AI and Scientific Research: Three Questions About Data, Creativity, and Serendipity
What recent experiments with large language models reveal—and what they may never fully replace in the practice of science.
Testing Whether AI Can Do Science
Celina Zhao’s report in the Feb. 27, 2026 issue of Science, How will we know if AI is smart enough to do science, describes tests to determine how large language models can help scientific research. The tests described range from very detailed fact-checking to simulations that model how scientists formulate and test hypotheses.
It’s likely that such tests—initially encouraging—will become more robust, detailed, and realistic over time. Whether they can actually replace human scientists is another matter and raises questions about creativity and human thinking processes.
There is little doubt that much of the grunt work associated with data collection and analysis can be enhanced and streamlined by AI, as can the retrieval and interpretation of details from the published literature. In AI, journals, and the evolving knowledge divide I discussed how AI-based tools are being introduced that can interact intelligently with scientific literature. It’s likely that such tools will evolve rapidly. Still, several questions deserve careful consideration when discussing the role of AI in scientific research.
Question 1: Access to Data
The first concerns data access. If an AI tool is used by a researcher at any stage of a research project, will access to the published literature be sufficient, or will access to the underlying data that supported that research also be required?
There is no simple answer to this question, given the complexities associated with accessing and interpreting original research data. One issue involves how easily an AI agent might obtain and analyze such underlying research data, as well as the cost and resources required to do so. In theory, an intelligent AI system might seek access to original data if, for example, it detects inconsistencies or possible errors in what it is reviewing. Yet the logistics, cost, and resources involved—especially if the data reside in proprietary systems or involve human subjects—could prove problematic and require human intervention in the process.
Question 2: Creativity and the “A-ha Moment”
The second issue associated with AI involvement in research concerns creativity—how we define it and what level of creativity we expect an AI system to exhibit. Identifying hidden similarities or relationships is not the same thing as creating something genuinely new. Can AI assist in the creative processes involved in scientific research? The answer is almost certainly yes. But can AI generate something truly novel—something we might describe as an “a-ha moment,” a sudden insight or realization? I’m less certain.
Question 3: Serendipity in Scientific Discovery
The third concern relates to serendipity and the role it plays in scientific discovery. I first encountered this when I was a graduate student interviewing scientists about how they located and used articles in scientific journals. Some were adamant they did not want to rely solely on stored computerized searches or narrowly defined logical queries. They wanted exposure to research outside their immediate focus areas, precisely because they might encounter something from an “unrelated” field that could ultimately prove relevant to their own work.
Could an AI agent be programmed to scan unrelated areas of research in the hope of finding something useful for a particular researcher’s interest? It’s plausible that an AI system could be “fine-tuned” to look beyond a researcher’s primary focus area in search of potentially relevant insights and then learn from human feedback which discoveries were actually useful. Perhaps such an approach, which isn’t that different from scanning related areas out of curiosity, should then be called “planned serendipity.”
Who Benefits First?
When all is said and done, there is little doubt that AI can help—and already is helping—researchers around the world. One group that may benefit especially in the near term is graduate students, who have long played a central role in research projects by carrying out much of the behind-the-scenes work that scientific progress requires. When this happens, the educational process will also need to include how best to manage AI in research work.
Copyright © 2026 by Dennis D. McDonald
