Serge Belongie is professor in machine learning at the University of Copenhagen and Director of the Pioneer Centre for Artificial Intelligence in Denmark. He will be a keynote speaker at the STOA workshop ‘Generative AI and scientific development’ on 29 April 2025.
We’ve heard a lot about artificial intelligence (AI) transforming science in recent years. Can you give an example from your own work that exemplifies these changes?
Serge Belongie: I’m a co-Principal Investigator on a long-running research project called Visipedia, 
which has the goal of capturing and sharing visual expertise. We embarked on collaborations with the popular citizen science apps Merlin Bird ID and iNaturalist to incorporate AI functionality into their systems.
The occurrence data produced by these apps feeds into the Global Biodiversity Information Facility (GBIF) and has enabled numerous studies involving climate change, detection of invasive species, and geospatial modeling.
What have been the most notable AI-driven changes to the way science is conducted and communicated?
Serge Belongie: The recent Nobel Prize in Chemistry for protein prediction is, of course, a very high profile example of AI-assisted scientific discovery. The approach used in that groundbreaking work is emblematic of many modern research approaches that draw heavily on data-driven machine learning approaches, building and testing models using millions of noisily labeled observations.
Human scientists remain critically important in terms of seeing the big picture, thinking outside the box, and connecting the dots with insights, but AI is also starting to help with hypothesis generation and experiment proposal.
In terms of science communication, popular Large Language Models (LLMs) like ChatGPT and Claude are increasingly used by researchers and journalists for writing, editing, and summarizing findings in a variety of audience-appropriate ways. NotebookLM, a tool from Google that generates a two-host podcast from a scientific paper, has also become a popular means of disseminating ideas.
What are the advantages of using AI in science?
Serge Belongie: LLMs are significantly lowering the barrier of entry for people to write code for everything from statistical analysis to iPhone apps. As a result, a molecular biologist, for example, with no background in computer science, can now write nontrivial code to assist in drug discovery.
AI will allow small teams of researchers with limited budgets to make more progress, focus on the core scientific problems rather than arcane programming language syntax, and discover promising patterns in ever-growing amounts of data in each field.
On the other hand, what are the risks of using AI in science?
Serge Belongie: The increased access to ever more powerful AI tools has given rise to an “AI slop” problem, such as fake conference papers, confabulated data, plagiarism, and citations of non-existent references.
The most popular AI approaches today have well known problems involving confabulation and, more generally, uncertainty quantification. That means they must be used in the hands of scientists with high AI literacy. In the wrong hands, these techniques can produce wild goose chases with regrettable carbon footprints.
AI is helping, above all, with relatively menial tasks. Do you worry that a distancing from elemental tasks could reduce the capacity of human scientists to come up with new ideas?
Serge Belongie: On the contrary, I foresee that scientists will be able to focus more on the core scientific questions. That said, it’s a blessing and a curse that LLMs have a confabulation problem. It’s a curse because we wish these models wouldn’t make such mistakes. But it’s a blessing because it means scientists need to think carefully about unit tests, experimental design, and cultivating trust in readers and reviewers.
The rise of AI-generated and peer-reviewed scientific papers has sparked significant debate. Is it a problem that scientific papers are written and/or reviewed using LLMs? If so, why?
Serge Belongie: It is a significant problem, since science is built on a foundation of peer review, and these reviewers overwhelmingly volunteer their time for the benefit of their community. When authors and/or reviewers get lazy or sloppy with AI tools, it erodes trust and saps motivation from scientists who may already feel overworked and underpaid.
Do you think certain parts, or perhaps all, of scientific research could become entirely automated in the future, reducing the need for human scientists?
Serge Belongie: Humans – scientists or otherwise – are prediction machines. We move about the world, taking in stimuli, anticipating what might happen, acting on the data between that prediction and the reality that unfolds. We do this when driving and we do this when listening to music.
Machines are, arguably, even better at prediction. They are tough to beat when they have enough data and enough compute. But humans have abilities beyond prediction that machines have yet to capture. Humans excel at making leaps of intuition, connecting the dots between seemingly unrelated problems.
A nice example of this is when Leslie Lamport took inspiration from special relativity in his seminal work on distributed computing. AI is at a stage of development in which it can assist in the creative process, by suggesting hypotheses or promising directions of solution, but it’s still the case that clever people are needed to orchestrate such systems.
Do you think AI can, or could ever, rival human scientists in creativity? If not, is there a risk that AI-driven science will become more uniform and less innovative?
Serge Belongie: Some R&D teams are working toward this future. Sakana.ai, for example, has the aim of fully automating open-ended scientific discovery by means of an “AI Scientist.” Their high profile experiment of submitting a paper to an ICLR workshop ran into some hurdles, however, that suggest the path to complete automation is still science fiction. So I don’t foresee a Nobel Prize going to an AI Scientist anytime soon.
What can be done at an EU-level to reap the rewards of AI in science, whilst mitigating the risks?
Serge Belongie: Training Large Models – for language, vision, or other modalities – is notoriously compute-hungry. This means that European research labs need low-friction access to supercomputing facilities.
One solution is to lower barriers to supercomputing access, e.g., via the AI Factories Initiative.
Another is to increase wages for scientists who work at the intersection of Science and AI. Finally, Europe should seize the “Brain Gain” opportunity from the US that has emerged from recent geopolitical shifts.