Anthropic and OpenAI have been touting their models' ability to speed up AI research. A new study from Princeton and the UK AI Security Institute put that claim to the test using unpublished papers, and came away with a different picture.

Why a new method was needed

Solid evidence for claims about automated AI research has been mostly absent, the authors argue. Existing evaluations either test agents on narrow, verifiable tasks or submit AI-generated papers to peer review — a process the researchers describe as "overstretched, stochastic, and suffering from poor review quality."

Instead the team built an approach they call shadow evaluation. An agent receives the core research question from an unpublished paper. The original authors, who spent months on that same question, then evaluate the result as conference reviewers would. Because the results are not on the web yet, the agent cannot fall back on training data.

How the experiment was set up

The researchers partnered with the authors of two NeurIPS 2026 submissions. The first examines how the personality traits of language models can be steered through their weights. The second develops a method called TabPFN that detects when a tabular prediction model meets deployment data differing sharply from its training data, tanking its accuracy.

The main experiments used Claude Opus 4.8 with Extra-High Reasoning. Each agent received:

  • Six days of wall-clock time
  • $3,000 in API credits
  • A GPU budget
  • Full access to a virtual machine and the open web

The agent ran inside a scaffold — a software environment that orchestrates model calls and provides tools — letting it delegate to subagents and monitor its own resource usage. It could also consult external AI review tools. The scaffold was OpenClaw, an open-source, vendor-neutral agent framework built by Austrian developer Peter Steinberger.

Two rejections

The original authors reviewed the finished papers as conference reviewers would and rejected both. One received a strong reject. They criticised poorly motivated data and experiments, unreadable prose, and no new contributions. One reviewer called the reasoning a "proof by example" fallacy that was "highly non-scientific." Another called the experiment choices "bizarre" and the results clearly the product of "post hoc choices."

Failure modes that keep showing up

Analysis of the agent logs reveals systematic weaknesses. The agents lacked judgment about what meets the bar for publishable research: they generated plausible hypotheses, then discarded them on the basis of small, hand-curated or synthetic datasets. They also failed at creative problem-solving — when initial hypotheses were falsified, they could not open a new line of attack.

The upshot is that frontier models can handle research engineering while failing at the parts of the research process that actually matter.