TLDR: A study by Hoq, Facciani, and Weninger reveals that AI-generated credibility scores significantly influence how users perceive political news on social media, often surpassing the impact of institutional labels and traditional engagement metrics like likes and shares. The research, involving 1,000 participants, found that AI feedback consistently moderated partisan bias and institutional distrust, suggesting a growing trust in algorithmic assessments and highlighting the need for transparent AI design to balance influence with user autonomy.
In today’s fast-paced digital world, social media has become a primary source of news for many. However, understanding what makes news credible on these platforms is a complex challenge. A recent study titled AI Credibility Signals Outrank Institutions and Engagement in Shaping News Perception on Social Media by Adnan Hoq, Matthew Facciani, and Tim Weninger from the University of Notre Dame sheds light on how different signals influence our perception of news accuracy, especially with the rise of AI-generated content.
The researchers conducted a large-scale experiment involving 1,000 participants to explore how AI-generated credibility scores impact how users view political news. They compared the influence of these AI scores against traditional signals like likes and shares, as well as institutional credibility markers.
The Study’s Approach
The experiment was designed to test four main hypotheses. First, the Ingroup Bias Hypothesis suggested that people would rate news headlines aligning with their political views as more accurate. Second, the Institutional Distrust Hypothesis proposed that AI-generated credibility assessments would be more persuasive than those from traditional institutional sources. Third, the Social Proof Hypothesis anticipated that high social engagement (likes, shares) would make headlines seem more credible. Finally, the AI Persuasion Hypothesis predicted that participants would adjust their credibility and sharing intentions based on AI-generated labels.
Participants were randomly assigned to one of four conditions: a control group with no credibility signal, a group receiving real-world media bias classifications from GroundNews, a group with reversed GroundNews labels, and a group receiving AI-generated assessments from ChatGPT. Each participant rated 21 political news headlines, which were also shown with varying levels of simulated social engagement (none, low, medium, or high).
Key Findings
The study yielded several significant insights:
First, political identity indeed plays a crucial role in how people judge news credibility. Participants consistently rated headlines aligning with their political affiliation as more accurate. Interestingly, this effect was strongest among moderates, while liberals and conservatives showed a preference for center-aligned content over both ingroup and outgroup headlines. When it came to sharing, only political moderates showed a significant inclination to share content that aligned with their political identity.
Second, both AI-generated feedback and institutional signals significantly influenced judgments of accuracy. However, their impact varied. While institutional cues from GroundNews were highly effective when aligned with a user’s ideology, they were often dismissed when misaligned. In stark contrast, ChatGPT feedback consistently boosted accuracy ratings across different political alignments, suggesting that AI assessments are perceived with higher neutrality. ChatGPT feedback also increased willingness to share headlines across the board, whereas institutional cues were less effective and sometimes even counterproductive for certain groups.
Third, and perhaps surprisingly, social engagement metrics like likes, shares, and comments had very little measurable effect on perceived accuracy or shareability. This finding challenges the common assumption that popularity metrics on social media strongly influence how credible news is perceived, especially when these cues are presented outside of their native platform environments.
Finally, the AI Persuasion Hypothesis was strongly supported. Participants adjusted both their credibility and sharing judgments in response to ChatGPT’s labels (e.g., accurate, inaccurate). This effect was particularly strong when the content aligned with their political identity, indicating that AI feedback can either amplify or dampen trust based on existing beliefs.
Also Read:
- Unveiling AI Personalities: A New Interface for Understanding Chatbot Behavior
- The Shifting Sands of AI Beliefs: How Context Changes Language Models
Implications for the Digital News Landscape
These findings have profound implications for how we design and implement credibility interventions in our increasingly AI-driven information ecosystems. AI-generated cues have the potential to mitigate bias and improve discernment, but their influence is not uniform and carries ethical risks, such as over-reliance. The study suggests that AI feedback, when framed with clear explanations, can be a more adaptable and inclusive tool for assessing credibility in politically diverse environments.
The research underscores the need for transparency, explainability, and careful framing when deploying AI credibility signals. Future efforts should focus on creating AI tools that are persuasive yet transparent, adaptable to diverse users, and designed to enhance, rather than override, human judgment and critical thinking.
