Deepfakes Fool Voters — But Not More Than Headlines or Audio

📰 What Was Tested

A large, representative sample (N = 5,750) of American respondents was exposed to fabricated videos of public officials synthesized by deep learning (“deepfakes”) and to comparable misinformation in other formats (text headlines, audio). Two experiments used a novel collection of deepfakes created in collaboration with tech-industry partners to compare credibility judgments, affective reactions, and detection accuracy across media types.

🔬 How Exposure and Detection Were Measured

• Randomized exposure to deepfake videos, authentic videos, text headlines, and audio recordings.
• Credibility and affective response ratings collected after exposure.
• A forced-choice detection task asking viewers to identify real versus fabricated videos.
• Experimental manipulations included brief informational messages and accuracy primes.
• Covariates included partisanship, political literacy, and digital-technology literacy.

📊 Key Findings

• Up to 50% of the representative sample judged deepfakes to be credible.
• Deepfakes were no more credible than equivalent misinformation presented as text headlines or audio recordings.
• No meaningful subgroup heterogeneities in credibility perceptions or larger affective responses for deepfakes relative to other media.
• Partisanship strongly predicts a large gap in detection accuracy for real videos, but not for deepfakes: partisan differences emerge when assessing authentic footage, not fabricated footage.
• Brief informational messages or accuracy primes only sometimes—and only modestly—reduce the perceived credibility or impact of deepfakes.
• Higher levels of political literacy and digital-technology literacy consistently improve the ability to distinguish deepfakes from authentic videos of political elites.

💡 Why It Matters

Deepfakes are widely believable to many viewers, but they are not uniquely persuasive compared with other forms of misinformation. The strongest mitigators of mistaken belief are broader political and digital literacy rather than short informational prompts, and partisan bias shapes detection of authentic political video more than it shapes detection of fabricated video. These results come from two randomized experiments using industry-produced deepfakes and have direct implications for misinformation policy and media-literacy interventions.