There’s no first-mover advantage in military AI – leadership’ is irrelevant

Much of the debate over military AI leadership and U.S. technological competition with China hinges on the assumption that there is a significant first-mover advantage when it comes to these technologies, meaning the first to develop them could reap substantial economic and military effects. However, fear of pronounced AI first-mover advantages instead reflects how AI is prone to “overhyping” where incredibly high expectations of capabilities surpass the reality of what is possible. Overhyping can obscure real progress and generate an inappropriate perception of an ​​​​​​AI arms race that “misrepresents the competition going on among countries.” Any possible first-mover advantage for AI would be unsustainable.

AI is a general-purpose, enabling technology not dissimilar to electricity. Moreover, the private sector drives its development, rather than the defense sector. While technologies that are singularly applicable to military contexts diffuse more slowly, those that are multi-use like AI have the added prodding of market incentives to speed up their spread.

AI Is Open Source and Available to All

Even when compared to other private sector-driven technologies, AI could spread even faster, since most AI research and development is open source with an unprecedented exchange of code and talent between tech companies and academia. This is a relatively new phenomenon in tech - “there is no business need to make closed infrastructure solutions, because within a few months everything will be totally different,” which has led to actors releasing even the most cutting-edge, proprietary AI. In 2015, Google opened up its sourcing framework TensorFlow. Facebook followed suit just a few years later with Caffe2 and PyTorch. OpenAI published GPT-2 in 2019, a large language processing model. The culture of keeping this work open-source and collaborative is widespread. A survey of AI and machine learning researchers showed that a majority believed that both a high-level and detailed description of methods, the results, and the actual algorithms should always be published, absent compelling risks from openness.