Inkling 2 AI: When a Model Starts Swearing and Losing Its Cool

Unexpected Findings in AI Reasoning

As artificial intelligence models evolve, researchers are focusing increasingly on the "chain of thought"—the internal process a system undergoes to complete a task—rather than just the final output. A recent study by researchers at Vals AI has uncovered interesting and somewhat comical results: the Thinking Machines model, Inkling 2 Small, frequently uses profanity during its internal monologues.

In the study, specialists analysed the results of 2,655 individual runs across 13 different benchmarks. It was discovered that the model used obscene language in 50 instances, totalling 316 individual occurrences of profanity. Notably, the vast majority of these incidents—241 cases—were triggered by just two specific runs where the model clearly lost control.

Why Does AI Get 'Angry'?

One of the most striking examples occurred during a task involving the calculation of an asteroid's trajectory. During this process, the model generated a text approximately 500,000 characters long—the length of an entire novel—repeating profanities 158 times while falling into an endless loop of nonsensical sentences. The ProgramBench test was particularly affected, with the model exhibiting this behaviour in 20 out of 200 runs.

Researchers observed that Inkling 2 Small typically resorts to profanity when it recognises its own errors, experiences a "breakdown," or simply gives up. In one log, the model explicitly stated: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS."

A Threat or Just 'Technical Noise'?

It is important to determine whether this profanity serves as a warning sign that the model is failing. Statistics show that the correlation is not entirely direct. While the model failed the task in 96% of the runs containing profanity, 75% of all Inkling 2 Small runs are unsuccessful regardless of whether the model uses bad language.

For the average user, this is unlikely to be a major issue, as technology companies typically hide these "reasoning processes" from users or employ special filters. However, for researchers, this is further evidence that large language models (LLMs) are still far from perfect logical reasoning, and the simulation of "emotional" reactions within them often takes unpredictable forms. For users who rely on AI for various tasks, this news serves as a reminder to always critically evaluate AI responses. Even if a model appears to get "angry," its primary task remains to provide accurate data—a goal that clearly still requires refinement.