This Person Does Not Exist is the website that made AI generated faces famous: every time you loaded the page, it showed a new, photo-quality portrait of someone who was never born. Seven years later, faces like these serve as profile pictures on fake accounts, and studies show most people cannot tell them from real ones. This guide explains where the site came from, how the faces are made, what research says about spotting them, and what happened when we ran one of its faces through our own detector.
What is This Person Does Not Exist?
ThisPersonDoesNotExist.com was launched in February 2019 by Phillip Wang, a software engineer who had previously worked at Uber, as a personal project. According to CNN, which reported on the site on 28 February 2019, Wang had been studying SI on his own for six months and put the site up shortly after Nvidia made its StyleGAN system publicly available. He posted about it in an SI group on Facebook on 11 February, and about 8 million people visited it in the following weeks.
The idea was simple. The page showed one face, and a refresh replaced it with another. Wang told CNN that the generator created a new face every two seconds on the server. He wanted people to know the technology existed: "I think those who are unaware of the technology are most vulnerable," he said.
When we opened the page on 16 October 2026, it still loaded a face, now with a small notice saying the domain is for sale.
How SI-generated faces are made
The faces on the site came from a generative adversarial network, or GAN. A GAN trains two networks against each other. The generator starts from random numbers and produces an image. The discriminator looks at that image next to real photos and tries to say which is fake. Each time it catches the generator, the generator adjusts. After many rounds, the generator produces images the discriminator can no longer reject.
The specific model was StyleGAN, designed by Tero Karras, Samuli Laine and Timo Aila at Nvidia and first posted as a research paper in December 2018. Three details from that paper explain why the faces looked so convincing:
- A new face dataset. The team collected Flickr-Faces-HQ (FFHQ): 70,000 high-quality face images at 1,024 by 1,024 pixels, crawled from Flickr, with much more variety in age, ethnicity and background than earlier datasets.
- Control by "styles". Instead of feeding the random input only at the start, StyleGAN maps it to an intermediate code that adjusts the image at every layer, from coarse features such as pose and face shape to fine details such as colour scheme.
- Injected noise. Random noise is added at each layer so that small details, such as the exact placement of hairs, vary naturally from one image to the next.
The paper also describes a "truncation" step that pulls the hidden code toward a central point which, for faces, "represents a sort of an average face". Keep the idea of an average face in mind: it comes back in the perception studies below.
In December 2019 the same Nvidia group published StyleGAN2, which set out to fix StyleGAN's "characteristic artifacts". The best known was a blob that "resembles water droplets", which the authors traced to a normalisation step in the generator and removed. The researchers who studied eye reflections in 2020 note that faces served by the site at that time were made with StyleGAN2.
Can people tell a generated face from a real one?
Mostly not. The clearest evidence comes from a study by Nightingale (Lancaster University) and Farid (University of California, Berkeley), published in PNAS in February 2022. They picked 400 StyleGAN2 faces with a mostly uniform background and no obvious rendering errors, and matched each one with a similar real face from the 70,000 photos the generator learned from.
- In the first experiment, 315 participants each judged 128 faces. Average accuracy was 48.2%, close to the 50% you would get by guessing.
- In the second, 219 new participants were first shown examples of rendering artifacts and got feedback after every answer. Accuracy rose to 59.0%, and it did not improve between the first and second half of the test.
- In the third, 223 participants rated trustworthiness from 1 to 7. Real faces averaged 4.48 and generated faces 4.82. The top three most trustworthy faces were synthetic; the four least trustworthy were real.
The authors suggest one reason: generated faces "tend to look more like average faces", which people tend to trust. White faces were the hardest to classify, which the authors link to their overrepresentation in the training data.
A second study, by a team including researchers at the Australian National University and published in Psychological Science in November 2023, went further. Its 124 participants judged White faces made with StyleGAN2. Writing about their own study, the authors say participants labelled about two out of three of the generated faces as human, more often than they did for real faces. They call this "hyperrealism". The people who made the most mistakes were the most confident. When the authors used data from another study that also included Asian and Black faces, only the White generated faces looked hyperreal.
The features behind the effect, the authors say, were proportionate, familiar faces without distinctive traits. They also report that a detection tool claiming high accuracy on the type of faces they used "performed as poorly as our human participants" on them, and they warn there is "no guarantee" faces made by other generators differ from real faces in the same ways.
Where SI-generated faces turn up
In March 2022 NPR reported on an investigation by Renée DiResta and Josh Goldstein of the Stanford Internet Observatory, who found more than 1,000 LinkedIn profiles using what appear to be generated faces. It started with a sales message from a "Keenan Ramsey". DiResta looked at the profile picture and noticed that the woman wore only one earring, that bits of her hair disappeared and reappeared, and that her eyes sat right in the middle of the image.
NPR found that many of these profiles were used to generate sales leads: more than 70 businesses were listed as employers, and several said they had hired outside marketers. LinkedIn said it removed the profiles that broke its rules. Such faces had also been used to push disinformation and harass activists, NPR noted. Romance and investment scams use the same trick, as we describe in our guide to SI scams with fake photos and fake profiles.
How to spot an SI-generated face
Research has found several tells in GAN faces. None is proof on its own, and each has known limits.
| Tell | What to look for | Limit |
|---|---|---|
| Eye position | Eyes sitting right in the middle of the image, a placement DiResta says becomes familiar once you have seen many such faces | Cropping can move the eyes, and real headshots are often centred too |
| Corneal highlights | Light reflections that differ between the two eyes in number, shape or position | Only valid for a frontal portrait with distant light sources visible to both eyes |
| Pupil shape | Pupils with irregular, non-elliptical edges | Eye disease, blur, occlusion or a strong angle can do the same |
| Accessories and edges | One earring only, bits of hair that disappear and reappear | Easy to crop out or retouch |
| Background and clothing | Smeared shapes, blurry or swirly clothing, objects that are not objects | A plain background hides it; the PNAS study picked faces with uniform backgrounds |
The eyes
In a 2020 paper, Shu Hu, Yuezun Li and Siwei Lyu of the University at Buffalo compared the bright reflections on the corneas of both eyes. In a real portrait taken facing the camera, with lights some distance away, both eyes see the same scene, so the reflections look alike. The model has no notion of that physics. Testing real FFHQ photos against faces from This Person Does Not Exist, their automatic method reached an AUC score of 0.94. The same authors list the limits: false positives when a light is very close to the subject or visible to only one eye, and no use at all when there are no reflections.
A 2021 paper by Hui Guo and colleagues at the University at Albany and Buffalo looked at pupils instead. Real pupils are close to round or elliptical; in GAN faces they often have irregular boundaries. On 1,600 real and 1,600 generated faces at 1,024 pixels, their method scored an AUC of 0.91. They warn that diseased or infected eyes, occlusions and skewed views can also produce irregular pupils.
Teeth, earrings and the rest
StyleGAN2's authors documented a "phase" artifact in the earlier model: details such as teeth did not move with the face, so in their example the teeth stayed aligned to the camera while the head turned. CNN's 2019 report noted teeth that "often look a bit strange", earrings on just one ear, unusual skin marks and blurry or swirly clothing. Newer models and a quick retouch remove many such flaws. Our general guide on how to spot SI images covers the checks that also apply to other kinds of generated pictures.
Most tells above were measured on GAN faces from 2019 to 2021. Faces from newer generators may not share them, and Nightingale and Farid deliberately chose faces with no obvious errors because a deceptive user would do the same. Treat a missing tell as "no evidence either way", not as proof that a face is real.
Checks that do not depend on the pixels
- Look for a history. Run a reverse search on the picture (see how to reverse image search). A real person usually appears in more than one place, at different ages and in different settings. A generated face often appears only on the profile you are looking at.
- Check the claims around the face. In the LinkedIn case, NPR found that the employer and the university listed on the profile had no record of the person.
- Use a tool as one signal. A detector such as our SI image detector looks at patterns in the pixels and reads the file's metadata. It gives a probability, not a verdict on a person's honesty.
We tested a face from the site
Wikimedia Commons hosts a face credited to thispersondoesnotexist.com and to StyleGAN2, dated 2022 and marked public domain. Because nobody is in the picture, it is a fair sample: there is no real person to protect, and the origin is known. We ran this copy through our SI detection tool, exactly as a visitor would.
Result: 99% likely SI. The classifier found the statistical patterns of an SI-generated image in the pixels of this copy.
Other hints shown: no camera EXIF data (a weak hint, since many real photos lose it when shared) and a close-up face filling about 40% of the image.
Metadata: JPEG, 1,024 by 1,024 pixels, about 465 KB, no camera EXIF data, no Content Credentials, no SI tag in the XMP. Nothing in the file says how it was made.
What the test covers: one face from one generator, in the copy stored on Commons. It does not tell you how the tool performs on faces from other generators, on heavily compressed screenshots or on retouched pictures.
We also looked at the picture ourselves. The eyes are the surprise: zoomed in, the reflections in both corneas look much alike, so the eye test from the 2020 paper would not have caught this one. The giveaways are elsewhere. A brown rod-like shape rises from the top of the head into the frame, a shapeless pink mass sits at the lower left where a shoulder or object should be, and the dark background has no recognisable detail. All of that could be cropped away in seconds.
What This Person Does Not Exist taught us
The site was built to raise awareness, and the research it fed into delivered an uncomfortable message. People judge generated faces as at least as trustworthy as real ones, and the most confident judges are often the most wrong. Visual tells exist, but they belong to particular models.
The practical answer is to stop treating a face as evidence. A profile picture shows that someone chose a picture, not that the person exists. Check where else the image appears and whether the details around it hold up. Nightingale and Farid also call for robust watermarks built into generators, the kind of provenance signal we explain in our guide to C2PA and Content Credentials. If you want to see how hard the task is for yourself, try our SI or Not game, where real and generated pictures are mixed together.
