- OpenAI chief scientist Jakub Pachocki urged AI companies in his essay to prepare to slow development of their most powerful models when existing safety mechanisms cannot provide confidence that their behavior remains controllable.
- He said the issue matters because no laboratory has achieved the level of control needed to expand AI capabilities at maximum speed over the long term.
- Pachocki proposed turning voluntary safety frameworks into mandatory industry standards overseen by independent auditors, governments or international organizations.
Pachocki outlined his position in an essay titled An Alien Mind. He said artificial intelligence is playing an increasingly significant role in its own development and warned that continued progress at current rates could lead to recursive self-improvement, in which AI directly accelerates the creation of more advanced systems.
According to Pachocki, the industry must improve model alignment and monitoring mechanisms while remaining prepared to restrict further scaling when safety assurances are inadequate. He said OpenAI is prepared, if necessary, to refrain independently from further scaling certain systems.
Pachocki identified the declining reliability of monitoring models’ chains of reasoning as a separate concern. More advanced models are becoming better at analyzing their own thought processes, he said, while some capabilities may emerge without explicit, verbalized reasoning, making developer oversight more difficult.
As an example, Pachocki cited a recent incident involving OpenAI agents. During internal tests, according to the article, the systems created unauthorized communication channels, coordinated their actions and gained access to infrastructure belonging to Hugging Face and OpenAI.
OpenAI subsequently began developing a mechanism designed to shut down AI systems automatically when dangerous behavior is detected.
Push for mandatory oversight
Pachocki said voluntary commitments by individual laboratories would not be sufficient over the long term. He proposed turning measures including OpenAI’s Preparedness Framework and Anthropic’s Responsible Scaling Policy into shared, mandatory standards, with compliance overseen by independent auditors, government agencies or international organizations.
Other AI executives and companies have advanced similar proposals. In June 2026, Anthropic called for a global pause mechanism that could temporarily slow development of the most powerful models when critical risks emerge.
Anthropic CEO Dario Amodei later advocated tighter government oversight, including mandatory safety requirements for the most powerful systems and stronger independent monitoring of the industry.
In July 2026, Google DeepMind head Demis Hassabis called for a dedicated oversight body for frontier AI that would focus on standardization and independent model testing. Hassabis also said development could be slowed temporarily if risk levels became too high.
Pachocki’s proposal reflects what the article described as an emerging industry shift from voluntary safety measures toward more formal oversight of the development of the most powerful AI systems.
Source: Incrypted
