Can you actually trust AI? (And just how much is your favorite influencer using it?)

ChatGPT and other generative artificial intelligence (AI) platforms make creating content, images, and even videos easier than ever. But what was once a platform for improving writing or content creation has ballooned into a decision-making engine. A recent study conducted by OpenAI (ChatGPT’s parent company) found that most people use ChatGPT for not only writing but also for practical guidance and seeking information. Despite AI platforms’ value built on synthesizing vast amounts of information, the question becomes: Can you trust the advice ChatGPT and other AI platforms provide?

This question about the validity of AI-generated information, and larger questions about misinformation spread on social media, is what we sought to understand in our research sponsored by the Page Center. Specifically, we looked at social media influencer (SMI) distribution of misinformation as well as their use of AI. In this study, we partnered with Mario Nicolini, senior public policy expert at Ambrela; Beaudine Verhoek, strategic communications officer at North Atlantic Treaty Organization (NATO); and Angela Dwyer, head of insights at Fullintel.

First, we explored AI use among SMI posts in a context that would be chock-full of misinformation – the 2024 U.S. Presidential campaigns. Specifically, we chose two highly public events thinking they would be ripe for misinformation spread – the attempted assassination of Donald Trump and the Kamala Harris’ announcement of her presidential candidacy. We collected data during the height of both events (July – August, 2024). We tracked proven points of misinformation in SMI posts on X including, in the Trump shooting, that the shooting was staged, images were doctored, and extra security was declined. In Harris’ candidacy, there were false claims about her ethnic heritage, an extra-marital affair and that Biden's death had led to her candidacy.

Results were unexpected. Despite our assumption that misinformation would be rampant, it was actually quite low. Only 16% of influencer content we analyzed featured misinformation. The most misinformation was found in Trump’s shooting (13.5%). The percentage of misinformation in posts about Kamala was almost nil (3.3%). Evidence of AI content was even lower at 1.25%. We did, however, find a potentially significant trend when examining AI posts for misinformation. 14% of AI content featured misinformation in our study. So, influencer use of AI and spread of misinformation may have been minimal during the U.S. Presidential campaigns, at least across the topics and influencers we studied.

Of course, one case isn’t enough to prove how often influencers use AI or share faulty information. So, we connected with influencers discussing political topics — with the expectation that politics features a high level of potential misinformation. In-depth interviews with ten influencers gave us rationale regarding AI use and why or why not they share misinformation. Most told us they don’t use AI because it threatens their credibility and rarely reflects their voice. Misinformation was avoided for the same reason.

Many said that they fact-checked their information to maintain trust with their followers. However, influencers rarely correct misinformation they find. Even though most were highly involved in promoting factual information, they rarely saw it as their responsibility to correct other’s mistakes by addressing faulty information through their own social media activity. Representation was also an issue for influencers in this study — they were also less-inclined to want to promote influencers spreading misinformation by correcting the faulty information in their own posts.

So, despite the potential of influencers to use AI and spread misinformation, we found little evidence that this was the case. In future studies, we plan to expand our reach on this issue to track AI-generated misinformation more broadly on social media.

For more information about this study, email Brian Smith at bgsmithphd@gmail.com. This project was supported by a 2024 Page/Johnson Legacy Scholar Grant from the Arthur W. Page Center.