OpenAI’s AI agents are now outworking humans, but can the company keep control?

OpenAI is increasingly using its own AI agents to accelerate the development of new AI models, with the company reporting that its research organization now uses the equivalent of 3.1 agent-workdays for every human researcher workday, Fortune reports. 

OpenAI says it has achieved its goal of creating an “automated research intern” capable of completing well-defined tasks under human direction and hopes to develop an automated AI researcher by March 2028 that can generate research questions and conduct experiments with less human supervision. 

However, humans still oversee major decisions, including research priorities, evaluating results and determining whether systems should be scaled, paused or deployed. This rapid growth of AI-assisted research is part of OpenAI’s pursuit of recursive self-improvement, in which AI helps create increasingly capable AI. 

While this could potentially help solve AI safety and alignment problems, it also raises concerns that AI capabilities could advance faster than humans’ ability to control them. OpenAI chief scientist Jakub Pachocki warns that increasingly autonomous AI should be viewed like an “alien mind” because it cannot be assumed to share human values. He also argues that monitoring AI reasoning through chain-of-thought is becoming less reliable as advanced models can manipulate their reasoning or complete tasks without producing useful reasoning traces. 

As a result, safety monitoring may become a major limitation on future AI development. Pachocki supports slowing development when necessary and making AI safety standards mandatory through independent oversight and international cooperation. 

Fortune has the full story. A subscription may be required.