AI's role in scientific discovery raises questions about what counts as understanding
A new article from researchers at the University of Cambridge asks whether scientific knowledge produced by AI agents can be considered genuine understanding if humans cannot penetrate how those agents reached their conclusions. The paper, published this month in the Royal Statistical Society's journal Data Science and Artificial Intelligence, argues that AI's potential to create a new scientific approach demands coordinated responses from across the research community.
The authors include Neil Lawrence, DeepMind professor of machine learning at Cambridge, and Jessica Montgomery, director of ai@cam, the university's strategic mission to develop AI technologies that serve science, citizens and society. Their central question is direct: "Is scientific knowledge constructed and used by AI agents considered scientific understanding if it is impenetrable to humans, or does scientific understanding refer to an activity that is intrinsically human?"
What the paper argues
The article, titled AI for science: reframing AI's role in discovery, positions AI as more than a tool for accelerating existing research methods. The authors suggest the technology could shift how scientific knowledge itself is produced and validated - a change with consequences for funding bodies, universities, and working scientists.
Rather than treating AI systems as neutral instruments, the paper calls for researchers to examine what happens when machine-generated findings outpace human ability to verify them. That tension sits at the heart of the debate over whether scientific understanding remains a human activity or becomes something machines can claim.
Support gaps in the research community
The prospect of AI driving a new scientific approach puts pressure on institutions that have been slow to adapt. Researchers need training, computational resources, and clear guidance on when and how to trust AI-generated results. The article frames this as a structural problem, not an individual one - individual scientists cannot resolve questions about epistemic authority on their own.
For those working in AI for Science & Research, the implications are practical. Teams need to decide how to document AI contributions to findings, how peer review should handle results that cannot be independently reproduced by humans, and what skills researchers must develop to evaluate machine-generated claims. A structured AI Learning Path for Research Scientists addresses some of these gaps, but the paper suggests the challenge runs deeper than individual upskilling.
Why this matters for science and research professionals
If AI systems begin producing scientific results that humans cannot verify, the definition of expertise in research roles changes. Scientists and research managers will need to decide - likely sooner than expected - whether their job is to understand phenomena directly or to design, supervise, and interpret AI systems that do the understanding for them. The Cambridge paper does not answer that question, but it puts it squarely on the agenda for anyone who funds, publishes, or conducts research.
Your membership also unlocks:
