AI Videos Shake Trust in Real Videos Even When Labeled, Study Finds

Researchers at MIT Media Lab and collaborators conducted a two-phase experimental study investigating how exposure to photorealistic AI-generated videos affects viewers' subsequent perception of authentic human-generated footage. One hundred participants were divided into two groups: one viewed AI-generated videos clearly disclosed as synthetic, while the control group watched human-generated videos. Both groups then transitioned to viewing human-generated content.

The results showed a clear divergence between the two groups' experience of the real videos. The AI-exposure group reported significantly increased doubt about the authenticity of the subsequent human-generated videos they watched. Beyond skepticism, they also experienced reduced confidence in their own judgments about what they were seeing, greater perceptual disruption—suggesting a disorientation in how they processed visual information—and notably lower feelings of social connectedness while viewing the real videos.

The persistence of these effects despite explicit labeling of the AI content is the study's key finding. Platform design and content moderation have historically focused on either improving detection of synthetic media or ensuring disclosure when AI-generated content appears. This research indicates that these approaches alone are insufficient: viewers can intellectually understand that they are watching synthetic content, yet still experience a degradation of trust and confidence when they subsequently encounter genuine material. The researchers conclude that addressing this challenge requires design strategies that prioritize what they term "perceptual safety"—protecting not only viewers' ability to distinguish real from fake, but also their psychological sense of security and connectedness when consuming visual media.

The study challenges a common assumption in content moderation and platform design: that clear labeling of synthetic media is sufficient to protect viewer perception. Prior research has emphasized detection and disclosure of AI-generated content, but this work reveals that the experiential act of consuming realistic synthetic videos creates a lingering doubt that carries over to genuine media. The two-phase design—exposing one group to AI video before both groups consumed real video—isolates the specific effect of prior exposure, showing it is not merely about deception but about how the plausibility of synthetic content reshapes confidence in visual authenticity itself.

The findings suggest that perceptual safety requires approaches beyond technical detection or text labels. The psychological impact appears to operate at a level deeper than conscious awareness of a label: participants knew what they were watching, yet their subsequent judgment of real videos was still affected. This distinction between intellectual knowledge ("this is synthetic") and felt experience ("I no longer trust what I see") points to a gap in current design strategies that treat the problem as primarily a detection or disclosure challenge.

No comments yet. Be the first to share your thoughts!