The rise of machine learning challenges the classic scientific method of hypothesizing, predicting, and testing.
In the past, breakthroughs in science, like Isaac Newton’s laws of motion, were rooted in theory. Newton famously formulated the relationship between force, mass, and acceleration after an apple fell on his head, using data and experimentation to predict behaviors beyond the apple. However, today’s advancements, particularly in machine learning, have shifted how science is conducted.
Programs like Facebook’s machine learning tools, which predict user preferences, and AlphaFold from DeepMind, which predicts protein structures, don’t rely on traditional theories. Instead, they simply “work” without offering explanations for why or how. The accuracy of these predictions has left many questioning whether we need theory at all. They are deeply effective, yet theory-free.
This evolution has caused discomfort among those accustomed to the scientific method. Unlike in Newton’s time, today’s data-driven approach often bypasses hypotheses and direct causality in favor of correlations. In 2008, Wired’s Chris Anderson predicted the demise of scientific theory, suggesting that as data becomes more abundant, computers would uncover relationships far more efficiently than humans could through traditional theories.
Experts like Peter Dayan, a computational neuroscientist, argue that this data explosion has made traditional theory-making insufficient. We no longer possess the tools to write the theories that could describe the complex relationships emerging from vast datasets. With machine learning continuing to outperform traditional methods, we may indeed be witnessing the dawn of post-theory science.
