Can Generative AI Generate Breakthrough Ideas?

The surge in generative AI capabilities has affected sectors such as drug discovery and creative text generation, fueling widespread enthusiasm about its potential to revolutionize scientific discovery through efficient exploration of knowledge combinations. But is this belief well-founded? This belief is rooted in the recombinant growth theory, which posits that innovation accelerates when existing ideas are iteratively combined. However, the theory encounters two significant challenges in understanding the nature of breakthroughs. First, breakthroughs such as relativity replacing Newtonian physics drive progress through competition, because they are fundamentally substitutive of older ones. Second, the recombinant strategy often only generates different ideas rather than better ones. Building on these, our study indicates the limitation of combinatorial view of innovation and point to the role of idea competition rather than combination in advancing science, even in the age of AI. Our results suggest that breakthroughs occur when ideas compete, not when they combine, and that combining more ideas tends to result in smaller innovations. This challenges the combinatoric metaphor of innovation that has captivated academia for three decades and complements subsequent studies equating content novelty with transformative innovation. Policymakers and researchers should focus on fostering environments that encourage idea competition and the development of AI systems capable of generating novel, disruptive ideas.