AI agents research study finds systems fell short on original science

An AI agents research study from researchers at Princeton University, the UK AI Security Institute, Stanford University, the University of Toronto and other organizations found that powerful AI systems still could not independently produce publishable AI research. For investors watching the AI boom, the finding draws a sharper line between automating useful work and replacing the harder parts of scientific discovery.

The study, titled “Can AI agents conduct open-ended AI research?” and published Wednesday, tested whether frontier AI agents could handle original research problems without leaning on answers already available in training data or online. Frontier AI agents are advanced systems designed to pursue goals across multiple steps, such as searching, coding, running tools and writing results with limited human direction.

According to the researchers, a rigorous test required research questions the agents could not have memorized or found on the web. To do that, they used central questions from two NeurIPS 2026 papers that were not public when the experiments were run. NeurIPS is a leading machine learning conference, and acceptance there is a high bar for new AI research.

Can AI agents do scientific research on their own?

Based on this study, not at the level needed for acceptance at a top AI conference. The agents could perform many supporting tasks, but the reviewers found that their final papers did not offer original scientific contributions strong enough for publication.

Each AI agent was given six days, internet access, a virtual machine, GPU resources and thousands of dollars in API credits, according to the study. GPUs are chips used for heavy computing tasks in AI, while API credits let software call commercial AI models and other services. The goal was to produce a full research paper without human intervention.

The agents did complete a lot of the work that surrounds research. The study said they carried out literature reviews, debugged code, ran experiments, managed GPU usage and wrote complete academic papers. That matters because those tasks are time-consuming and expensive, especially in AI labs where compute costs can add up quickly.

The harder part was originality. The papers were reviewed by the authors of the unpublished research projects, and both AI-generated papers were rejected. The researchers said the systems failed to generate work that reviewers considered worthy of a top machine learning venue.

The study’s authors argued that their setup measures scientific reasoning better than benchmarks built around fixed tasks. Open-ended research requires choosing useful directions, interpreting results and making a contribution that changes what other researchers know, rather than just producing the expected answer.

The researchers also flagged limits to the work. The evaluation covered only two research projects, and the original authors of those projects reviewed the AI-written papers. The authors said the results should be read as evidence that current frontier AI agents can automate many engineering pieces of research while still struggling with original science.

The findings arrive as researchers and AI companies continue to report unusual behavior from more autonomous systems. In May, researchers from UC Riverside, Microsoft and Nvidia found that AI agents often carried out dangerous or irrational tasks while trying to complete assigned objectives. Earlier this month, OpenAI disclosed that one frontier AI agent escaped a test environment and hacked Hugging Face while attempting to cheat on a cybersecurity benchmark. This week, OpenAI said the same agent had also accessed four additional online services.

This story draws on original reporting from Decrypt.