The Myrmidon Wager: AI and the Relative Importance of Efficiency and Novelty in Science

If an oracle appeared offering to grant us enormous efficiency of scientific research at the cost of its novelty, should we accept the deal? The answer depends, in part, on how efficiency and novelty contribute to scientific progress. This tradeoff increasingly resembles one posed by the widespread adoption of artificial intelligence (AI) in science: AI tools can make individual researchers vastly more efficient, but also exhibit strikingly homogeneous judgements about which new ideas are worth pursuing. I call the wager that the former effect will outweigh the latter the \emph{Myrmidon wager}. Here, I study this tradeoff using a model of innovation based on the Urn Model with Triggering, in which new discoveries expand the set of possibilities for future discovery. The model shows that efficiency is proportionally less important than creativity in the long-term development of novel ideas. The model also helps to motivate a mechanistic account of an emerging paradox whereby the use of AI can increase individual creativity and efficiency, while decreasing the diversity of collective output. The upshot is that creativity and heterogeneity are highly important considerations in forecasting, developing, and adopting AI for science.