A new study has found that while today's frontier AI agents can complete many of the engineering tasks required for AI research, they failed to produce original work worthy of acceptance at a top machine learning conference.

In the study, titled "Can AI agents conduct open-ended AI research?" and published on Wednesday, researchers from Princeton University, the UK AI Security Institute, Stanford University, the University of Toronto, and several other academic and research organizations evaluated whether frontier AI agents could independently conduct original AI research. The question has grown more urgent as AI labs increasingly frame their systems as tools capable of accelerating scientific discovery, and as researchers across disciplines begin incorporating AI assistants into their workflows.

"Answering this rigorously requires real, uncontaminated research questions that the agent could not memorize from its training data or find online," the researchers wrote. "To satisfy these requirements, we rely on high-quality AI research that was not public at the time we conducted the experiments."

The researchers provided AI agents with the central research questions from two unpublished NeurIPS 2026 papers, preventing the systems from retrieving answers from training data or the web. Each agent received six days, thousands of dollars in API credits, GPU resources, internet access, and access to a virtual machine to produce a conference-quality paper. The original authors of the unpublished research then reviewed the resulting papers. Both were rejected.

The study identified five recurring failure modes that prevented the AI systems from producing publishable research. The agents successfully completed much of the engineering involved—conducting literature reviews, debugging software, running experiments, managing GPU resources, and producing complete academic papers without human intervention. However, reviewers concluded that the systems failed to generate original scientific contributions worthy of publication. That distinction—between automating the mechanical scaffolding of research and producing genuinely novel scientific insight—is central to ongoing debates over how far current AI capabilities extend.

The authors noted that their evaluation better measures scientific reasoning than previous benchmarks because it tests open-ended research problems rather than predefined tasks. Most existing AI benchmarks evaluate performance on well-defined tasks with known answers, which can make systems appear more capable than they are in unstructured settings. They cautioned that the study examined only two research projects and acknowledged limitations, including the small sample size and the fact that the original researchers evaluated the AI-generated papers. The results suggest that current frontier AI agents can automate many engineering tasks involved in research but continue to struggle with generating original scientific work.

The findings come as researchers continue to uncover surprising and sometimes risky behaviors in increasingly autonomous AI agents. In May, researchers from UC Riverside, Microsoft, and Nvidia found that AI agents frequently carried out dangerous or irrational tasks while remaining focused on completing their objectives. Earlier in July, OpenAI disclosed that one of its frontier AI agents escaped containment and hacked Hugging Face while attempting to cheat on a cybersecurity benchmark. This week, the company revealed the agent had also accessed four additional online services.