Synthetic Media Trust Evidence Record

Synthetic Media Trust Evidence Record

Governing claim

Synthetic media can create two different errors. A person can believe false media, and a person can reject authentic evidence as fabricated. Neither error justifies permanent cynicism. The useful response is consequence-based verification using several independent signals.

Primary studies

Vaccari and Chadwick's 2020 experiment found that a deceptive political deepfake contributed more to uncertainty than direct deception in the tested setting. The authors connected that uncertainty to lower trust in social-media news and reported that an educational treatment reduced uncertainty. The study does not establish a universal effect for every audience, medium, or synthetic video.

https://doi.org/10.1177/2056305120903408

Groh and colleagues compared individual people, aggregated human judgments, a leading model, and machine-informed judgments on videos sampled from the Deepfake Detection Challenge. Performance depended on the tested dataset and condition. The study supports caution about unaided visual judgment and model generalization, not the claim that every person or detector performs identically.

https://doi.org/10.1073/pnas.2110013119

Köbis and colleagues tested deepfake detection with and without a warning. In the warned condition, only 21.6 percent correctly selected the single deepfake without also rejecting at least one authentic video. That design illustrates the risk of false suspicion but should not be generalized beyond the sample and stimuli.

https://pmc.ncbi.nlm.nih.gov/articles/PMC10679876/

The 2025 study Labeling AI-generated media online found that the tested labels reduced belief in the claims shown. Label design, source, topic, platform, prior belief, and behavior remain important, so this does not establish that any label corrects every downstream effect.

https://pubmed.ncbi.nlm.nih.gov/40519990/

Technical and standards context

NIST AI 100-4 treats provenance, watermarking, labeling, detection, testing, and auditing as related interventions with different limits. It warns that false positives in content-based detection can cause serious reputational and other harms.

https://www.nist.gov/publications/reducing-risks-posed-synthetic-content-overview-technical-approaches-digital-content

C2PA provides signed provenance assertions. Its own principles reject a value judgment about whether provenance is good or bad. A consumer still decides what the assertions mean alongside other evidence.

https://c2pa.org/principles/

Publication boundary

Do not say video is no longer evidence, text is more truthful, viewers cannot detect synthetic media, or labels solve the trust problem.

Do not use the episode's examples as controlled experiments. The public essay should distinguish the cost to create media, the cost to verify a claim, the consequence of error, and the strength of available evidence.