When Science Locks In, Can AI Help?

Over the past century, the purpose of scientific practices has undergone a profound transformation, from an intellectual pursuit to a problem-solving enterprise, leading to an increasing risk of path dependency in scientific endeavors. Our research, analyzing 41 million research articles from the past six decades, emphasizes the concern of science lock-in by identifying and measuring two types of innovations in science: complementary and substitutive. Over the past six decades, the fraction of complementary innovation has risen to become the majority, while the fraction of substitutive innovation has declined to become the minority, revealing two different mechanisms of innovation: recombination and substitution. Amidst the increasing knowledge burden for innovation, while current AIs may excel in idea recombination due to their strong memory capabilities, their potential to contribute to the substitution mechanism remains uncertain, as they lack the ability to forget.