.webp)
The short answer is unsettling: to the human eye, barely at all. In controlled studies, people asked to tell AI-generated faces from real ones perform at about the level of a coin flip, and, more strikingly, they tend to rate the synthetic faces as more trustworthy than the real ones. The faces produced by today's generators have not just closed the gap with reality; in some respects they have overshot it.
This piece answers how detectable AI-generated faces really are, for people and for machines. It covers what the research shows about human perception, why synthetic faces are so convincing, which visual tells remain and how few of them survive in modern models, how well automated detection does by comparison, and what that means as fabricated faces spread through fraud, fake profiles, and disinformation.
The uncomfortable theme throughout is a mismatch between confidence and ability. People believe they can spot a fake face, and the research says most of them cannot.
- To the human eye, AI-generated faces are barely detectable; in studies, people perform at or near chance.
- Synthetic faces are often judged more trustworthy than real faces, a phenomenon called AI hyperrealism.
- This happens because AI faces tend to look more average, familiar, and attractive, which reads as normal and trustworthy.
- A July 2026 study found people detected diffusion-generated faces only slightly above chance, at about 58%, and rated them the most trustworthy of all.
- The old visual tells, like odd eyes, warped backgrounds, and mismatched earrings, have largely been fixed in modern generators.
- Machines detect AI faces far better than people, often above 95% in the lab, but accuracy drops on unseen models and compressed images.
- The people most confident they can spot AI faces are frequently the ones most fooled.
- Because human judgment is unreliable, catching synthetic faces at scale, in fraud, fake profiles, and disinformation, requires automated detection.
The Short Answer: The Human Eye Fails
The landmark finding came from a 2022 study by Sophie Nightingale and Hany Farid, which concluded that AI-synthesized faces are indistinguishable from real faces and, remarkably, are judged more trustworthy. Participants distinguishing GAN-generated faces from real ones scored around 50%, no better than guessing, and the researchers described synthesis engines as having passed through the uncanny valley entirely. Follow-up work labeled the deeper effect AI hyperrealism: not merely that people cannot tell, but that synthetic faces are often perceived as more real than genuine ones.
The pattern has held as the technology advanced. A July 2026 study, the first to test faces made with the latest diffusion models, found that people detected AI faces only slightly above chance, at about 58%, and that both GAN and diffusion faces were rated more trustworthy than real faces, with diffusion faces rated the most trustworthy of all despite being judged slightly less realistic. In other words, realism and trust are separate judgments, and synthetic faces win on both in ways that favor the deceiver. Individual ability varies a lot, with accuracy in some studies ranging from below chance to around 80%, but the central conclusion is consistent: unaided human perception is not a reliable defense.
Why AI Faces Fool Us
The reason is counterintuitive. Synthetic faces fool people not because they are exotic but because they are, in a sense, too ordinary. Research attributes hyperrealism to the fact that AI faces tend to be more average, more familiar-looking, more attractive, and less memorable than real human faces. A face generated to sit near the statistical center of its training data looks like everyone and no one, and because the human brain treats an average face as the norm, it reads as both real and trustworthy. We assess a face in as little as 100 milliseconds, and that fast, automatic judgment is exactly what an average synthetic face is well suited to pass.
This produces a dangerous confidence gap. Studies have found that the people most likely to be fooled by AI faces are the least likely to realize it, and that people generally miscalibrate their confidence and misread the cues they think they are using. There is one hopeful wrinkle: recent research suggests that training people to attend to their overall impression of a face, rather than to hunt for specific artifacts, can nearly double accuracy. But left to instinct and a search for glitches, most people fail.
The Tells That Sometimes Remain
For a few years there were reliable giveaways, especially in faces from earlier GAN models like the ones behind the "this person does not exist" demonstrations. Because those generators aligned faces in a fixed layout, the eyes tended to sit in the same position in every image. Backgrounds were often warped or surreal, since the model focused on the face. Accessories betrayed the fake, with mismatched earrings or asymmetric glasses, and the reflections in the two eyes frequently failed to match, since each was rendered somewhat independently. Teeth and hair, being fine and irregular, blurred or merged.
The problem is that these tells are fading fast, and they were never reliable for a casual viewer even at their best. Modern diffusion-based generators produce clean, coherent backgrounds, consistent accessories, and plausible eyes, erasing most of the cues that guides used to list. Relying on a checklist of visual artifacts is now a losing strategy, because the checklist describes yesterday's generators, and any tell that becomes well known is a flaw the next model is trained to fix.
Machines Detect Better, But Not Perfectly
Automated detection substantially outperforms human perception, but it is not a solved problem. Detectors can read signals the eye cannot, most importantly the frequency-domain fingerprints that generators leave in the statistical structure of an image, and on known generators they routinely exceed 95% accuracy in laboratory tests. That is the good news, and it is why serious defenses lean on software rather than human reviewers.
The caveats are the same ones that apply across deepfake detection. Accuracy drops sharply on generators the detector was not trained on, the generalization gap that lets a brand-new model slip through. Diffusion faces are intrinsically harder to detect than older GAN faces because their statistics sit closer to real images. Compression and re-encoding, which every social platform applies, degrade the subtle signals detectors depend on. And detection models can be opaque and vulnerable to deliberate manipulation. Notably, humans and machines tend to make different errors, which is why the strongest results often come from combining machine detection with human review rather than relying on either alone.
Why It Matters: Synthetic Faces in the Wild
This is not an academic curiosity, because a fabricated face that no one can flag is a near-perfect tool for deception. AI faces now anchor fake social-media and professional profiles, including the bot networks that have populated platforms with plausible but non-existent people. They are the photo behind romance scams and catfishing, the selfie attached to a synthetic identity applying for credit or an account, and the fabricated spokesperson in a disinformation campaign. Because the faces are generated on demand and belong to no real person, there is no original photo to trace and no victim whose likeness was stolen to raise the alarm.
The trust finding makes this worse. If synthetic faces are not merely undetectable but actively perceived as more trustworthy, then a scammer using an AI face may start with an advantage over one using a real photo. As the researchers behind the 2026 diffusion study warned, these tools have democratized the ability to create convincing fake faces for anyone, with no technical skill required. For a concrete example of how this feeds a larger fraud, see our guide to how synthetic identities are generated.
How to Actually Detect an AI Face
The practical guidance follows directly from the evidence. First, do not trust your instinct, because the research is clear that a confident gut feeling is not correlated with being right, and a face that feels especially trustworthy deserves more scrutiny, not less. Second, do not rely on a checklist of visual glitches, since modern generators have fixed most of them. Reverse image search can occasionally help by revealing that a photo has been reused, but a freshly generated face has no earlier source to find. Provenance markers such as C2PA credentials confirm origin when they are present, but they are opt-in and easily stripped.
That leaves automated detection as the only method that scales and that reads the signals humans cannot. The realistic posture is to treat any consequential face, in onboarding, in a new contact, in a viral image, as unverified until checked by a tool built for the purpose. This is where DuckDuckGoose, based in Delft, focuses: its DeepDetector analyzes images for the statistical signatures of synthetic media, flagging AI-generated faces that people, by the evidence, simply cannot reliably catch. Mordor Intelligence names DuckDuckGoose AI among the major companies in the fake image detection market.
Frequently Asked Questions
Can people tell AI-generated faces from real ones?
Generally, no. In controlled studies, people distinguish AI faces from real ones at about the level of chance, and a 2026 study on diffusion-generated faces found detection only slightly above chance at around 58%. Ability varies between individuals, but unaided human perception is not a reliable way to catch a synthetic face.
Why are AI faces judged as more trustworthy than real faces?
Because AI faces tend to be more average, familiar, and attractive than real ones, and the brain treats an average face as the norm, which reads as both real and trustworthy. This effect, called AI hyperrealism, means a synthetic face can seem more genuine than a real photograph, which is precisely what makes it useful for fraud.
What are the visual signs of an AI-generated face?
Older GAN faces sometimes showed mismatched eye reflections, warped backgrounds, asymmetric earrings or glasses, and blurred teeth or hair, along with eyes fixed in the same position across images. These tells have largely been eliminated in modern diffusion generators, so a checklist of visual glitches is no longer a dependable method.
Are diffusion-generated faces harder to detect than GAN faces?
Yes, for both people and machines. Diffusion models produce output that sits statistically closer to real images and lacks the grid-like frequency fingerprints of GANs. The 2026 study found diffusion faces were rated the most trustworthy of all, and automated detectors trained on older models lose accuracy on diffusion output.
Can software detect AI-generated faces reliably?
Better than people, but not perfectly. On generators it has been trained on, automated detection often exceeds 95% accuracy in the lab by reading frequency-domain signals invisible to the eye. Accuracy falls on unseen generators and on compressed images, which is why reliable systems combine multiple signals and are updated as new generators appear.
Where are AI-generated faces used maliciously?
They anchor fake social and professional profiles and bot networks, serve as the photos in romance scams and catfishing, act as the selfie behind synthetic identities in fraud, and provide fabricated spokespeople for disinformation. Because the face belongs to no real person, there is no original to trace and no victim to report the misuse.
How can I check whether a profile photo is an AI face?
Do not rely on your instinct or on spotting glitches, both of which are unreliable now. Reverse image search occasionally helps but fails for freshly generated faces, and provenance markers help only when present. The dependable approach is an automated detection tool that analyzes the image for the statistical signatures of synthetic media.
About the author
