AI model swore 316 times in its hidden reasoning logs
Vals AI researchers found Thinking Machines' Inkling 2 Small swore in 50 of 2,655 benchmark runs, according to The Decoder.
The facts
- Vals AI researchers read thinking logs from 2,655 runs across 13 benchmarks and found swearing in 50 runs, totalling 316 instances, according to The Decoder.
- 241 of those instances came from just two runs that spiralled out of control; one asteroid-trajectory task produced a single 500,000-character response with 158 expletives.
- ProgramBench was hit hardest, with 20 of 200 runs affected.
- 46 of 48 scored runs with swearing failed, but about 75 percent of all scored runs fail anyway, and 96 percent of failed runs had no expletives.
Case notes · Danny's take
Why it matters
The swearing is funny. The real story is what it shows about reading a model's reasoning. Those logs are evidence, and most people never look at them. Here the profanity turned out to be a poor warning signal: 96 percent of failed runs had none. Noise is not a diagnosis. Users usually don't see these thinking steps anyway, so what you see is a filtered version of what happened.
For builders
If you build on reasoning models, log and inspect the thinking traces, especially for runs that loop or run very long. A 500,000-character spiral costs tokens and time. Set output limits and loop detection instead of waiting for odd language to flag a failure.
Not financial advice. Do your own investigation.
This story was summarised with AI from The Decoder's reporting and checked against our editorial policy. Danny's take reflects his views. Disclaimer.
Danny Walter
Founder of Moon Republic, Volt Terminal and Moon Republic Academy. Author of The Market Is a Crime Scene and After the Applause.