The challenge of AI-driven Automation in science
Artificial intelligence promises to revolutionise science by increasing the speed and effectiveness of the discovery process. At the heart of this promise are so-called “AI Scientists” – systems that can autonomously perform research tasks. To realise the dream of AI Scientists we not only need advances in AI, but also tools that allow us to understand and assess the capacities these agents display. The Leverhulme-funded project "Agents that can err: building the foundations for autonomous AI Scientists" will bring together expertise in philosophy, computer science, and the natural sciences to tackle this second challenge. More specifically, over the course of 48 months my team will develop the conceptual foundations and analytic tools required to assess AI agents for their ability to reason about experimental error. The goal of this project is to develop the foundations for error-reasoning agents (ERAs) and to thereby support the careful and effective introduction of AI agents into the research process.
Related publications:
Guttinger, S. "AI, automation, and the problem of troubleshooting in science." Forthcoming in The Role
of Artificial Intelligence in Science: Methodological and Epistemological Studies, eds. André
Curtis-Trudel, Darrell Rowbottom, and David Barack. Routledge
Guttinger, S. "Surveillance in the lab? How datafication is changing the research landscape." EMBO
Reports, https://doi.org/10.1038/s44319-024-00153-2 (2024)
Philosophy of Scientific Practice:
Reliable Data, replication, and the idea of control
Claims about a replication crisis in the experimental sciences have recently led to serious soul-searching amongst scientists, funders and other stakeholders. This led to calls for fundamental changes to the way research is being conducted and funded.
The aim of this strand of my research is to re-assess the very foundations on which these wide-ranging policy debates are based. A key driving force behind the crisis narrative is the worry that current practices in the experimental sciences, and in particular in fields such as pre-clinical cancer research, are not producing reliable data; there is not enough stringent control - on different levels of experimental practice - to ensure the production of trustworthy outputs.
However, the idea of control and how it can improve (or hamper) research is complex and not well-understood. The main objective of this strand of my research is to develop a deeper understanding of control practices in science and how they could inform the assessment and improvement of reliable data production.
Related publications:
Guttinger, S. "The limits of replicability." European Journal for Philosophy of Science, 10(10). (2020)
Guttinger S. "A new account of replication in the experimental life sciences." Philosophy of
Science, 86 (3), 453-471.(2019)
Guttinger S. "Replications everywhere." BioEssays, 40:1800055, (2018)