Policymakers who believe that flagging information as recommended by artificial intelligence (AI) will curb the spread of fake news on social media might need to think again. A new paper from a research team including Ivy College of Business Assistant Professor Alec Dennis suggests that labeling social media stories as recommended by AI only encourages users to share information, whether fake or not.
In their study, “Artificial Intelligence Recommendations Amplify the Sharing of True and Fake News on Social Media by Appealing to Fast Cognition,” published in the Journal of Management Information Systems, the researchers used social influence theory to test how labeling stories as recommended by AI triggers user cognition.
Specifically, they explored whether social media users defaulted to quick and intuitive comprehension (System 1 cognition) or slow and rational comprehension (System 2 cognition) when encountering something recommended by AI. In other words, are users more discerning and less likely to share fake news when platforms are transparent about using AI?
There is a practical assumption that labeling news stories as recommended by AI will prompt social media users to devote more attention to understanding the information before sharing. This has led policymakers in the European Union and China to pursue laws requiring platforms to adopt AI content labels. This assumption is based on a concept known as algorithm aversion, which is the human tendency to distrust decisions made by algorithms. However, there is also the opposing concept of algorithm appreciation.
“We were curious if this would be a case of algorithmic aversion, as politicians were hoping it would be, or a case of appreciation where, rather than serving as a red flag and causing users to double check their sources, it would instead cause people to see AI as a seal of approval,” said Dennis, an assistant professor of information systems and business analytics.
“We could use an AI tool to actually identify when there is fake news and say, ‘Hey, an AI tool has verified this and this is not actually correct, so you should ignore this.’”
— Alec Dennis
Following three online experiments, the researchers concluded that users defaulted to fast, System 1 cognition and were therefore more likely to share fake news when it was recommended by AI.
“In practice, people don’t like thinking deeply about everything they come across,” said Dennis.
Whether this means users in the context of social media are “cognitively lazy,” as System 1 cognition has been described in previous literature, engaging in a subconscious labor-saving strategy, or simply unaware of AI’s limitations are potential topics for future research. For now, the implications of this study are clear: AI content labels will not thwart the spread of fake news and can actually increase it.
For his part, Dennis is now researching how AI can be used to fight fire with fire. “We are taking this leaning toward algorithmic appreciation and the fact that people do defer to AI to explore how we can use this to combat the problem rather than promote it,” he said. “We could use an AI tool to actually identify when there is fake news and say, ‘Hey, an AI tool has verified this and this is not actually correct, so you should ignore this.’”
