“Those who expected to hear research findings presented by avatars or computer voices were disappointed. Instead, there were people in Zoom tiles explaining PowerPoint presentations. The online conference Agents4Science, which took place last week and was organized by researchers from Stanford University, looked quite ordinary. Nevertheless, it was a first: the focus was on the capabilities of artificial intelligence in science. Only papers with AI systems as first authors were allowed. Human researchers were only permitted as "supervisors."
Everyone uses it, but nobody wants to talk about it: The use of AI methods is widespread in science. It's no longer just about analyzing unwieldy large datasets. Large language models are now being used to summarize the current state of research, generate research questions, formulate hypotheses, plan the research process, write up the results, and evaluate the papers. However, according to a study from this May, admitting to using AI in one's research risks diminishing the perceived value of that work. Added to this is the confusion of guidelines and rules for the use of AI in research and teaching that has emerged in recent years. Sometimes the use of chatbots is permitted, sometimes even encouraged, and sometimes it's forbidden. The result is a great deal of secrecy.
It was all the more commendable, therefore, to reverse the perspective. Together AI, a cloud infrastructure provider, donated ten thousand dollars for the three best papers. This resulted in 315 submissions, primarily from computer science. These were first evaluated by several AI systems. Several systems were used because the models differed significantly in their assessments: Gemini 2.5 accepted many submissions, while GPT-5 rejected many. Not exactly a ringing endorsement for the use of these programs in the review process. The eighty highest-rated papers were then also reviewed by human experts. People submitted proposals, and 48 were ultimately selected for the conference; all are accessible on the conference website: https://agents4science.stanford.edu/index.html. The prize money went to research on the behavior of AI agents in digital marketplaces, the effects of towing fees in San Francisco, and the question of whether chatbots can write flawed but convincing research papers that deceive other AI agents.
The latter work addresses the biggest concern associated with the use of generative AI in science: falling for flawed but seemingly credible research. This could lead to an inextricable mess of half-truths in science. The fears proved justified: the bots accepted flawed research papers 67 percent of the time. Humans need to pay close attention to ensure that the programs adhere to the criteria given to them, said Fengqing Jiang, a doctoral student at Washington University, who presented the study. The algorithms proved helpful in checking details: Are all the numbers consistent?
Risa Wechsler, a professor of astrophysics at Stanford University, said that in her field of research, papers written with AI are neither interesting nor important. The programs lack a sense for good questions. This was confirmed by some of the presenters, who candidly reported using language models to work in areas outside their own fields of expertise, since the AI takes care of the content.
People feel better when they work with AI, but their results are often worse, said James Evans, director of the Knowledge Lab at the University of Chicago. He feared that the diversity of perspectives and questions that makes progress possible would be lost if research increasingly followed the dictates of algorithms. However, he hoped that the algorithms could also remedy this deficiency: if they could simulate different agents with different backgrounds who then work together on a problem.
Language models could also facilitate communication about data, methods, and techniques across disciplinary boundaries. This would make things easier, says Jonah Cool, Head of Life Sciences Partnerships at the AI company Anthropic. Perhaps in the future, it will be possible to transform papers on scientific methods into AI agents that then work together to tackle a problem using their respective approaches, suggests moderator and co-organizer James Zou. So far, the models are simply too agreeable to advocate for their own perspective.
Whatever may constitute real benefit and what remains a future project: if these attempts are openly discussed, science has already won. Agents4Science is therefore intended to be just the beginning of a series of conferences.” [1]
1. Zu nett für einen eigenen Standpunkt: Wider die Heimlichtuerei: Tagung über die Rolle von KI in der Forschung. Frankfurter Allgemeine Zeitung; Frankfurt. 29 Oct 2025: N4. MANUELA LENZEN
Komentarų nėra:
Rašyti komentarą