Openness causes unsafe AI development---extinction.

Is faster technological progress and rollout of AI capabilities desirable?

This brings us to the second question about the short and near-term desirability of openness: supposing openness would speed technical progress and rollout of AI capabilities, would that be socially beneficial?

It is clear that machine intelligence holds great promise for positive applications across many sectors of the economy and society, including transportation, healthcare, the environment, entertainment, security, and scientific discovery. For instance, an estimated 1.2 million people die every year in road accidents around the world, a number that could eventually be reduced to a low level as AI-enabled vehicles take over more functions from human drivers (Goldman Sachs, 2015). A report by McKinsey estimates an economic impact of several trillions of dollars annually from AI-related technologies by 2025.8 A full review of the potential positive applications is outside the scope of this paper.

As with any general-purpose technology, it is possible to identify concerns around particular applications. It has been argued, for example, that military applications of AI, including lethal autonomous weapons, might incite new arms races, or lower the threshold for nations to go to war, or give terrorists and assassins new tools for violence (Future of Life Institute, 2015). AI techniques could also be used to launch cyber attacks. Facial recognition, sentiment analysis, and data mining algorithms could be used to discriminate against disfavoured groups, or invade people's privacy, or enable oppressive regimes to more effectively target political dissidents (Balkin, 2008). Increased reliance on complex autonomous systems for many essential economic and infrastructural functions may create novel kinds of systemic accident risk or present vulnerabilities that could be exploited by hackers or cyber-warriors (See Perrow, 1984).

Insofar as it is possible to fine-tune openness choices so as to differentially expedite specific kinds of AI applications, these concerns might indicate the need for making exceptions to a generally pro-openness stance. For example, open-sourcing the code for autonomous weapons seems undesirable, and we have not heard anybody calling for that to be done. But basic research in AI is typically not application-specific in this way. Rather, to the extent that it succeeds, it will deliver algorithms and techniques that could be used in a very wide range of applications. This holds, in particular, for most work in current focal areas such as deep learning and reinforcement learning: that work is exciting precisely because it seeks general solutions to learning problems that occur in a wide range of tasks and environments.