Open Source AI and Automated Science

Abstract

Our university has launched the first academic, large-scale automated science facility in the US. This facility will include shared data, compute, protocols, AI/ML, etc. that will empower open research. By emphasizing open data, open source software, and open source AI, this automated science facility represents a major opportunity for developing and sharing open research practices. This talk describes our university's approach and platform, along with potential opportunities and challenges related to reproducibility, defining openness for AI, and examining changes in research practice.

Early examples include the benefits of audit trails or provenance with automated tools, balanced out by challenges associated with the opaqueness of AI. Additionally, there is early evidence that training of students might change as automated tools are used for tasks previously implemented by people. One of our university professors has developed an AI "co-scientist" to manage laboratory tasks. We are partnering with a researcher at another university, who is conducting an ethnographic study of how research and education practices are changing.

Regarding AI, our university is leading a multi-institution program focused on defining openness in AI, and outlining the various benefits and possible risks. This work builds upon the early work and initial version of the Open Source AI Definition, which was designed through a community process led by the Open Source Initiative. This concept of openness in AI has been noted in various AI forums, including the Paris AI Action Summit.

This talk relates to the conference topic of how technology, including AI, is helping or hindering open research good practice, particularly as it relates to automated science, across a range of scientific disciplines and lab practices for both research and education.