Inkling 2 Small sometimes swears in its internal reasoning logs. A review by Vals AI found 316 instances of “fuck” or variations across benchmark runs, but the profanity does not reliably indicate whether the model has failed. Most failed runs contained no expletives, and users typically do not see these thinking steps.
Profanity was concentrated in a few runs
Vals AI researchers examined thinking logs from 2,655 runs across 13 benchmarks. They found profanity in 50 runs. The total reached 316 instances, but 241 came from just two runs that spiraled out of control.
One unusually long run followed a task to calculate an asteroid’s trajectory. The model produced a single 500,000-character response, roughly novel-length, repeating the same sentences and including 158 expletives. That example shows how a small number of extreme outputs can account for a large share of the total.
ProgramBench had the highest count of affected runs: 20 of 200. The source does not report comparable counts for each of the other benchmarks, so the overall pattern is clearer than any ranking beyond that result.
When the model swears
The logs suggest that Inkling 2 Small’s profanity often appears around its own perceived mistakes, apparent breakthroughs, or moments when it gives up. One run included the line: “FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS.”
That wording may sound like a candid status report, but the study’s figures show why it should not be treated as a dependable diagnosis. A model can express frustration in a log without that expression providing a useful or consistent account of the outcome.
Why profanity is a weak failure signal
Among scored runs, 46 of 48 runs with swearing failed. On its own, that figure might make profanity seem like a strong warning sign. But about 75 percent of all scored runs failed anyway, and 96 percent of failed runs contained no expletives.
The broader failure rate matters: when failures are common, profanity appearing in some failed runs does not make it a reliable way to identify them. The much larger share of failed runs without profanity also means that looking for swear words would miss most failures.
In practical terms, the logs reveal behavior, not a trustworthy alert. A reader who sees an expletive may be looking at a failed run, but the absence of profanity says little about whether a run succeeded. The figures do not support using the model’s language alone to sort outcomes.
What users are likely to see
These results concern internal thinking steps. Users typically do not see those steps, or see them only with a filter. That distinction matters: the reported profanity was found in the logs, and the source does not say that users routinely encounter it in ordinary responses.
The findings are therefore about how Inkling 2 Small’s internal text can look during benchmark work, including when runs become repetitive or unsuccessful. The extreme asteroid-trajectory example and the wider benchmark counts point to different aspects of the same issue: individual runs can produce striking language, while the overall data offers a more useful picture of how often it happens and what it can signal.
Across the 2,655 runs, profanity was limited to 50, and most failures were profanity-free. That combination makes the headline number memorable, but the more important takeaway is the limitation of the signal: swearing can accompany a failure, yet it does not reliably reveal one.