The perm adopts a “wait and see” approach to non deployment, but rollback is often impossible

(Annette Zimmermann is Lecturer in Philosophy at the University of York and Technology & Human Rights Fellow at the Carr Center for Human Rights Policy at the Harvard Kennedy School. 7-21)

A final obstacle to more robust ethical reflection on AI development is the presumption that we always have the option of non-deployment. If at some point in the future it turns out that an AI tool is having unacceptably bad consequences, some might say, we can simply decide to stop using the tool then. This may be true in some cases, but it is not clear why we should think it is always possible—especially without industry-wide regulation. The labor effects of automation, for example, may well be effectively irreversible. In current market conditions, it is hard to imagine how a company could take back its decision to replace a human-executed task with an AI-driven, automated process. Should the company face backlash over its AI tool, current incentives make it far likelier that it would seek to find another way to automate the task rather than rehire humans to execute it. The pressure to automate is now so strong in some sectors that some companies are pretending to have built and deployed AI. In 2016, for example, Bloomberg News reported that personal assistant startup X.ai was directing employees to simulate the work of AI chatbots, performing avalanches of mind-numbing, repetitive tasks such as generating auto-reply emails and scheduling appointments. It would be naïve to think that once such tools are actually built and deployed, the work force could easily revert to its pre-automated structure. For another example of the limits of non-deployment, consider DukeMTMC, a dataset of two million video frames recorded in public spaces on Duke University’s campus and made publicly available without protecting the identities of the people included in the videos. The data wound up being used for controversial research on computer vision-based surveillance technology, and it was taken down in June 2019 after significant public criticism. But as Princeton University researchers recently pointed out, at least 135 research papers utilized that dataset—and others derived from it—after it had been taken down. Non-deployment thus did not make the ethical and political risks associated with this technology disappear.