AI agents can’t yet do open-ended AI research that’s publishable

A new multi-institution study asks whether frontier AI agents can independently conduct open-ended AI research. Researchers from Princeton University, the UK AI Security Institute, Stanford University, the University of Toronto, and other organizations tested whether “AI agents” could produce a paper worthy of acceptance at a top machine learning conference. They restricted the setup to avoid memorized or publicly available answers by using the central questions from two unpublished NeurIPS 2026 papers. Each agent was given six days, thousands of dollars in API credits, GPU resources, internet access, and a virtual machine, then submitted a conference-quality draft. Result: the AI agents completed much of the engineering work—literature review, debugging, running experiments, managing compute, and producing full academic papers. But reviewers rejected both submissions, concluding the systems failed to generate original scientific contributions. The authors identified five recurring failure modes that prevented AI agents from producing publishable research. They also argue the evaluation better measures scientific reasoning than benchmarks based on predefined tasks. They caution that the study covers only two research projects and that the original researchers assessed the outputs. For traders: this is a tech-sector signal on AI automation limits rather than a direct crypto catalyst—yet it may shape sentiment around AI tooling, lab spending, and near-term AI-related risk appetite.