Artificial intelligence and the scientific method: How to cope with a complete oxymoron

Abstract

Artificial intelligence (AI) can be a powerful tool for data analysis, but it can also mislead investigators, due in part to a fundamental difference between classic data analysis and data analysis using AI. A more or less limited data set is analyzed in classic data analysis, and a hypothesis is generated. That hypothesis is then tested using a separate data set, and the data are examined again. The premise is either accepted or rejected with a value p, indicating that any difference observed is due merely to chance. By contrast, a new hypothesis is generated in AI as each datum is added to the data set. We explore this discrepancy and suggest means to overcome it.