Artificial intelligence and scientific infrastructures: Effects on resources, skills, and communications

• In Event: Infrastructure, Industrial Policy, and the Political Economy of AI

Abstract

Research and development (R&D) requires scarce and high-quality material and human resources, coordinated in specialized and dynamic organizations and ecosystems. This paper offers grounded insights into the effects of AI/ML on the resource requirements of science, and its effects upon scientific communication, through data from in-depth interviews with 32 academic researchers using AI/ML in the field of manufacturing and materials science (MMS). It illustrates a tendency for AI/ML computational modeling to shift resource requirements from physical experimentation to datasets and computational capacity, with attendant shifts in research capacity among compute-rich and compute-poor organizations. It also suggests that, while AI/ML are adding necessary skills to the conduct of research, they are not yet eliminating any.

Our findings demonstrate several ways in which AI/ML can affect the infrastructural needs of science. AI/ML facilitates more computation-efficient substitution for existing physics-based models, and permits computationally intensive modeling of phenomena which previously could not be modeled at all. They shift demand from resources required for physical experimentation to large, high-quality datasets and computational capacity. However, as reliable training datasets depend upon reliable physical theory, and ultimately upon empirical trial and validation, it is unclear how far this demand shift can go. AI/ML allow researchers to make more of empirical data and theory which already exist, not to eliminate empirical data and theory altogether. In human resource requirements, the modeling mode of use increases the requirements for computational skills among researchers, and does not (yet) eliminate requirements for other skills. Indeed, all research skills remain needed, with AI/ML more an additional tool in the toolbox than a complete replacement for any other research approach. In research communications, we reconfirm that LLMs show promise as tools for literature search, review, translation, and writing assistance, alongside potential problems of plagiarism and unreliability. More interesting is the finding that AI/ML as research tools alter the requirements for scientific communication and mutual assessment to increasingly prioritize novel, high-level model assessment skills and access to underlying datasets and codebases. Maintenance of effective scientific communication will require development of shared standards for archiving and reporting such.

Overall, AI/ML in MMS shifts infrastructural dependencies away from physical experimentation toward data, computational capacity, and new forms of expertise, while remaining anchored to empirical theory and validation. These technologies become interwoven with existing scientific infrastructures, deepening and extending the resource base required to do and evaluate science. At the same time, AI/ML alters the infrastructure of scientific communication by increasing reliance on shared datasets, codebases, and advanced model assessment practices, raising questions about robustness and reproducibility. Seen through an infrastructural lens, the significance of AI/ML lies not only in accelerating discovery, but in how it reshapes the taken-for-granted foundations on which future scientific knowledge will be built.

Author

• John Paul Nelson, Oregon State University
  Presenting Author