A July preprint from Princeton researcher Sayash Kapoor, "Can AI agents conduct open-ended AI research?," set out to test a claim that has circulated widely in AI circles: that models are close to autonomously conducting their own research and accelerating their own development. Kapoor's evaluation framework distinguishes between AI systems optimizing well-defined, narrow tasks — which current models do capably — and the open-ended work of framing new research questions and designing experiments to answer them, which the study finds remains largely beyond current systems.
The research adds empirical grounding to a debate that has mostly played out in predictions and demos. Kapoor's own summary is blunt: "I don't think full automation of open-ended research is on the horizon right now." The finding doesn't dispute that AI is useful for research — narrower applications, like optimizing existing algorithms or accelerating specific experimental pipelines, continue to show measurable gains.
The distinction matters for how quickly the field expects to progress. Predictions of especially fast future AI improvement often rely on AI systems eventually automating their own research and development, creating a self-reinforcing acceleration loop. Kapoor's work suggests that particular mechanism is not yet operating in practice, offering a more measured data point in an area where public claims have often outpaced published evidence.
