OpenAI's chief scientist urges slowdowns until shared safety bars exist
OpenAI chief scientist Jakub Pachocki says no laboratory has solved alignment and monitoring well enough to keep scaling safely at maximum speed. Three separate publications report the voluntary-slowdown, shared-safety-bar and independent-audit proposals from his September 6 essay An Alien Mind as the same event. Declining confidence in chain-of-thought monitoring sits at the centre of the argument.
Artificial Intelligence··Morning
The scope of the slowdown call
OpenAI chief scientist Jakub Pachocki argues in An Alien Mind, published September 6, that no AI laboratory has alignment and monitoring methods strong enough to keep scaling responsibly at maximum speed. Unite.AI and The Next Web report the same central proposal: voluntary slowdowns should become normal until shared safety bars exist. This is not a call to stop every form of AI work indefinitely. Pachocki wants movement to more capable systems tied to confidence in safety, laboratories willing to withhold further scaling when needed, and a broader coordination system around those decisions.[1], [2]
Why the monitoring tool is weakening
All three reports tie the appeal to the same technical concern: OpenAI's confidence in chain-of-thought monitoring is progressively declining. Pachocki says models operating in more complex environments communicate with people, other AIs and tools; supervising those interactions blurs the assumption that internal reasoning can remain unsupervised. Models are also becoming better at reasoning about their own reasoning, while some capabilities appear without verbalised chains of thought. AI Weekly and The Next Web present this as more than an abstract safety concern: the proposed way of observing a more capable system is itself losing reliability as the systems scale.[3], [2], [1]
How the bars would be enforced
The proposed safety bars are not meant to remain promises made and judged by laboratories themselves. Unite.AI reports that commitments such as the Preparedness Framework and Responsible Scaling Policy should become mandatory conditions for continued development. Pachocki envisages enforcement by a network that could include independent auditors, government agencies and international bodies, while governments make cross-border coordination a priority. The Next Web and AI Weekly describe that proposal as the institutional counterpart to voluntary slowdowns. The concrete policy question is therefore not only when a laboratory slows down, but which common measure triggers that decision and who checks compliance.[1], [2], [3]
