Deep learning diffusion by infusion into preexisting technologies—implications for users and society at large

abstract Artificial Intelligence (AI) in the form of Deep Learning (DL) technology has diffused in the consumer domain in a unique way as compared to previous general-purpose technologies. DL has often spread by infusion, i.e., by being added to preexisting technologies. We find that DL-algorithms for recommendations or ranking have been infused into all the 15 most popular mobile applications (apps) in the U.S. (as of May 2019). For example, when a DL-system was infused into YouTube, it almost immediately reached a third of the world's population. We argue that existing theories of innovation diffusion and adoption have limited relevance for DL-infusion, because it is a process that is driven by enterprises rather than individuals. We also discuss its social and ethical implications. First, consumers have a limited ability to detect and evaluate an infused technology. DL-infusion may thus help to explain why AI's presence in society has not been challenged by many. Second, conflicts of interest are inevitable, since the goals of the DL-supplier and its users are not always aligned. Third, as compared to other innovations, infusion is likely to be a particularly important process for DL-applications, because they need large data sets to function well, which can be drawn from preexisting users. Related, DL-infusion seems to comparatively benefit large technology companies that already have many users. This suggests that the value drawn from DL is likely to follow a Matthew Effect of accumulated advantage: many preexisting users bring about better DL-driven features, which attracts more users, which results in better DL-features, etc. This self-reinforcing process could limit the possibilities for new companies to compete. This way, the notion of DL-infusion may put light on the power shift that comes with the presence of AI in society.