Skip to content

OpenAI is throwing everything into building a fully automated researcher

Key Ideas

  • OpenAI has made building a fully automated AI researcher its "North Star" for the next few years, pulling together work on reasoning models, agents, and interpretability.
  • The timeline: an "autonomous AI research intern" (handling tasks that would take a person a few days) by September 2026, and a full multi-agent research system by 2028.
  • OpenAI's chief scientist Jakub Pachocki envisions "a whole research lab in a data center" -- systems that work indefinitely and coherently like people, with humans setting goals.
  • The approach builds on Codex (agent-based coding tool) and extends coding problem-solving to general scientific research across math, physics, biology, chemistry, and policy.
  • Key risks include the system going off the rails, being hacked, or misunderstanding instructions; OpenAI's primary mitigation is chain-of-thought monitoring using other LLMs.
  • Even Pachocki does not expect systems "as smart as people in all ways" by 2028, but argues they do not need to be in order to be transformative.

Summary

Will Douglas Heaven reports on OpenAI's strategic pivot toward building a fully automated AI researcher, based on an exclusive interview with chief scientist Jakub Pachocki. OpenAI plans to build an autonomous research intern by September 2026 -- a system that can independently tackle tasks requiring a few days of human effort -- as a stepping stone to a full multi-agent research system by 2028 capable of handling problems too large or complex for humans. The approach extends the capabilities of Codex, OpenAI's agent-based coding tool, to general scientific problem-solving.

Pachocki argues that continued scaling of reasoning models, combined with training on complex tasks like math and coding competitions, will produce systems that can work for longer periods with less human guidance. He points to recent successes where GPT-5 discovered new solutions to unsolved math problems and advanced work in biology, chemistry, and physics. However, external researchers like Doug Downey of the Allen Institute for AI caution that chaining multiple tasks together causes error rates to compound, and current models still make many mistakes on scientific benchmarks.

The article also addresses safety concerns. Pachocki acknowledges "serious unanswered questions" about autonomous systems and describes chain-of-thought monitoring -- using other LLMs to review a model's reasoning scratch pad -- as the primary safeguard. He advocates for sandboxing very powerful models and calls for government involvement in governance, while acknowledging the tension with recent OpenAI decisions like signing a Pentagon deal after Anthropic declined.

Relevance to Economics Research

If OpenAI's vision materializes, it would fundamentally change the economics research production function. An AI system that can independently conduct multi-day research tasks would affect every stage of the pipeline from literature review through empirical analysis. The timeline (intern by late 2026, full researcher by 2028) suggests these capabilities could be available within the career horizon of current PhD students. The article also raises important questions about concentration of research capability and governance of powerful AI systems.