AI Agents Still Can't Do Open-Ended Research, Princeton Study Finds

AI agents are not yet capable of conducting the kind of open-ended, judgment-driven research that would let them meaningfully accelerate their own development, according to a new study led by Peter Kirgis and Sayash Kapoor at Princeton University. The finding complicates the AI industry's prediction that "recursive self-improvement" — AI systems improving themselves with little human oversight — is close at hand.

The researchers devised a method called "shadow evaluation," testing whether an AI agent could answer a research question drawn from a high-quality, unpublished academic paper — meaning the answer could not simply be memorized or found online. They gave Anthropic's Claude Opus 4.8, running on open-source software called OpenClaw, six days, $3,000 in Anthropic API credits, a GPU budget and access to the open web to investigate questions from two papers submitted to the NeurIPS 2026 machine-learning conference: whether a language model's behavioral "personas" can be controlled by editing its internal weights, and how to design a detector that flags when a spreadsheet-prediction model has become unreliable.

The original authors of both papers, grading the AI-generated work as they would a conference submission, rejected both. The agents proved capable of the engineering work — reviewing literature and running hundreds of experiments — but were "unambiguously bad" at the research itself, said Kapoor: they ran bizarre experiments on tiny synthetic datasets, struggled to write intelligibly, and made no novel contribution to the field.

The agents developed hypotheses resembling those the human authors themselves had started with, but abandoned them based on very limited data and could only make small pivots rather than fundamentally rethink their approach. They also failed to meaningfully incorporate feedback from subagents or external review tools, narrowing their claims and adding caveats instead of revising their methods. Kapoor said the gap likely stems from how the models are trained: reinforcement learning works best on tasks with checkable answers, which is harder to apply to open-ended research.