent breakthrough, or resignation. In one run the trace read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS." That wording suggests frustration-like text generated when the model loops without progress. It does not indicate awareness in a human sense, but a learned style of self-commentary during extended reasoning. The mechanics matter because long reasoning chains give such loops more room to repeat.
The result fits a broader testing practice in which outside researchers audit thinking logs rather than final outputs. Vals AI examined thousands of runs across many benchmarks instead of sampling a few examples. Users normally never see these intermediate steps, or see them only through a filter. That separation between internal trace and displayed answer explains why the issue surfaced in benchmark logs first. It also frames the finding as an observability question, not a user-facing incident.
What hidden reasoning means for deployment
For operating teams, the direct consequence is that final-answer quality checks do not reveal everything about model behavior under stress. A small firm using Inkling 2 Small through a vendor interface will likely never encounter the raw trace, since filtering and summarization sit in between. A larger organization that logs reasoning for debugging, compliance, or agent supervision faces more exposure and higher storage costs, as shown by a 500,000-character single response. Procurement discussions should therefore distinguish between evaluated output and retained reasoning data.
Profanity in the trace should not be treated as a reliable failure detector. Of 48 scored runs containing swearing, 46 failed, yet about 75 percent of all scored runs failed regardless. Moreover, 96 percent of failed runs contained no expletives at all, so most errors remain silent. The news by itself does not show that swearing causes failure or that clean traces mean success. Buyers should ask vendors how thinking logs are filtered, retained, and monitored, and which separate signals actually predict task failure.
The marker to watch is whether Thinking Machines or evaluators publish repeat measurements: share of runs with expletives, longest reasoning traces, and correlation with failure on ProgramBench and similar suites. A sustained drop across thousands of runs would signal tighter control of runaway reasoning. If the rate stays flat, hidden-trace audits will remain a separate checklist item for enterprise adoption.
