OpenAI bets on autonomous AI researcher by 2028

OpenAI plans to replace some of the drama of discovery with an autonomous AI researcher, shipping a first “research intern” by September and aiming for a fully automated multi-agent system by 2028.

The plan, laid out in a rare glimpse into the company’s long-range ambitions, is less about a single model and more about a north star: build an agent-based system that can tackle large, complex problems with minimal human prompting. In a high-profile interview, OpenAI chief scientist Jakub Pachocki described the path from a single, autonomous intern to a self-directing, multi-agent lab assistant that can read papers, design experiments, and propose next steps without daily handholding. The stakes are high: the team is betting that a sequence of increasingly capable autonomous agents can accelerate science much the way autotools and automation transformed software engineering a decade ago.

What makes this compelling—and risky—for the AI engineering community is not just the ambition, but the frame. The “autonomous AI researcher” is not a plug-and-play feature; it is a systemic shift toward agent coordination, self-critique, and safety rails in pursuit of credible, reproducible results. The September milestone is positioned as a practical, testable step: a researcher intern capable of handling a small set of problems, with the longer horizon of a fully automated lab partner by 2028. In that light, the project resembles building a self-driving research program—the car, for now, is still in a controlled lot.

The tech narrative here sits alongside a cautionary note about the limits of automation in high-stakes biology. The Download also spotlights the growing blind spots in psychedelic drug trials—an industry where automated literature review and experiment design could both help and hurt. The juxtaposition matters. If the same chemistry that makes a paper look good can mislead a trial under pressure, how do you ensure that an autonomous researcher won’t chase spurious signals or unsafe hypotheses? It’s a reminder that automation amplifies both insight and error, which is why governance, interpretability, and human-in-the-loop oversight aren’t optional.

Two practitioners’ takeaways stand out for product teams racing to ship AI-assisted research tools this quarter:

Build evaluation and safety into the loop early. The autonomous researcher will need robust, auditable decision logs, reproducible experiment plans, and clear failure modes. The aim is not a “black box scientist” but an auditable, incremental system whose outputs can be traced and challenged by humans. Expect new benchmarks that measure not just accuracy, but hypothesis quality, experiment validity, and reproducibility across agents.

Expect a compute-and-data triple constraint. Multi-agent coordination—literally dozens of micro-agents reading papers, drafting designs, and scheduling experiments—will demand substantial compute and access to diverse data sources. The cost isn’t just GPUs; it’s memory, latency, data licensing, and safety layers. For startups, the payoff is faster literature synthesis and hypothesis drafting, but at the risk of stale data, tool misuse, or overconfident conclusions without external validation.

Analytically, the analogy helps: imagine giving a lab bench, a whiteboard, and a telescope to a single AI that can both invent experiments and critique its own ideas. The result could be a dramatic acceleration of how researchers move from problem framing to actionable experiments—but only if the system stays honest about its uncertainties and remains tethered to human judgment during critical decisions.

What this means for products shipping this quarter is nuanced. Expect AI-driven research assistants to appear first as assistive tools—papers summarized, experiments proposed, and exploratory analyses automated—while safety, auditability, and governance keep a tight rein. The long arc toward a fully autonomous lab partner will unfold across years, with early pilots likely to set a framework for safer, more scalable AI-driven discovery.

Sources

The Download: OpenAI is building a fully automated researcher, and a psychedelic trial blind spot