Reframing AI’s role in scientific discovery
A paper in RSS Data Science and Artificial Intelligence opens a call for papers on how AI changes scientific practice, and where the limits of that change lie.
Overview
Artificial intelligence is creating new capabilities for scientific research, from analysing complex datasets and supporting modelling and simulation to helping researchers design experiments and navigate scientific knowledge. But advances in AI capability do not translate automatically into scientific progress. Realising their value depends on how new tools become integrated into scientific practice, the capabilities of the institutions and communities adopting them, and the relationship between AI systems and human scientific judgement.
The Accelerate Programme for Scientific Discovery was established at the University of Cambridge in 2020 to support the use of machine learning and AI across the sciences. Led by Jessica Montgomery, the programme has provided a practical setting in which to explore how AI capabilities diffuse across disciplines and become useful in research. Its experience has informed a wider research agenda at ML@CL concerned with the conditions for effective and responsible adoption of AI in science.
This work examines AI adoption as a systems problem, shaped by the interaction of technical capability with scientific knowledge and practice, data and experimental infrastructure, engineering capacity, skills, incentives, and institutional arrangements. A central question is how new capabilities spread through the scientific system. Research arising from Accelerate has examined the importance of interdisciplinary communities, open research practices, reusable tools, and data stewardship in creating a ‘diffusion engine’ through which methods and knowledge can move between machine learning and scientific disciplines. This shifts attention from isolated demonstrations of AI capability towards the infrastructure and practices needed for widespread scientific use.
More recent work considers the integration of AI into scientific workflows and the consequences for scientific judgement and understanding. As AI systems become capable of undertaking larger parts of the research process, questions arise about which tasks can usefully be delegated, how outputs should be evaluated, where human expertise remains essential, and what it means for scientific knowledge to be produced through systems whose reasoning may not be accessible to researchers.
Together, this work asks a broader question: what needs to be true of the scientific system for advances in AI to translate into better science? Accelerate provides one empirical foundation for answering that question, connecting experience of AI adoption in scientific practice with wider ML@CL research on AI, institutions and technological change.
In the process, Accelerate Science trained over 2500 researchers in how to use AI in their science, and incubated over 50 new AI for science projects.
Related People
Aditya Ravuri
Research Scientist, Isomorphic Labs
Ahmad Abu-Khazneh
Senior Machine Learning Engineer, Accelerate Programme, Cambridge University
Bianca Dumitrascu
Departmental Early Career Academic Fellow, Accelerate Programme, Cambridge University
Carl Henrik Ek
Senior Lecturer, Cambridge University
Challenger Mishra
Departmental Early Career Academic Fellow, Accelerate Programme, Cambridge University
Isaac Sebenius
MD Candidate, Harvard Medical School
Jess Montgomery
Director, ai@cam and Accelerate Science, Cambridge University
Justin Tan
PhD Student, Cambridge University
Neil D. Lawrence
The DeepMind Professor of Machine Learning, Cambridge University
Related News
3 September 2026
21 August 2024
Accelerating AI for science
A Royal Society Open Science paper sets out a framework for moving AI from isolated successes to wider use in research.
8 March 2023
AI for Science — an emerging agenda
A report from Dagstuhl Seminar 22382 sets out how machine learning and domain science can work together on complex scientific systems.
Recent Publications
Natural Language Processing markers in First Episode Psychosis and People at Clinical High-risk
Translational Psychiatry, 11(630):
Multimodal Graph Coarsening for Interpretable, MRI-Based Brain Graph Neural Network
IEEE 31st International Workshop on Machine Learning for Signal Processing (MLSP), :
Deep learning for Bioimage Analysis in Developmental Biology
Development, 148(18):
Solving Schrödinger Bridges via Maximum Likelihood
Entropy, 23(9):1134
Optimal marker gene selection for cell type discrimination in single cell analyses
Nature Communications, 12(1186):