Can Artificial Intelligence Make
Scientific Discoveries?

Abstract

Artificial intelligence is causing a paradigm shift in the scientific method.
Traditionally scientific discovery has been a task performed by human scientists using diverse processes from mathematical formalisms and mental
representations to abduction and moments of lucky Gestalt shift. After rationalist pursuits of a logic of discovery waned, accounts of scientific discovery
largely shifted from philosophy over to psychology, with discovery being
increasingly conceptualized as a human psychological process. Yet the application of increasingly successful artificial intelligence programs to science
has brought about claims in both philosophy and science that AI is making scientific discoveries. In this thesis I will examine several philosophical
conceptions of scientific discovery, particularly Kuhn’s taxonomy of puzzle-solving versus revolutionary science. Using these definitions, I clarify
the scope of AI’s ability to contribute to scientific discovery. I argue that
the discoveries which AI can make have distinct characteristics that correlate with Kuhn’s notion of the puzzle-solving discoveries that occur during
normal science. In contrast, I will argue that AI currently lacks capabilities
necessary for a kind of discovery that I call conceptual discovery. Current
constraints on the scope of AI’s contributions to discovery that I will discuss
include the frame problem, lack of a proper representation language, and in
particular, the challenge of formalizing abductive and analogical reasoning