The use of artificial intelligence (AI) in science has the potential to revolutionize the field, from helping doctors find early markers of disease to aiding policymakers in avoiding decisions that lead to war. However, a growing body of evidence has revealed deep flaws in how machine learning is used in science, implicating thousands of erroneous papers across various disciplines. To address this issue, an interdisciplinary team of 19 researchers, led by Princeton University computer scientists Arvind Narayanan and Sayash Kapoor, has published guidelines for the responsible use of machine learning in science. The team aims to improve scientific standards and reporting practices to prevent crises of credibility in research.
The authors believe that the current crisis of credibility, stemming from the use of machine learning in science, could become more serious than past replication crises, such as the one that emerged in social psychology over a decade ago. With machine learning being adopted across many scientific disciplines without universal standards safeguarding the integrity of these methods, the authors emphasize the importance of addressing this issue before it escalates further. Their guidelines, detailed in a paper published in the journal Science Advances, seek to stamp out this crisis of credibility and establish best practices for the responsible use of machine learning in research.
The checklist developed by the team focuses on ensuring the integrity of research that utilizes machine learning. Transparency is a key component of the guidelines, requiring researchers to provide detailed descriptions of their machine learning models, including code, data used for training and testing, hardware specifications, experimental design, project goals, and study limitations. The authors stress that these standards are flexible and can accommodate a wide range of nuances, such as private datasets and complex hardware configurations, in order to promote reproducibility and validate research claims.
While adherence to these new standards may slow down the publication process for individual studies, the authors believe that widespread adoption of these guidelines could ultimately increase the overall rate of discovery and innovation in science. By focusing on the quality of published papers and ensuring they serve as a solid foundation for future research, the authors hope to accelerate the pace of scientific progress. They emphasize the importance of prioritizing scientific progress over simply publishing papers, as errors in research can lead to wasted time, money, and resources with potentially catastrophic downstream effects.
In working towards a consensus on the guidelines, the authors aimed to strike a balance between simplicity and comprehensiveness, making the standards accessible to a wide range of researchers, peer reviewers, and journals. They suggest that researchers can use the guidelines to improve their own work, peer reviewers can assess papers more effectively, and journals can adopt the standards as a requirement for publication. By promoting honesty and integrity in scientific research, the authors hope to mitigate the prevalence of avoidable errors in the scientific literature and uphold the credibility of machine learning-based studies.
