The Promise and Perils of AI in Science
-
AI tools like ChatGPT could increase scientific productivity but pose risks like publishing more errors and reducing understanding.
-
The authors categorize 4 visions for AI in science Oracle, Surrogate, Quant, and Arbiter.
-
Over-reliance on AI tools can create "illusions of understanding" - believed depth, breadth, and objectivity.
-
The goal of science is understanding the world in all its complexity, which may be constrained by AI capabilities.
-
The authors don't oppose using AI in science but want more conversation about associated epistemic risks.