186 comments

[ 3.2 ms ] story [ 224 ms ] thread
At first I was tricked a handful of times, but I trained myself in what to look for. At first uneven blemishes proved a useful heuristic, but then when I looked deeper I found the edges and backgrounds were even more effective. The fakes somehow feel like they are in this... Oily world of illusions.
I got the first 5 wrong because I thought I was meant to pick the computer, not the real person. After that I did ~20, got them all correct.

There's alot of tells. Glasses make the edge of the eyes look strange. Around the ears, hats, sometimes the backgrounds etc, the blurs are wrong or corrupt.

However, if I didnt know 1 of the 2 were generated, I wouldn't look for anything and would probably just assume it's real, unless there was real obvious corruption on the face.

On a streak so far.

I'm playing "can you tell which picture has a non blurry background and has no artifacts?"

edit: My first mistake is when I thought a piece of fabric on a human was unusually warped.

My game is looking at the eyes, seems like the model has a tendency to make faces that have pretty much mirrored eye shapes and sockets, pupils being clear of imperfections or just plain circles or even both pupils being identical to each other, tilt is also a huge issue, most of the AI images have their output "looking directly into the lens" and nearly perpendicular to the aperture, real humans are off center in more ways than one in all of these aspects and more, as well as having numerous orientations

To me, looking at the background is kind of cheating to sus out facial features, after all we are trying to figure out if the face is real not the background

True, in most cases, backgrounds with a lot of elements (ex: a house in the background) were part of real pictures
Typically in AI generated images another giveaway is that when the head covers the entire height of the picture, the background may randomly change between the left and right side in implausible ways.
>My game is looking at the eyes

I look at the hairs. In real images you can see their fine threaded structure, in fake ones it's rather blurry and inconsistent.

Looking at the ears has worked for me 100% of time.
Outside of obvious facial artifacts the background was the next give away. I got it correct every time.
Same here. Just looked at the background and got all (10/10) right.
Reflections (highlights) in the eyes being different, artifacts and so on. But that said, if I weren't looking for fakes I'd probably accept them as real enough.
As soon as I started paying attention to the faces I lost my streak; they are basically perfectly realistic. But the backgrounds are a dead giveaway.
so next: generate green backgrounds and use the same background in both pictures
This. The generator is really bad at compositing people into the image. So while from the actual face it's sometimes hard to tell, backdrop and foreground items (like a mic or toy) are a giveaway. So is face paint or unusual props (fake mustarch or carnival custome). Especially since a lot of images from the real humans dataset seem to contain these.

So next time you're on a video call with someone and you're unsure if they're human or not, ask them to draw a letter on their face or have them dress like a pirate ;-)

>So next time you're on a video call with someone and you're unsure if they're human or not, ask them to draw a letter on their face or have them dress like a pirate ;-)

is that a thing now? my cursory search for "deepfake video call" gave me https://www.youtube.com/watch?v=wYSmp-nrJ7M but other than that, there is just youtubers goofing around with the tech. Do you know of a "good quality" deepfake video call that can fool us like the whilefaceisreal does sometimes?

Some German politicians thought they had video calls with Klitschko. But it was a deep fake.

Maybe there are more instances, but this made the local news.

I only get localized results on the phone, so here is a German link. Use deepl or Google translate: https://www.tagesschau.de/investigativ/rbb/deep-fake-klitsch...

ah.... so it "has" already caused some problems. in that case it does look like an interesting idea....

btw, firefox has a first party translator that is on-device so that works nicely, btw

Yeah. In that case I think the intention was to have them look bad. And since the faked person is a celebrity, enough data was available to produce a fake of suitable quality.

Maybe we will see this in the future for CEO scams. Though in that case maybe a good UI that clearly indicates that the victim is called by an external user "Mr. Big CEO <hackerperson@totally-not-s.us>" might already be helpful.

human trains GAN model --> GAN model trains human
I'll simply observe that it is easy to tell a fake face when presented an either/or choice and when specifically asked to. Most of the time we aren't looking as closely, so while I see some commenters being very happy about their accomplishments, I don't personally see a reason to rejoice.

Regardless, the AP news article[1] linked under the "methods" page provides some useful reading on how to detect these faces, for anyone interested.

[1] https://apnews.com/article/ap-top-news-artificial-intelligen...

This is a very important distinction. With deliberate attention, you can indeed catch many fakes in this kind of scenario (to me it seems the background is often a giveaway, but you do need to focus on it)

But in passing, accompanying a news article, a tweet or an instagram post, are you paying as much attention? Those are the scenarios where the potential for harm is much bigger.

Here's the important paragraph (I made it bullet points):

... reeled off a list of digital tells that he believes show the Jones photo was created by a computer program, including

- inconsistencies around Jones’ eyes,

- the ethereal glow around her hair and

- smudge marks on her left cheek

For me I didn't even look at the face. To me the obvious giveaway that made me spot all of them was, first of all the unnatural bokeh that you see in all AI images that don't look like anything a camera would produce. And the second thing is looking at clothing that folds in strange ways.
Exactly my thought - I scored 5/6 choosing very quickly the one which seemed more imperfect at a glance, but I'm no Reddit-expert photoshop-identifier or whatever. They all looked real, I'd have assumed any of them were without the 'one of these is fake' context.
I got 5/5 correct just by looking for weird artefacts in the hair and background. Just looking at the eyes alone was much harder (I sometimes couldn't tell).
Ears and neckline are also good indicators.

Face on, ears seem to become smudges.

My personal observation is that these generators fail miserably when generating low-detail parts and hair. In many of these pictures you don't have to look at the face at all, but rather look at the background and the one with heavy artifacting will be fake. In "enterprise" style pictures one can look at hairs and find heavy artifacting there.
for me the eyes looked really off in all the fakes, the pupils seems like the wrong size for the level of light and looked like the shape of the eye didn't fit into the bone structure of the face.
Problem is most people don't look heavily at people's hair and backgrounds when seeing a random photo somewhere, they look at the face!

I imagine this kind of stuff will trick a lot of people in practice.

Sure, I completely agree on this one - AI generated faces have been pretty decent for years now. The quality is currently in some weird space: with careful preselection they can fool unsuspecting reader passing by and yet at the same time reader looking for fakes will detect them (given high enough size/quality) with high confidence.
Yeah, I had exactly the same reaction. When I take the time to scan for artifacts, I get close to 100%, but when I try to do it quickly, I get close to 50%.

That 100% will gradually come down as the tech improves. And I'd guess the tech is already good enough that most people won't be able to improve on 50% success at first glance -- I don't think my instinct would noticeably improve with practice.

You can also look for artifacts but yeah, it could have both faces generated by AI and in many cases you wouldn't be able to tell.

If you know how to look for imperfections and quirks it's easier but ain't nobody doing it for just a image without "this might be AI" context

I think that's true to an extent, but serious errors such as the woman with what looked like a horn growing out of her cheek to match a partially-occluded earring seem to happen often enough that they would call attention to people being fake.

Apart from happiness (smiling), all of the deep fakes showed a blunted affect. Genuine humans tend to have quite expressive faces, and the many of the fakes looked like NPCs from an Elder Scrolls game.

These lead me to believe that a situation where deep fakes might matter e.g. security video presented as evidence in court, it would be possible to start picking up the deepfake artifacts/signatures even for a human expert.

Look not at the face but around the ears and the backgrounds.
(comment deleted)
Seems like I have been looking and playing around with too many of these AI-Generated imagery! I got the first 15-odd all correct.
Since this uses StyleGAN, it's relatively easy to tell when an image is fake or real, since the networks seems to have trouble with backgrounds and faces that are directly adjacent to the main face.

However, since diffusion models are all the rage now, I think we would perform significantly worse with landscapes or images of fruits and animals, especially if the task is "distinguish between the real and fake art".

It's usually the glitchy artifacts which give the clues.

Not the face itself, bur what's around it/background/other objects etc.

Check for weird-looking "something's not right" objects/backgrounds and you'll get most of it fine.

I got five in a row without looking for glitches or background objects, then stopped.

Look really carefully at a small area of skin. See if wrinkles, pores, hairs, and minor skin imperfections are present. See if they make sense in the context of the rest of the face.

It asks to click on the person who is "real", which gets a bit strange when "real" seems to be someone in green contacts and a wig for cosplay: https://www.whichfaceisreal.com/realimages/12481.jpeg

Biggest issue seems to be a number of images of people consuming their deformed selves: https://www.whichfaceisreal.com/fakeimages/image-2019-02-18_... https://www.whichfaceisreal.com/fakeimages/image-2019-02-17_...

Well, even Cos-players are real persons...
They're a real person doing their best to look like a fake person.
This game seems quite easy. When I know one of them is computer generated and one not, it's easy to pick the real one. Like a multiple choice question is much easier than otherwise.

I didnt get a single one wrong, and am now playing with the rule that I have to decide within a few seconds, still all right.

Still, they're pretty good, If one of the CG images came up by itself in the course of other business I wouldn't bat an eyelid.

I suggest to develop an AI that can tell which one is a robot or a human
Isn't that what we are training when we complete tasks on these sites?!
Eyes, teeth, and Ears, that’s my endgame… plenty of anomalies there, took me about one per second to get 20 right in a rows then I stopped playing. I figured I was trining the the system… your welcome!
My hypothesis was "children are always real". Got about 15 right before a fake child appeared.
Perfect score, but it’s such an impressive technology. Traditional graphics with triangles and ray tracing are fake at a glance, but here you need attention to detail and a bit of wit.
I find the result impressive.

I'm sure this fools a majority of people, contrary to the comments here. Obviously, with detailed analysis, you can probably spot the difference, but in day-to-day activity, and without knowing that one picture is fake, you will fool even more.

In this context you are looking for fakes, but on a website with Sales Person Avatars e.g. in a Chat I wouldn't search for fakes.

I wouldn't assume the Avatars are the real persons either but yeah persons...

After doing 30 I was able to differentiate very quickly, it's surprising how easy it is to detect these. You can tell by abnormalities in ears and AI probably wont show you hands because it struggles a lot. The backgrounds often look correct but dont make architectural sense. I also noticed if I dont look at the person in the eyes it sometimes is a tell, I'm not sure why though.
After a while I could determine the fakes by carefully scrutinizing the ears. On first impression I can't tell them apart though
Same as the other commenters, I got the first couple wrong, but then quickly realised what I was looking for. You can see artefacts in the skin of many of the faces, and often the ears were the giveaway.
I found that in direct comparison, the background often was enough to tell the difference - but that was mostly because one of the images had a detailed background with text or architecture, which I know the AI would struggle with.

I think a similar test that is not asking for a direct comparison but just "is this image real?" would be much harder, since there is no better "safe" choice to fall back on.

20/20 chosen correctly. Yes I think I can. The AI images are very impressive, but there are always artifacts that give it away.
Is this a serious topic worthy of serious responses from high ranking HN readers?

Depressing. They're both photos. A photo (of 'reality') is at its very best, at very best, already just a representation of the subject. Both are (technically) fake, aren't they?

There's a rather important distinction between an image of reality and a purely synthetic image.

I'm really not sure what point you are making.

A photograph of reality can and often does look 'unreal', or odd, or fake. So many, many, aspects cause this; lighting, expressions caught mid point, even the colour. My point is, the invitation is nether a robust or an intriguing one. Nevertheless, it is a success for other reasons.
It seems like the real face is always the less "perfect" one