HOUSE_OVERSIGHT_014716.jpg

Origins February 24 — 26, 2017 PROJECT An Origins Project Scientific Workshop ARIZONA STATE UNIVERSITY Challenges of Artificial Intelligence: Envisioning and Addressing Adverse Outcomes obesity, non-communicable disease; greater longevity; and lower rates of infant mortality — effectively stop any serious challenges to the systems being used and further developed. By 2050, life styles and healthcare across the US and many other parts of the world are governed by Al systems that have their roots in the early Al-consult technologies. The advice given to people, the actions that are imposed on them, the way people are persuaded and encouraged to live their lives in certain ways, are opaque, and are no longer under transparent direct human control. However, most people live longer, healthier and happier lives as a result. There remain several concerns: e There remains some differentiation in health and well-being related quality of life within society. Some communities and individuals opt out of Al-consult control, although their health-metrics are typically very poor in comparison with the rest of society. e Perhaps troublingly, there are some trends that are hard to make sense of. For instance, there seem to be fewer cases of mental and physical disability than might be expected. However, with Al-consult controlling healthcare (and health data) across the board, there are few ways for people to analyze and study these possible trends. e Lack of transparency can be a starting point for many adverse outcomes. e Autonomous devices rely on collecting personal data for performing their tasks. But what happens when a device starts to know more about its owner than the human itself? How do we ensure the device does not act in ways that would not act in ways that the owner would not want it to? (Of course the important question of making sure the data under consideration is protected and does not fall into malicious hands is a whole other discussion, but let us table that for now.) The classic story of the Target ads comes to mind, where a teenager was sent ads for pregnancy related products, however, she had not told her family about the pregnancy. e Systems might as above might move beyond such areas of health, and provide advice to people on both their daily decisions and longer-term planning. Such systems might evolve to become personal advocates who represent people to third parties. This would include both giving advice, and formulating arguments to make to others, or in making those arguments directly as your representative. These advocate bots will gradually be useful to a larger and larger fraction of the population, eventually being useful even as corporate legal counsel and as advisers to CEOs. Strong systems and reliance will raise reasonable alarms about Al control of people and society. How can we be sure that our these highly relied upon systems are genuinely advocating for us rather than the interests of others? DISCUSSION How can we characterize potential high-threat areas and stay aware of these possibilities even if these effects are insidious, and occur over long periods of time. What might be done to address potential poor outcomes? How can people maintain skills, agency, and be empowered, and aware over time with the expected growth and eventual ubiquity of Al systems that advise and guide? 20 HOUSE_OVERSIGHT_014716