AI Generated Faces: How to Spot a Person Who Does Not Exist
In a controlled study, people could not tell AI generated faces from real ones, and they rated the synthetic faces as more trustworthy. That is the whole problem in one sentence. Here is what AI generated faces are, how they are made, where they turn up, and how to catch one.
AI generated faces are photorealistic portraits of people who do not exist, produced by models such as NVIDIA's StyleGAN or modern diffusion systems like Midjourney and Stable Diffusion. You can spot many of them by checking eye alignment, earrings, hair edges and backgrounds, and you can verify any face by scanning it with an AI image detector.
- What are AI generated faces?
- How are AI generated faces made?
- Where do fake faces show up?
- How do you spot an AI generated face?
- Why do AI faces look trustworthy?
- How do you verify whether a face is real?
- FAQ
- Conclusion
What are AI generated faces?
An AI generated face is a portrait invented by a machine learning model. The model blends patterns learned from millions of real photographs into a new face with realistic skin, eyes and hair. Nobody posed for the shot, because there is no person.
The phenomenon went mainstream in 2019 with thispersondoesnotexist.com, a site that loads a new synthetic face on every refresh. It ran on StyleGAN, a generator architecture published by NVIDIA researchers in December 2018 (Karras, Laine and Aila, arXiv 1812.04948). The site made one point and made it well: a free webpage can produce an endless supply of convincing strangers.
The technology has since shifted. The crisp, occasionally glitchy StyleGAN look has given way to diffusion-era faces from the same family as Midjourney, DALL-E 3 and Stable Diffusion. These fix many of the old giveaways, which is why a detection-first approach now matters more than "just look closely".
How are AI generated faces made?
You do not need to build a generator to catch one, so this stays conceptual. There are two broad approaches.
A GAN, or generative adversarial network, pits two models against each other. One paints faces. The other tries to tell fakes from real photos. They push each other until the painter produces images the critic can no longer flag. StyleGAN works this way, sampling points in a mathematical latent space where each point maps to a different face.
A diffusion model, the dominant approach today, starts with pure noise and removes it step by step until a coherent face emerges, guided by what it learned about how faces look. Diffusion is the engine behind most of the most realistic synthetic people now in circulation.
Both methods produce a person from statistics, not from a camera pointed at a human. That single fact is the root of every tell below. For the broader mechanics, see GAN vs diffusion deepfakes.
Where do fake faces show up?
A face that belongs to no real person and cannot be reverse-searched is a gift to anyone running a scam or an influence operation.
Catfish and fake profiles
Romance scammers prefer AI faces because a stolen photo can be reverse-searched back to its owner and a generated face cannot. If you are vetting a profile photo, our guide to AI generated profile pictures walks through the workflow.
Disinformation and bot networks
State-linked influence operations use synthetic faces routinely. Graphika has documented coordinated networks in which dozens of accounts used GAN-generated profile pictures to pose as local residents (Graphika reports). A tactic that was a novelty in 2019 now appears in most operations investigators analyse.
Synthetic identity fraud
Fraudsters pair a generated portrait with stolen or invented details to pass remote identity checks and open accounts. This is a pressure point for any business running remote onboarding. See KYC and identity verification for the onboarding angle.
Fake reviews, testimonials and team pages
Synthetic headshots populate fake testimonial sliders and the "our team" pages of shell companies. A roster of friendly faces builds instant trust and none of them has to exist. A polished, perfectly diverse staff with no traceable LinkedIn presence deserves a second look.
How do you spot an AI generated face?
No single tell is proof, and the newer the model, the more of these it fixes. Treat them as a flag count. The era tags show which giveaways are fading.
| Tell | What to look for | Era | Reliability in 2026 |
|---|---|---|---|
| Identical eye positioning | Pupils at the same coordinates across different faces. Overlay two and they line up | GAN | Fading |
| Earring and glasses asymmetry | One earring differs from the other; frames change thickness across the bridge | GAN and diffusion | Still reliable |
| Hair melting into skin | Strands dissolve into the forehead or background instead of ending | GAN and diffusion | Still reliable |
| Background soup | Sharp subject, background warps into half-objects | GAN and diffusion | Still reliable |
| Collar and clothing chaos | Necklines and seams mismatch side to side | GAN | Fading |
| Teeth and ear irregularities | Too many teeth, fused teeth, misshapen ears | GAN and diffusion | Still reliable |
| Too-perfect symmetry and lighting | Flawless symmetry, poreless skin, even glossy light on a candid-looking shot | Diffusion | Supporting only |
Two or more flags on a single face is a strong signal, especially earrings, hairline and background together. For the general-purpose list across all image types, see how to tell if an image is AI generated.
Why do AI faces look trustworthy?
Here is the finding that should change how you treat any face online. In a 2022 PNAS study, Sophie Nightingale and Hany Farid found that participants identified AI-synthesised faces with 48.2% accuracy, below chance. In a second experiment, participants rated the synthetic faces as more trustworthy than real ones (Nightingale and Farid, PNAS, 2022).
The likely reason is that generators converge on average, symmetrical features, and human brains read averageness as trustworthy. Your instinct is not merely unreliable here. It is biased toward believing the fake. "It looks real to me" is not a verification method.
How do you verify whether a face is real?
Reverse image search, and expect it to fail
A stolen real photo often surfaces its true owner. A purely generated face returns nothing, because it exists nowhere else. A blank result is suspicious, not reassuring.
Count the tells
Walk the table above. Two or more flags is a strong signal.
Scan it with an AI image detector
Upload the image to the AI image detector. The pixel model returns a whole-image verdict and a 0 to 100 score. If you have the original file, the source check reads EXIF, XMP and C2PA metadata and can name the generator with an evidence level. A face saved from a profile page is usually stripped of metadata, so expect none there and rely on the score.
The detector accepts JPG, PNG, WEBP, BMP, TIFF, HEIC and HEIF up to 12 MB. Free accounts include 50 detections per month, no card required. For how the two layers work, read how AI image detectors work. To compare tools before you pick one, see the AI image detectors that tested best.
FAQ
Are AI generated faces real people?
No. A generated face is a statistical blend of features learned from many real photographs, so it belongs to nobody. By chance it can resemble a real person, and researchers have documented near-matches, but there is no individual behind the image. Any account presenting it as a real person is misrepresenting itself.
Is thispersondoesnotexist still online?
The site was still serving StyleGAN faces at the time of writing (checked 20 October 2026). It remains the simplest demonstration of the phenomenon: refresh the page and a new person who never existed appears. The faces it produces carry the classic GAN-era tells, including the centred eyes.
Can AI generated faces be detected?
Yes. Detectors do well on face imagery because faces are a narrow, heavily studied domain. GAN faces leave strong, well-characterised fingerprints. Diffusion faces are harder but still detectable from pixel statistics, and when the original file survives, metadata and C2PA records can name the generator outright.
Is it legal to use an AI generated face?
Using a generated face is legal in itself. It becomes illegal when used for fraud, impersonation of a real person, deceptive advertising or identity crimes, and most platforms ban undisclosed synthetic profile photos under their terms. See deepfake laws by state for the US picture.
Why do AI faces look so trustworthy?
Because generators average. Averaged, symmetrical faces read as trustworthy to human perception, a bias measured directly by Nightingale and Farid in 2022, whose participants rated synthetic faces as more trustworthy than real ones. The effect is a feature of how our brains judge faces, not something you can train away quickly.
Conclusion: assume nothing, verify faces
AI generated faces are people who do not exist, made by StyleGAN-style GANs and now by diffusion models, and they are good enough to beat human judgement and even earn more trust than real faces. The counter is simple: reverse-search, count the tells, and scan the image with a detector. Treat any face you cannot verify as unverified, not as real.
Check a face now. Free accounts include 50 detections per month.