Nudging Algorithms by Influencing Human Behavior: Effects of Encouraging Fact-Checking on News Rankings

As society relies on algorithms to guide decisions from observations of humans, attempts to influence human behavior can also influence those algorithms. Here I report a field experiment that observes an “AI nudge,” an intervention influencing algorithm behavior by nudging human behavior. In an online community of 14 million, I test if encouraging readers to fact-check articles causes recommendation algorithms to interpret verification as popularity and promote those articles. Interventions encouraged readers to (a) fact-check articles or (b) fact-check and vote to influence a recommendation algorithm. While both encouragements increased fact-checking behavior, only the fact-checking condition reduced an article’s algorithmic ranking on average over time, contrary to expectations. Since AI nudges are possible, they have pragmatic and theoretical importance for understanding human and machine behavior.

MARCH 26, 2020: I am making available this pre-print during the COVID-19 pandemic to encourage people working on fact-checking to also consider second-order outcomes on ranking algorithms.

After completing the draft in 2018, I had set this dissertation paper aside while on the academic job market. I am planning to submit it for review and publication once things slow down.

随着社会依赖算法对人类观察结果进行决策，试图影响人类行为的行为也可能影响这些算法。在此，我报道了一项实地实验，观察了所谓的“AI微调”，即通过微调人类行为来影响算法行为的干预措施。在一个拥有1400万用户的在线社区中，我测试了鼓励读者进行事实核查是否会导致推荐算法将验证视为流行度并推广这些文章。干预措施鼓励读者（a）对文章进行事实核查或（b）进行事实核查并投票以影响推荐算法。尽管两种鼓励都增加了事实核查行为，但只有事实核查条件在平均时间上降低了文章的算法排名，这与预期相反。鉴于AI微调的可能性，它们在理解人类和机器行为方面具有重要的实践和理论意义。

2020年3月26日：我正在将这篇预印本在COVID-19大流行期间提供给人们，以鼓励从事事实核查的人们同时考虑对排名算法的二级影响。

在2018年完成草稿后，我在学术求职市场上将这篇博士论文草案搁置一旁。我计划在事情放缓后提交它进行评审和发表。