A Japanese startup, Sakana, recently announced that its AI successfully generated a peer-reviewed scientific paper. While this claim is technically accurate, there are several key nuances to consider.
The Growing Debate Over AI in Scientific Research
Discussions about AI’s role in the scientific process are becoming more intense. Many researchers argue that AI isn’t yet capable of serving as a true “co-scientist.” Others see its potential but acknowledge that it’s still in the early stages of development. Sakana aligns with the latter view.
The company utilized its AI system, The AI Scientist-v2, to create a research paper, which was then submitted to a workshop at the International Conference on Learning Representations (ICLR)—a respected AI conference. Sakana claims that both the workshop organizers and ICLR’s leadership agreed to experiment with double-blind reviewing AI-generated papers.
AI-Generated Papers Submitted for Peer Review
Sakana collaborated with researchers from the University of British Columbia and the University of Oxford to submit three AI-generated papers for peer review. According to the company, The AI Scientist-v2 autonomously handled all aspects of the papers, from forming scientific hypotheses to conducting experiments, analyzing data, generating visualizations, and writing the entire text.
“We generated research ideas by feeding the workshop’s abstract and description into the AI,” said Robert Lange, a research scientist and founding member at Sakana, in an email to TechCrunch. “This ensured that the papers aligned with the workshop’s themes and were suitable for submission.”
Out of the three submitted papers, one was accepted for the ICLR workshop. This particular paper critically examined training techniques used in AI models. However, in a move to maintain transparency and respect ICLR’s conventions, Sakana voluntarily withdrew the paper before its official publication.
How Impressive Was This Achievement?
While Sakana’s AI-generated paper passing peer review may sound groundbreaking, the reality is more complex.
In a blog post, Sakana admitted that its AI system made several citation errors, including mistakenly attributing a method to a 2016 paper instead of the original 1997 research. This raises concerns about the reliability of AI-generated scientific content.
Additionally, the level of scrutiny the paper received was less rigorous than standard peer-reviewed publications. Because the company withdrew the paper after the initial review phase, it never underwent a meta-review—a final round of evaluation where workshop organizers could have potentially rejected it.
Another crucial point is that conference workshops generally have higher acceptance rates than main conference tracks. Sakana acknowledged in its blog post that none of its AI-generated papers met the standards required for ICLR’s main conference track publication.
Expert Opinions: AI Is a Tool, Not a Scientist
Several AI researchers weighed in on Sakana’s claim, cautioning against overhyping the results.
Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta, called the results “misleading.” He pointed out that Sakana’s team manually selected papers from a batch of AI-generated ones, meaning human judgment played a key role in picking submissions likely to be accepted.
“What this really shows is that humans and AI together can be effective,” Guzdial said. “It doesn’t prove that AI alone can drive scientific progress.”
Similarly, Mike Cook, a research fellow at King’s College London, questioned the rigor of the review process.
“New workshops like this one are often reviewed by more junior researchers,” he explained. “Also, this workshop focused on negative results and challenges, which makes it easier for an AI to convincingly write about failures.”
Cook added that AI’s ability to generate human-like prose makes it unsurprising that an AI-generated paper could pass peer review. However, he warned that this raises concerns about AI potentially flooding scientific literature with noise instead of advancing knowledge.
The Bigger Picture: AI’s Role in Scientific Research
The ability of AI to generate scientific papers raises both technical and ethical concerns. AI models are known to hallucinate facts and introduce errors, making many researchers skeptical about using them in serious scientific work. Additionally, if AI-generated research becomes common, it could undermine the credibility of peer review and dilute the quality of published findings.
“We need to ask whether this experiment demonstrates AI’s ability to conduct scientific research—or just its ability to persuade human reviewers,” Cook said. “There’s a big difference between passing peer review and actually contributing to a field.”
To its credit, Sakana makes no exaggerated claims about its AI’s capabilities. The company states that the experiment was meant to evaluate AI-generated research quality and encourage discussions about establishing clear norms for AI in science.
“There are important questions about whether AI-generated science should be judged on its own merits first, to prevent bias against it,” Sakana wrote. “Moving forward, we aim to engage with the research community to ensure AI’s role in science is meaningful and doesn’t devolve into a tool solely for passing peer review.”
Sakana’s AI-generated paper passing peer review is an interesting milestone, but it comes with several caveats. Human oversight was necessary, and the review process was not as stringent as standard peer-reviewed studies. The experiment highlights AI’s growing role in research, but also the need for clear ethical guidelines and scientific standards to ensure that AI enhances—rather than diminishes—scientific integrity.
Keep up with all AI news here.
