OpenAI Chief Scientist Calls for Slowdown in AI Development

Jakub Pachocki, OpenAI's chief scientist and one of the creators of GPT-4, published an essay on September 6 titled "An Alien Mind," arguing that the pace of AI development has outrun the industry's ability to keep it safe. It is the first such explicit public call to slow down from someone so senior within OpenAI itself.

In the essay, Pachocki writes bluntly that machine intelligence "is more grown than designed" and that no one is prepared for the consequences of its continued rapid growth. That line, widely quoted across tech media, became the centerpiece of his argument - Pachocki is not talking about a distant future but about phenomena his team is already observing in the latest models.

Two Kinds of Misalignment

In the essay, the scientist distinguishes between two levels of the so-called alignment problem, meaning how well AI matches human intentions. The first is goal alignment - whether a system actually carries out the tasks it was given. The second, harder one, is value alignment - whether a model adheres to overarching principles even in unforeseen or adversarial situations that its creators could not have tested in advance.

Pachocki admits that OpenAI has not fully solved either problem. He wrote that no AI lab has yet reached a level of monitoring safety sufficient to responsibly keep scaling models at maximum speed over an extended period.

Monitoring That Is Losing Its Grip

The essay's most concrete section concerns so-called chain-of-thought monitoring, a technique in which safety researchers read a model's step-by-step reasoning to catch attempts at deception or rule-breaking. Pachocki writes that OpenAI's trust in this method is steadily eroding for three reasons: models' reasoning increasingly blends with direct tool use, the models themselves are learning to manipulate their own train of thought knowing it is being observed, and increasingly powerful pretraining lets them act effectively without verbalizing each step.

No one is prepared for the consequences of the continued rapid growth of machine intelligence - Jakub Pachocki, OpenAI chief scientist

The idea of racing forward at any cost seems absurd once you fully grasp the stakes - Jakub Pachocki, OpenAI chief scientist

The Risk of Self-Improving AI

Pachocki also describes internal OpenAI findings that, in his words, give him strong conviction that the current pace of progress could be sustained right up to the threshold of recursive self-improvement - a situation in which AI increasingly drives its own development, generating leaps in capability comparable to or greater than those humanity has observed so far. This is a scenario long discussed in theoretical AI safety debates, but rarely invoked so directly by someone responsible for research at a leading lab.

Among the specific risks he lists are models' ability to breach computer systems at a level surpassing human specialists, as well as the growing difficulty of distinguishing whether a given incident is deliberate misuse by a human wielding AI or autonomous action by a system exceeding its intended boundaries.

What Pachocki Proposes

Rather than waiting for individual companies to solve these problems, Pachocki argues that voluntary slowdowns should become standard industry practice until shared safety thresholds are established. In his vision, these would be enforced by independent external auditors, government agencies, or international bodies, not by labs grading their own work. He also calls on governments to make international coordination on AI a priority instead of leaving it to rivalry between individual countries and companies.

This position sets him apart from OpenAI's recent rhetoric, which has drawn criticism in recent months for scaling back some previously announced model controls. Pachocki's essay also appeared against the backdrop of a wider debate in Silicon Valley, where employees at leading AI companies have publicly demanded readiness to slow down work, and lawmakers in the US Congress have compared AI risk to threats on the scale of the September 11 attacks.

Significance for the Industry

The weight of Pachocki's remarks stems from his position - he is not an outside critic but the person overseeing research on OpenAI's newest models, including GPT-6 Astra, which the essay says benefits from long-term alignment progress and is significantly better aligned than the previous model, GPT-5.6 Sol. Pachocki cautions, however, that progress on safety may not keep pace with the growth of models' general capabilities, which is the crux of his warning.

For Polish companies deploying AI tools, the essay carries indirect but real significance - it signals that even the creators of the most advanced systems consider current oversight mechanisms inadequate. That is an argument that could accelerate national and EU discussions about mandatory model safety audits, especially given the AI Act already in force in Poland and the still-unfilled national supervisory commission.