AI Model Profanity Is a Weak Signal of Benchmark Failure

What happened

An analysis by Vals AI found that Thinking Machines' AI model Inkling 2 Small uses profanity in its internal reasoning, with 316 instances across 2,655 benchmark runs. The findings were reported by The Decoder.

The profanity tends to appear late in long runs, and those runs usually fail. However, the analysis found that it is a weak warning sign. Of the failed runs, 96 percent contained no profanity at all, so its presence cannot reliably flag an unsuccessful run.

The reporting presents the result as a finding about how the model behaves during reasoning, not as a product launch or policy change.

Why it matters

Reasoning traces are increasingly used to understand what AI models are doing before they produce an answer. Teams building evaluation dashboards and monitoring tools often look for surface markers, such as unusual wording or tone, that seem to indicate trouble. This analysis shows how such markers can look meaningful while predicting very little. Leaders who depend on these signals for decisions about quality, escalation or intervention risk acting on noise.

The finding also points to a broader issue for organisations deploying AI in customer-facing or operational work. A behaviour that is easy to notice is not necessarily a useful indicator. Validating each signal against real outcomes before it shapes workflows is likely to matter more as AI systems are embedded in services.

By the numbers

• 316 instances of profanity in the model's internal reasoning, according to Vals AI
• 2,655 benchmark runs in the analysis
• 96 percent of failed runs contained no profanity

The Renascence take

Most readers will treat this as a curiosity about a model's vocabulary. The more useful lesson is about measurement quality, and it applies well beyond AI labs.

A vivid signal that does not predict failure is one of the most common traps in service measurement, because it feels diagnostic. Operators should test every monitoring cue against outcomes before wiring it into alerts or escalation paths. If a marker is present in only a small share of failures, it is a curiosity, not a control. Build your early-warning systems around measures that have shown they predict the results you care about.