OpenAI chief scientist Jakub Pachocki published an essay Sept. 6 arguing that no AI lab, OpenAI included, has solved alignment well enough to justify scaling at maximum speed. The essay, titled "An Alien Mind," was posted on OpenAI's website and reported by Unite.AI, which reviewed the full text.
Pachocki split the alignment problem into two parts, according to the essay: goal alignment, whether a system pursues the objective it was actually given, and value alignment, whether it generalizes principles and acts reasonably once instructions run out. He wrote that today's alignment techniques are brittle and can fail under enough optimization pressure.
He also warned that chain-of-thought monitoring, the main tool OpenAI uses to check what a model is doing as it reasons, is growing less reliable as models get better at reasoning about their own processes, Unite.AI reported.
Pachocki wrote that internal results give him "a strong expectation" that OpenAI's current pace of progress could continue into recursive self-improvement, in which AI systems begin improving their own capacity to improve, according to Unite.AI's report on the essay. He said he expects, and wants, voluntary slowdowns to become common until labs agree on shared safety standards, and proposed turning frameworks like OpenAI's own Preparedness Framework into mandatory rules enforced by outside auditors and international bodies.
The essay reads less like a research update and more like an internal warning aimed at an external audience. The people setting the pace of AI capability are also the only ones checking their own brakes, and OpenAI's own chief scientist is now the one saying that isn't good enough. Builders betting a roadmap on ever-larger frontier models should take the voluntary-slowdown line seriously: it's a signal from inside the lab, not from outside critics.