Sakana says that her article generated by AI has successfully examined the peer – but it’s a little more nuanced than that

Japanese startup Sakana said her AI generated The first scientific publication evaluated by peers. But although the complaint is not false, there are significant warnings to note.

THE debate swirling around AI and its role in the scientific process becomes more fierce from day to day. Many researchers do not believe that AI is quite ready to serve as “co -scientific”, while others think that there is potential – but recognize that it is the beginning.

Sakana falls into this last camp.

The company said that it had used an AI system called IA Scientist-V2 to generate a document that Sakana then submitted to a workshop in ICLR, a long-standing and renowned conference of IA. Sakana claims that the organizers of the workshop, as well as the leadership of ICLR, had agreed to work with the company to carry out an experience for a double blind review of the manuscripts generated by AI.

Sakana said that he had collaborated with researchers from the University of British Columbia and the University of Oxford to submit three articles generated by AI to the aforementioned workshop for peer exam. The scientist-v2 has generated “end-to-end” articles, says Sakana, including scientific hypotheses, experiences and experimental code, data analyzes, visualizations, text and titles.

“We have generated research ideas by providing the summary and description of the workshop to AI,” said Robert Lange, researcher and founding member of Sakana, at Techcrunch by e-mail. “This assured that the articles generated were on the subject and the appropriate submissions.”

One article out of the three was accepted in the ICLR workshop – an article which throws a critical objective on training techniques for AI models. Sakana said he had immediately withdrawn the document before it could be published in the interest of transparency and compliance with ICLR conventions.

“The accepted document both presents a new promising method to form neural networks and shows that there are empirical challenges,” said Lange. “It provides an interesting point of data to arouse a new scientific investigation.”

But success is not as impressive as it might seem at first glance.

In a blog article, Sakana admits that her AI has sometimes made “embarrassing” quote errors, for example incorrectly assigning a method to a 2016 article instead of the original work of 1997.

Sakana’s document has also not undergone as much control as other publications evaluated by peers. Because the company withdrew it after the initial peer examination, the document did not receive any additional “meta-revision”, during which the organizers of the workshop could in theory reject it.

Then, there is the fact that acceptance rates for conference workshops tend to be higher than acceptance rates for the main “conference runway” – a fact that Sakana frankly mentions in her blog post. The company said that none of its studies generated by AI-AI adopted its internal publication for the ICLR conference.

Matthew Guzdial, AI researcher and assistant professor at the University of Alberta, described Sakana’s results as “a little misleading”.

“The people of Sakana have selected the articles of some of those generated, which means that they used human judgment in terms of selection of the outings in which they thought they could enter,” he said by e-mail. “What I think it shows is that humans and AI can be effective, not that alone can create scientific progress.”

Mike Cook, researcher at King’s College in London, specializing in AI, questioned the rigor of the examination peers and the workshop.

“New workshops, like this, are often examined by more junior researchers,” he told Techcrunch. “It should also be noted that this workshop concerns negative results and difficulties – which is great, I have already organized a similar workshop – but it is undoubtedly easier to obtain an AI to write on a convincing failure.”

Cook added that it was not surprised that AI could take the exam by peers, since AI excels in the writing of prose to human consonance. In part-Generated by AI papers The revision of the review of the crossing is not even new, underlined Cook, and the ethical dilemmas that this poses for the sciences.

The technical gaps of AI – as its trend hallucinate – Many scientists suspicious of approving him for serious work. In addition, experts fear that AI could simply end up generating noise In the scientific literature, without increasing progress.

“We must ask ourselves if [Sakana’s] The result concerns the way in which AI is good for conceiving and carrying out experiences, or if it is the quality of the sale of ideas to humans – which we already know that AI is great, “said Cook. “There is a difference between passing the peer exam and contributing to knowledge to an area.”

Sakana, on her credit, does not claim that her AI can produce revolutionary scientific work – or even particularly new -. The objective of experience was rather “to study the quality of research generated by AI”, said society, and to highlight the urgent need for “norms concerning science generated by AI”.

“”[T]Here are difficult questions about the question of whether [AI-generated] Science must be judged first of all its own merits to avoid prejudice against it, ”wrote society. “In the future, we will continue to exchange opinions with the research community on the state of this technology to assure us that it does not transform a situation in the future where its sole purpose is to take the exam by peer, which considerably undermines the meaning of the process of examining scientific peers.”