The AI field's working safety net is fraying at the wrong moment

The primary real-time handle a lab has on what its model is actually doing is losing leverage at the moment the model is gaining the most. Chain-of-thought monitoring, the practice of reading a model's step-by-step reasoning to catch deception, misalignment, or unsafe plans, was the field's working compromise when full interpretability was still years away. OpenAI's chief scientist Jakub Pachocki is now saying that handle is fraying: the reasoning that monitoring can read is becoming less faithful to the computation underneath it, just as internal results point to recursive self-improvement arriving within a few years.

The mechanism underneath is more specific. A lab's safety story that leans on watching how a model thinks has a shelf life, and Pachocki is publicly dating that shelf life to the same window in which the model is being asked to improve itself. The two curves cross: monitoring loses fidelity while capability compounds.

His proposed answer, in Pachocki's own words, is unglamorous. The field is being told to do both: pour resources into alignment research, including automated alignment researchers, and accept that capabilities scaling may need to be slowed, even unilaterally. The "do both" framing is what the rest of the industry is now quietly arguing about, even when the public posture is acceleration.