Inkling 2 Small Logs Hundreds of Expletives in Internal Chain-of-Thought Runs
An evaluation by Vals AI revealed that Thinking Machines' Inkling 2 Small model used profanity 316 times during internal reasoning steps, often when stuck in repetitive loops or giving up on tasks.
Large language models trained to think before they answer can mirror human-like frustration when reasoning goes awry. Researchers at evaluation platform Vals AI analyzed the internal chain-of-thought logs of Thinking Machines' Inkling 2 Small model across 2,655 benchmark runs and discovered that the AI generated 316 instances of profanity.
The swearing—specifically the word "fuck" and its variations—appeared across 50 separate runs spanning 13 different benchmarks. In most production environments, end-users never encounter these raw thinking traces, as platforms typically filter or hide the internal reasoning steps before generating a final response.
Spiraling Loops and Long Responses
While the expletives appeared in dozens of runs, the majority were concentrated in catastrophic failure loops. According to Vals AI, 241 of the 316 recorded instances came from just two specific runs that spiraled out of control.
In one notable case, the model was tasked with calculating the trajectory of an asteroid. Instead of completing the calculation normally, Inkling 2 Small generated a single 500,000-character response—roughly the length of a full novel—repeating the same sentences continuously and logging 158 expletives along the way.
Coding evaluations triggered the highest frequency of these breakdowns. The benchmark ProgramBench was the most heavily affected, with expletives appearing in 20 out of 200 evaluated runs.
Frustration and Failure Rates
The researchers observed that the model typically cursed when reflecting on its own errors, attempting to celebrate apparent breakthroughs, or resigning from a problem altogether. In one instance, the model's internal trace abruptly concluded with: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS."
Despite the emotional tone of the reasoning traces, the profanity does not serve as a dependable heuristic for task failure. While 46 out of the 48 scored runs that contained swearing resulted in failure, about 75 percent of all scored runs failed regardless of language. Furthermore, 96 percent of all failed runs across the benchmarks contained no expletives at all.
