In AI We Trust: Can AI Enhance Support for Online Moderation?

• In Event: Poster Session: Information Technology and Politics

Abstract

Political divides, populism and misinformation are on the rise. The nature of online discussions is often blamed for these democratic challenges, in that they tend to be politically like-minded, uncivil, and facilitate the spread of misinformation. Artificial intelligence (AI) is perceived as part of the problem, either through algorithms that expose users to like-minded content (Pariser, 2011) or through non-human automated bots that spread content at high speeds (Vosoughi, et al., 2018).

This project argues that AI has important democratic potential that remains un-explored in the area of political discussion. Across two pre-registered survey experiments in three countries that differ with regard to their attitudes toward AI (US, Spain, and Poland, total N = 6,000), we examine whether attributing (a) online content moderation, recommendation, and generation decisions as well as (b) politically pro- and counter-attitudinal statements to a human, AI, or human assisted by AI can influence (a) perception of the agent (in terms of fairness and political objectivity), (b) perception of the content moderating decision (in terms of its legitimacy and people’s trust), as well as people’s openness to political arguments (in terms of their agreement with the argument, its perceived quality, and perceived credibility and political objectivity of the agent posting the argument). We explore the tested effects across four different situations, in which AI is or could be used: (1) deleting hate speech and uncivil political posts, (2) reminding users to be civil and warning users that their behavior is inappropriate, (3) participating in online discussions to present political facts, and (4) making decisions about what news to show to readers based on information about what other content news users have read. Further, to establish that the effects on the reaction to political arguments apply to different issues, we test arguments on (a) a cold issue (social security), a (b) hot issue (immigration), and (c) an issue related to forecasting, the domain of AI.

For Experiment 1, respondents across the three countries were first randomly assigned to be in a human (33.33% of them), AI (33.333%), or AI-assisted human (33.333%) treatment conditions. For Experiment 2, respondents stayed in these “agent” conditions and were also randomly assigned to a pro- (50%) versus counter-attitudinal (50%) argument condition, in a 3 (statement source: human, AI, AI-assisted human) by 2 (statement leaning: pro, counter-attitudinal) between subjects design. They were told that they will read some political statements and that he comment was written/generated yesterday by “a person who / an AI that / a person assisted by an AI that tracks relevant political news” (depending on the condition). Depending on the condition, they saw either a liberal (e.g., pro-social security, pro-immigration, or contra-Trump economic policies in the US) or conservative (e.g., contra social security and immigration, and pro-Trump economic policies) statements of parallel length.

Our overarching expectation was that AI will have an advantage over human moderators and content generators in political online contexts. After all, humans are known to have political biases, to belong to political groups, and to be influenced by irrational factors in their decision-making. In contrast, people see algorithmically-driven search engines as fair, unbiased, and ideologically blind (Dutton et al. 2017; Fallows 2005; Sundar & Nass, 2001), and follow an advice from an algorithm to a greater extent than an advice from a human (Logg et al., 2019). Also, algorithms are programmed to follow the same procedures every time without being influenced by emotional factors and are perceived as lacking agency (Gray et al., 2007; Gray & Wegner, 2012; Waytz & Norton, 2014).

In experiment 1, we thus predicted that individuals will perceive an AI and an AI-assisted moderator as more fair and politically unbiased than a human moderator (H1) and see the decisions attributed to an AI and AI-assisted moderator as more trustworthy and legitimate (H2). Contrary to these expectations, we find that – on average and across all the political scenarios tested – human moderators were perceived as more fair and politically unbiased compared to AI moderators both for the US and Spanish samples (data from Poland have not yet been analyzed). In a similar vein, our respondents agreed more with the various moderation, recommendation, and content generation decisions and saw them as more trustworthy when these decisions were attributed to human moderators compared to AI moderators. The same pattern held for each individual scenario. We also inquired whether the agent and the decision will be perceived differently in the AI-assisted human treatment condition than in the AI treatment condition (RQ1). The data show that AI-assisted humans were perceived to be more fair and politically unbiased compared to AI.

Authors

• Magdalena Wojcieszak, UC Davis, U of Amsterdam

• Arti Thakur, University of California, Davis

• Andreu Casas, Vrije Universiteit Amsterdam

• Ericka Menchen-Trevino, American University

• Miriam Boon