Can people reliably spot an AI-generated face just by looking?

The key study is Nightingale and Farid, "AI-synthesized faces are indistinguishable from real faces and more trustworthy", published in PNAS in February 2022. Across three experiments, participants classified real and synthetic faces with accuracy close to 50%, the level you would get by guessing. A group that was trained on common artefacts and given feedback did only modestly better. In a separate experiment, participants judged the synthetic faces as more trustworthy than the real ones. A follow-up by Miller and colleagues, "AI Hyperrealism: Why AI Faces Are Perceived as More Real Than Human Ones", published in Psychological Science in 2023, went further: in their first experiment with 124 adults, White AI faces were judged as human more often than actual human faces, and the people who made the most errors were the most confident.

Key figures from Nightingale and Farid, 2022

Experiment	Participants	Result
Real or synthetic, no training	315 people	Accuracy close to 50%, the level of guessing
Real or synthetic, with training and feedback	219 people	Only a modest improvement over guessing
Trustworthiness ratings	223 people	Synthetic faces rated slightly more trustworthy than real faces

Why the old tips stop working

Advice like "check the ears", "look at the teeth" or "watch for mismatched earrings" came from earlier generators with obvious flaws. Each new generation of models fixes the most visible artefacts, so a checklist that worked two years ago quietly stops working. Worse, it gives people false confidence when an image passes the checklist.

What works better

• Provenance. Does the image carry Content Credentials, or come from a source with a verifiable publishing history?
• Context. Who first posted it, when, and does the account have a history that makes sense?
• Reverse image search. Real photographs of real people usually exist in more than one place.
• Forensic tools as a second opinion. Detection tools can add evidence, but treat their output as probabilistic, like text detectors.
• Understand the provenance route. The forum's thread on C2PA and Content Credentials explains how signed capture flips the question from detection to verification.