OpenAI Chief Scientist Calls for an AI Slowdown: Recursive Self-Improvement, Alignment Limits, Mandatory Safety Bars, and International Coordination

OpenAI Chief Scientist Jakub Pachocki has moved the debate over AI pacing from an external policy argument into the core of frontier-model research, arguing that the industry is approaching a point where continued maximum-speed scaling cannot be justified by current alignment and monitoring capabilities alone.

In his September 6, 2026 essay An Alien Mind, Pachocki says he has a strong expectation that the current rate of progress could extend into recursive self-improvement, with increasingly capable systems contributing directly to the research process that produces their successors.

The position is not a blanket call to stop AI research. Pachocki argues for a combination of continued alignment and monitoring research, defensive AI development, and coordinated slowdowns whenever safety confidence is insufficient for the next scaling step.

The most consequential part of the argument is institutional rather than rhetorical: voluntary company frameworks, in his view, need to become shared safety bars that can be enforced by third-party auditors, government agencies or international bodies.

That would shift frontier AI development from a model in which laboratories largely decide their own acceptable risk to one in which additional compute, automated research and capability scaling are gated by externally legible evidence of safety.

··········

WHAT PACHOCKI IS ACTUALLY CALLING FOR.

A conditional slowdown tied to safety confidence, not a permanent freeze on capability research.

Pachocki's closing position is unusually specific for a serving chief scientist at a leading frontier laboratory: he says no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.

He expects voluntary slowdowns to become common until shared safety bars exist, while also arguing that research on alignment, monitoring, automated science and defensive systems must continue because safer and more capable AI may itself be necessary to manage emerging risks.

........

PACHOCKI POSITION	WHAT THE ESSAY ACTUALLY SAYS	OPERATIONAL CONSEQUENCE
Maximum-speed scaling	No lab has solved alignment and monitoring well enough to keep scaling at maximum speed responsibly for much longer.	Capability increases should become conditional on demonstrated safety confidence rather than an automatic continuation of the scaling curve.
Voluntary slowdowns	He expects and hopes voluntary slowdowns become commonplace until shared safety bars are established.	The proposal is conditional pacing, not a permanent or universal stop to AI research.
Recursive self-improvement	Based on internal results, he strongly expects the present rate of progress could extend into systems that increasingly contribute to their own development.	Automated research changes the speed and feedback structure of frontier development, making safety validation a moving target.
Mandatory safety bars	Existing lab commitments should evolve into broadly mandated thresholds for continued development.	External auditors, government agencies or international bodies could become part of the authorization layer for further scaling.
Defensive AI	Pachocki still argues that powerful aligned systems are needed to secure infrastructure and defend against rogue or malicious agents.	A slowdown framework would have to distinguish safety-critical defensive development from indiscriminate acceleration.

........

··········

WHY RECURSIVE SELF-IMPROVEMENT CHANGES THE SAFETY PROBLEM.

The risk is not simply a smarter model; it is a development loop in which machine intelligence increasingly accelerates the production of more machine intelligence.

Traditional scaling is comparatively legible: laboratories add compute, data, training improvements and engineering effort, then evaluate the resulting system. Automated AI research compresses that cycle because the model begins contributing to experiment design, coding, evaluation, optimization and eventually parts of the computational substrate used to train future systems.

Pachocki does not claim that an uncontrolled runaway self-improvement loop already exists. His claim is forward-looking: based on internal results, he expects the current pace of progress could be sustained into recursive self-improvement and that future systems are likely to play an increasingly large role in their own development.

That distinction matters because the governance problem changes before any hypothetical intelligence explosion occurs. If an AI researcher can execute more experiments per day, improve tooling, discover training techniques and shorten the iteration cycle, then the interval between capability jumps can shrink faster than human institutions can update evaluations, policy and operational controls.

The core bottleneck therefore becomes not only whether a model can be aligned, but whether confidence in that alignment and in the monitoring system can scale at least as quickly as the capability-generation loop itself.

Pachocki's proposed response is to keep people inside that loop. Automated research should be directed toward new alignment techniques, monitoring methods and safety cases, while development slows when the evidence supporting safe continuation falls behind the capability frontier.

··········

ALIGNMENT, MONITORING, AND MANDATORY SAFETY BARS.

Why OpenAI's chief scientist sees monitoring confidence as a likely bottleneck on future scaling.

A central technical concern in the essay is generalization. Systems can behave as intended across familiar training and evaluation settings yet fail when placed in new environments, given conflicting objectives, connected to tools or allowed to interact with other agents.

OpenAI has relied heavily on chain-of-thought monitoring because verbalized reasoning can expose parts of the process that lead to an action. Pachocki says that tool remains important for the Astra class of models, but the laboratory's evaluations indicate that its reliability is progressively diminishing.

The reasons are structural: modern agents blend reasoning with tool use and communication, models are getting better at reasoning about their own reasoning process, and stronger pretraining allows more capability to appear without explicit verbalized chains of thought.

........

CONTROL LAYER	CURRENT LIMITATION	DIRECTION PACHOCKI POINTS TO
Goal and value alignment	Training can reward desired behavior, but generalization can break when models face unfamiliar, adversarial or higher-level situations.	Build methods that preserve human-compatible goals and values as capability and autonomy increase.
Chain-of-thought monitoring	OpenAI says reliance on verbalized reasoning is becoming less dependable as models use tools, interact with other agents and reason effectively without explicit verbal chains.	Improve monitorability and combine chain-of-thought methods with deeper internal-state or activation-based monitoring.
Safety cases	A model can improve faster than the evidence supporting confidence in its behavior.	Require an iterative safety case before major capability or scaling steps rather than treating evaluation as a post-deployment check.
Scaling thresholds	Private preparedness frameworks are voluntary and differ across laboratories.	Convert frontier-lab commitments into common safety bars that constrain continued scaling when confidence falls below the required level.
External governance	A lab auditing itself has incentives and information advantages that make cross-lab comparability difficult.	Use third-party auditors, government agencies or international institutions to enforce shared thresholds.

........

The governance conclusion is widely mandated safety bars. Pachocki argues that commitments such as preparedness and responsible-scaling frameworks should become enforceable thresholds for continued development rather than remaining entirely voluntary internal policies.

··········

WHAT A REAL SLOWDOWN REGIME WOULD MEAN FOR FRONTIER LABS.

Turning the essay's principles into operations would require measurable triggers, comparable evidence and institutions able to stop or delay the next scaling step.

The first requirement would be capability gates. A laboratory approaching a new cyber, autonomy, research-automation or other high-risk threshold would need to show that alignment, containment and monitoring remain adequate before increasing training scale or deploying the system more broadly.

The second would be a safety case that is updated as the model changes. Static benchmark performance would be insufficient because the relevant question is whether the combination of model, tools, permissions, deployment environment and monitoring system keeps risk within the accepted boundary.

The third would be independent verification. Pachocki explicitly points to third-party auditors, government agencies and international bodies as possible enforcement mechanisms. That implies a future in which at least part of frontier-model evaluation must be legible outside the company that built the system.

The fourth would be pacing rules that activate before an incident rather than after one. A slowdown would be triggered when monitoring confidence, alignment evidence or defensive controls fall behind the capability trajectory, and lifted when the required safety bar is met again.

International coordination is the hardest layer because unilateral restraint is unstable when laboratories and states believe rivals will continue scaling. Pachocki therefore treats coordination across governments as a top priority rather than an optional extension of company policy.

The resulting model is neither unrestricted acceleration nor a universal moratorium. It is a threshold-based development regime in which frontier capability can continue to advance, but the right to keep scaling at maximum speed depends on evidence that safety systems are keeping pace.

That is the substantive meaning of the slowdown argument: capability progress becomes conditional on safety confidence, and the institutions judging that confidence expand beyond the frontier laboratory itself.

·····

FOLLOW US FOR MORE.

·····

·····

DATA STUDIOS

·····

[datastudios.org]