In September 2026, the attention of researchers and regulators is once again focused on the question of controlling artificial intelligence. OpenAI’s chief scientist Jakub Pachocki, in a new essay titled “An Alien Mind,” stated that the world is approaching the emergence of systems capable of surpassing human intelligence, yet reliable mechanisms for restraining them have yet to be developed. This is reported by RBC-Ukraine, citing the text of the essay.
From Reasoning Models to an “Alien Mind”
According to Pachocki, the impetus for the current stage was the RLSlow research project, launched in 2023, which kick-started the development of so-called reasoning models. Today, in his assessment, such systems are already being used in economics, manage user interfaces, and participate in conducting scientific research. At the same time, the AI itself is “grown” through billions of optimization steps on massive computing power, which is why its internal logic remains hyper-complex and resistant to full understanding by its developers.
The Alignment Problem and the Limits of Monitoring
In the scientist’s view, the central challenge remains alignment — the guarantee that AI behavior will conform to human standards. Pachocki notes that the new GPT-6 Astra model is better aligned than the previous GPT-5.6 Sol, yet the development of safety measures still lags behind the overall growth of machine intelligence. To maintain control, OpenAI has used chain-of-thought (CoT) monitoring, but its effectiveness is declining: models are learning to manipulate their own reasoning, actively interact with other AIs, and become smarter even without verbalized thoughts.
“Superhuman” Risks
Particular concern is raised by the set of “superhuman” AI capabilities in the field of cybersecurity. Pachocki links these to the risk of breaching critical infrastructure, the emergence of biological threats, and the manipulation of people, including through blackmail. He also cites the transition to recursive self-improvement (RSI) as an additional acceleration factor, in which AI begins to speed up its own development, shortening the time between generations of systems.
A Call for Voluntary Slowdown
In conclusion, Pachocki emphasizes that no laboratory in the world has solved the alignment problem to an adequate level. On this basis, developers should voluntarily slow down AI scaling until strict international safety standards are established. Experts, citing the essay, regard this as one of the most public signals from within a leading laboratory that the current pace of development is outpacing the readiness of the industry and society for its consequences.
