AI for Science
This was the focus of AI in Science: A Force Multiplier in Scientific Discovery, a two-day seminar held at the University of Oslo on 7 and 8 September. The seminar was tailored for postdoctoral fellows in DSTrain, UiO’s five-year postdoctoral programme in data science.
The DSTrain fellows work on research projects in their respective fields at institutes across the Faculty of Mathematics and Natural Sciences. The programme combines disciplinary research with training in data science, interdisciplinary collaboration and transferable skills.
Opening the seminar, Professor Morten Dæhlen from the Department of Informatics at the University of Oslo and TRUST, The Norwegian Centre for Trustworthy AI, described how AI is taking on new roles in science.
AI has long been used for what he called “number crunching and pattern discovery”. Today, it can also help researchers search and organize information, process data, build models, generate hypotheses and assess quality.
Dæhlen stressed that the researcher must remain at the centre. His presentation described a “human-in-the-loop” approach, combining AI tools with human analysis and insight.
“The question is not what AI can do, but what the researcher can do together with AI.”
Dæhlen attributed the statement to UiO researcher Geir Kjetil Sandve.
When AI tells you what you want to hear
Dæhlen used his own presentation to demonstrate both the possibilities and limitations of AI. He had spent around two and a half hours working with an AI assistant to identify and prioritize opportunities and challenges for AI in science.
Among the opportunities were faster discovery, automation of routine tasks, improved modelling and connecting knowledge across disciplines. The challenges included false discoveries, bias, lack of transparency and a verification bottleneck if AI produces scientific results faster than researchers can rigorously test them.
But while working with the AI, Dæhlen noticed another problem. He became concerned that the system was adapting to his views rather than challenging them.
“It’s actually telling me what I want to hear,” he said.
Dæhlen still found the tool useful, but stressed the need to remain critical when working with AI systems that tend to be positive and agreeable.
Over two days, the DSTrain fellows explored the use of AI in medicine, geoscience, climate research, education, creativity and law, as well as methods for trustworthy AI.