Researchers used a generative adversarial network to produce images of synthesized faces that survey participants found more trustworthy than real human faces
and more trustworthy—than real faces.” https://singularityhub.com/...
Context & Ripple Effects
This finding closes a loop that opened with Nvidia's 2018 work generating convincing-but-fake faces from celebrity photo patterns: within four years, the same GAN lineage produced faces that survey participants not only accepted as real but rated more trustworthy than actual humans. A later 2023 study confirming AI-generated white faces were judged more real than photographs suggests the effect was durable, not a one-off artifact of one experiment.
First-order effects
- Human visual judgment fails as an authenticity check: participants actively preferred synthetic faces on trust, so any platform or process relying on 'does this person look genuine' is screening with a broken instrument.
- The named players here are researchers and survey participants, but the immediate exposure falls on anyone whose identity is verified by face photos — profile pictures, testimonials, dating and marketplace profiles all become cheap to synthesize convincingly.
Second-order effects
- Accessible tooling compounds the problem: [[a:985751|Stable Diffusion and DreamBooth make life-wrecking deepfakes from a handful of social-media photos]], meaning the trust advantage demonstrated in the lab is already deployable by ordinary users, forcing platforms toward provenance and detection systems rather than user vigilance.
- Trustworthy generation concentrates upstream: per the related coverage, building credible generative AI takes resources on the scale of Microsoft's and Google's, so the parties who can both create hyper-trustworthy faces and sell the verification for them gain outsized power over the trust layer itself.
Third-order effects
- If synthetic faces reliably outscore real ones on perceived trustworthiness, the industry's center of gravity shifts from perceptual authentication (looking real) to cryptographic provenance (being verifiably sourced) — content credentials and watermarking infrastructure become the load-bearing trust mechanism.
- Regulators and platforms face a likeness-governance problem distinct from fraud detection: when fakes outperform originals, consent and attribution rules around synthetic likenesses matter more than the real/fake boundary itself.
The trend: Synthetic faces have crossed from imitation to superiority in perceived trustworthiness, pushing digital identity verification from what images look like to where they provably came from.