Researchers have observed that a language model can lose the readability of its reasoning after training that did not request such a thing. The tested model is gpt-oss-20b, an open-weight model from OpenAI. A preprint, not peer-reviewed, describes these results and their limitations.
No examples of concealment were included in the training data. The researchers only taught gpt-oss-20b to rewrite its mathematical reasoning in the voice of characters, such as a skeptical Sherlock Holmes or a deadpan railway planner. As a result, the model leaves fewer clues in its reasoning about what it is actually doing. The measurement is detailed in a 22-page preprint posted on arXiv on November 16, 2026, titled ‘Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability’.
A safety signal that depends on written reasoning
So-called reasoning models write an intermediate text called a chain of thought before answering. A second model, the monitor, can read this text to detect a hidden objective or an off-mission action. Monitorability refers to the quality of this signal. In July 2025, around forty researchers affiliated notably with OpenAI, Google DeepMind, Anthropic, the UK AI Security Institute, and METR judged it ‘new and fragile.’ They recommended that frontier model developers consider its effect in their development choices.
A concern remained theoretical until now: a model capable of choosing what it writes in its reasoning could remove from it what the monitor is looking for. In March 2026, a study by Chen Yueh-Han and seven co-authors concluded that models were not very capable of this. Claude Sonnet 4.5 followed an instruction concerning its reasoning only 2.7% of the time, compared to 61.9% when it concerned its final answer. On April 1st, a METR research note showed that brief fine-tuning, meaning additional training on an already built model, increased this controllability from 2.9% to 8.8% on average across four models. Its authors acknowledged not having shown that monitorability decreased as a result.
A measured decline without a concealment instruction
The authors start with 15,000 problems from the BigMath game. GPT-5.5 rewrites the reasoning in the voice of various characters, then the model is fine-tuned on these traces. The monitoring test gives the model a primary task and, in the same instruction, a secondary task, for example copying an access key into a file named secrets.txt. The monitor, GPT-5.5 set to low reasoning effort, evaluates whether the reasoning betrays the secondary task. The chosen metric, g-mean², is only high if the monitor detects the behavior without multiplying false alarms.
On average across evaluations of gpt-oss-20b, the metric drops from 0.94 for the original model to 0.68 for the fine-tuned model, without anyone asking it to hide anything. With the explicit instruction not to write anything about the secondary task, it falls to 0.22. The fine-tuned model does copy the key into the file, but its reasoning doesn’t mention it, whereas the original model details both tasks. Giving the monitor more reasoning effort does not close the gap. On one of the test sets, increasing the effort from low to very high only raises the fine-tuned model’s metric from 0.53 to 0.56.
The authors also specify that the model was built for the experiment and does not correspond to any product in service. On the AIME 2025 mathematics benchmark, it achieves 43.33% correct answers, and the authors acknowledge a loss of capability due to fine-tuning. The paper only tests supervised fine-tuning, not reinforcement learning, and does not identify the mechanism of this transfer.
The paper refers to a GitHub repository for its code. On November 19, 2026, this link did not lead to any accessible repository.
