Fired OpenAI researchers ask the board to preserve the ability to monitor AI reasoning

Jasmine Wang, Tomek Korbak and Mikita Balesni have urged OpenAI’s board and safety committees to preserve chain-of-thought monitoring and cooperate with outside auditors, the Wall Street Journal reports. OpenAI says it agrees with their recommendations and says their dismissals concerned information-handling rules, not raising safety concerns. The full letter is not public.

The technical issue predates the dispute. Some reasoning models generate intermediate text while working through a task. A separate monitor can inspect that text for evidence of cheating or harmful intent. OpenAI’s own research found this could reveal misconduct that was harder to detect from the agent’s actions and final answer alone. It also found that directly penalising suspicious reasoning could teach a model to conceal its intent without eliminating the misconduct.

Later OpenAI research formalised a distinction that matters here: can a monitor detect a bad outcome, and can it recover why the model behaved that way? Its evaluation suite tests controlled changes to tasks, identifiable reasoning steps and observable outcome properties. Readable prose is therefore not sufficient evidence that a model remains reliably inspectable.

Analysis: agreement that monitoring is valuable leaves the governing decision unresolved. What measured deterioration would delay a model, who may inspect the evidence, and who can require a pause? Preserving a useful warning channel is a concrete objective; treating it as a guarantee of safety would be a different and much stronger claim.