HOUSE_OVERSIGHT_014699.jpg

, Origins February 24 — 26, 2017 PROJECT An Origins Project Scientific Workshop ARIZONA STATE UNIVERSITY Challenges of Artificial Intelligence: Envisioning and Addressing Adverse Outcomes 2) DEMOCRACY, INFORMATION, AND IDENTITY Al, Information, and Democracy (Incorporating contributions from Shahar Avin, Sean O hEigeartaigh, David McAllester, Eric Horvitz, and others) An informed public is important to the healthy functioning of democratic societies. We can expect potential forthcoming advances around the control of information feeds with applications in spreading propaganda, via spreading false or misleading information, creating anxiety, fueling conspiracy theories, and influencing voting. Such methods will bring key challenges to democracy. CHALLENGES AHEAD WITH AI, PROPOGANDA, AND PERSUASION Data-centric analyses have been long used in marketing, advertising, and campaigning over decades. However, over the past few years, we have seen the rise of the use of more powerful tools, including machine learning and inference aimed at algorithmic manipulation, with the target of influencing the thinking and actions of people. Some initial uses of these methods reportedly played a role in influencing the outcome of recent US presidential elections, as well as the elections in 2008 and 2012. We can expect to see an upswing in methods that manipulate states of information in a personalized automated manner. These systems can be designed and deployed as omnipresent/persistent, and aimed at specific goals for group- or person-centric persuasion. As our data and models of how people consume and act on information improve, and as an increasing portion of information consumption is mediated through digital systems managed by potentially opaque algorithms, it becomes increasingly conceivable that the information ecosystem would get captured by malicious actors deploying increasingly advanced tools to control, shape, forge and personalize information, from ads to news reports. Machine learning, in conjunction with active learning, expected value decision making, and optimization of allocations of key resources, such as dollars or human effort, can be targeted at monitoring, understanding, and then working to influence the beliefs and actions of large populations of people. Data can be collected from large-scale populations, across multiple devices and services, and used to make inferences about the psychologies and beliefs of people, and for designing and guiding persuasive flows of sequences of information. Uses of Al can include attempts to optimize stealthiness of the interventions. In the future, a great deal of the information consumed by citizens on personal devices is subject to alteration by information-engineers at media corporations and governmental propaganda offices, such that outside a few key positions of power no one really knows what is going on in the world. There is a danger of the growth of domination over time of large populations by a single dominant or a few systems. We can imagine methods that modify even such feeds as Wikipedia articles, creating personalized views—that subtly shift the version of the article seen by my HOUSE_OVERSIGHT_014699