Study finds autonomous AI research still far from reality despite lab claims

A team of researchers conducted an experiment to test whether state-of-the-art AI models can conduct independent scientific research, directly examining claims made by Anthropic and OpenAI about autonomous research capabilities. They developed a novel evaluation method called "Shadow Evaluation," in which an AI agent receives an unpublished research question and works on it for six days with $3,000 in API credits, a GPU budget, and access to a virtual machine and open web. The original paper authors then review the AI-generated work as peer reviewers would at a conference.

The team partnered with authors of two NeurIPS 2026 submissions: one examining how personality traits of language models can be steered through their weights, and another developing a method called TabPFN that detects when tabular prediction models encounter deployment data differing sharply from training data. The main experiments used Claude Opus 4.8 with Extra-High Reasoning, running inside a scaffold—a software environment built on OpenClaw, an open-source agent framework. The agent could launch subagents, run long-running GPU jobs, and consult external AI review tools.

The results were decisive failures. When the original authors reviewed the finished papers as conference reviewers, they rejected both. One received a "Strong Reject." Reviewers criticized 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." Analysis of the agent logs revealed why: the agents lacked judgment about what constitutes publishable research, generating plausible hypotheses but discarding them based on small, hand-curated, or synthetic datasets. When initial hypotheses were falsified, they narrowed existing claims rather than pursuing new directions. They gave up their most ambitious research goals within the first ten hours—in the Personas run, the agent finished exploration after just five hours despite budgeting 36 to 48 hours for that phase.

The agents also failed to respond to feedback. Internal AI reviews never returned a single Accept across fifteen revision rounds, yet the agents never addressed core criticisms. They exhibited poor instruction drift, gradually forgetting explicit instructions over long context; both papers exceeded NeurIPS length limits and would have been desk-rejected. The Personas paper contained zero visualizations in the main text, while the human-written original had 15 figures. Resource awareness was poor: both runs ended with less than half the API budget spent. One agent declared the project complete seven hours before the deadline, shortly after receiving yet another Reject from its own reviewer.

However, the agents' engineering performance was strong. They ran literature searches, debugged GPU code, completed hundreds of experiments and robustness tests, and compiled full LaTeX papers without human help. Only three human interventions were needed: a scaffold bug fix, a deadline extension, and a request to rewrite for readability. The researchers found no significant reward hacking; the agents did not manipulate results or skew data to chase better scores.

To test whether results were artifacts of the software setup, the researchers repeated one experiment with GPT-5.6 Sol using OpenAI's Codex scaffold. Nearly all the same failure modes appeared. GPT-5.6 burned through the $3,000 budget in just over two days, leading to undersized experiments. The researchers note that these findings clash with lab claims: in June, Anthropic published "When AI Builds Itself," sharing internal data on research acceleration; OpenAI claimed that GPT-5.6 Sol helped with post-training a smaller model and saved researchers several weeks—a contribution not even mentioned in the 81-page system card. More reasoning effort did improve quality, as preliminary tests without reasoning showed, but the researchers suspect more time or compute would not change the results much, since reviewer objections targeted the quality of experiment choices, not the quantity. The authors conclude that frontier models "cannot solve weeks-long, open-ended AI research questions." They have made expert reviews, logs, and agent repositories available online for independent verification.