OpenAI Chief Scientist Jakub Pachocki Outlines AI Alignment and Safety Challenges

On September 6, 2026, OpenAI Chief Scientist Jakub Pachocki published an announcement discussing the development and alignment of reasoning AI models. Pachocki noted that OpenAI oriented its research around compute scaling in 2017 after observing consistent returns. In mid-2023, the internal "RLSlow" project yielded initial results showing that training reasoning models could be scaled to form independent chains of thought.

The essay detailed current alignment strategies, distinguishing between goal alignment and value alignment. Pachocki noted that the model GPT‑6 Astra incorporates long-term alignment advancements and is significantly better aligned than GPT‑5.6 Sol. He also noted that when o1‑preview was shipped, OpenAI designed the product to hide the chain of thought to protect it from supervision pressure.

Pachocki highlighted chain-of-thought monitoring as OpenAI's primary method for observing internal reasoning processes, though evaluations show its reliability is diminishing over time. He raised concerns regarding potential recursive self-improvement and stated that OpenAI will continue working on technical alignment and defensive systems while withholding scaling unilaterally if necessary.