if there’s a picture, the person obviously exists.... right?
September 8, 2019 2:15 PM   Subscribe

Which Face Is Real trains viewers to discern between computer-generated faces and actual people's faces.
New adverserial machine learning algorithms allow people to rapidly generate synthetic 'photographs' of people who have never existed. Already faces of this sort are being used in espionage. Computers are good, but your visual processing systems are even better. If you know what to look for, you can spot these fakes at a single glance — at least for the time being. The hardware and software used to generate them will continue to improve, and it may be only a few years until humans fall behind in the arms race between forgery and detection.

Which Face Is Real uses images from thispersondoesnotexist.com.
Photorealistic face generation software StyleGAN uses a generative adversarial network (GAN) approach, in which two neural networks play a game of cat and mouse, one attempting to generate artificial images indistinguishable from real photographs.
The site was created by professors Jevin West and Carl Bergstrom as part of the Calling Bullshit project, previously on MeFi.
posted by spamandkimchi (48 comments total) 25 users marked this as a favorite
So, I assume that while we're doing this, we're also training the AI to make better fakes, right?
posted by jacquilynne at 2:19 PM on September 8, 2019 [44 favorites]

Jacquilynne, ordinarily I'd suspect something similar, but I used to be in the same department with Jevin West and Carl Bergstrom. IMO, their hearts are as pure as hearts in academia get.
posted by palmcorder_yajna at 2:24 PM on September 8, 2019 [5 favorites]

Hint: You can identify 90% of the real faces instantly by looking at their background. If the background is identifiable, it's a real person. If it looks like large scale jpeg artifacts, it's fake.
posted by justkevin at 2:26 PM on September 8, 2019 [16 favorites]

Also, if the face is obscured by something like hair, or if the subject isn't looking directly at the camera, it's much more likely to be a real person.
posted by J.K. Seazer at 2:31 PM on September 8, 2019 [3 favorites]

Ya the backgrounds are one giveaway but to me the tell is usually that generated hair still does kind of uncanny wispy wavy shit at the edges, like the smoke of a candle that's just been blown out. It could be compression artifacts but doesn't have quite the same feel.

It's funny because a lot of the people I know who went into hair/fur simulation also did research in smoke simulation. And computer generated hair ten+ years later still has a kind of ...smokey feel.
posted by potrzebie at 2:36 PM on September 8, 2019 [6 favorites]

Also, for a lot of the photos, something about the earlobes wasn't right. That, and the "person" looked to be in front of a crappy green screen
posted by NoMich at 2:37 PM on September 8, 2019

yesterday for April Fool’s my workplace had a short training article on recognizing computer-generated faces from real ones and one of the tricks mentioned was “count the teeth” and I just wanted to say that it’s both ironic and kind of horrifying how society has unwittingly cycled right back to IF YE MEET A MAN ON THE ROAD, COUNT HIS FINGERS LEST YE DEAL UNKNOWING WITH A FAE
posted by Rhaomi at 2:41 PM on September 8, 2019 [48 favorites]

Dang, only got 8 out of 10 right. No wonder I had so many bots in my friends list.
posted by Aya Hirano on the Astral Plane at 2:43 PM on September 8, 2019

Pretty impressive - it'll be harder when the AI models don't always look like their teeth are covered in clear fruit glaze. Zooming in on the AI skin shows up a lot “deep dream” type artifacts.
posted by bonobothegreat at 2:51 PM on September 8, 2019 [1 favorite]

Hell is other people.
posted by chavenet at 2:52 PM on September 8, 2019 [3 favorites]

I also noticed things like ears not being broadly symmetrical for no good reason, or jewellery that was different on different sides. The one that threw me was the man in the business suit compared against a girl with only half a silver chain around her neck, in some kind of blurry water-parky looking environment. I went with the business suit man. Still not sure how I could have spotted him.
posted by Merus at 3:00 PM on September 8, 2019

Some of those AI-generated faces have truly fabulous hats.
posted by ourobouros at 3:03 PM on September 8, 2019 [1 favorite]

All the better to distract you from its vestigial second neck, clearly.
posted by J.K. Seazer at 3:06 PM on September 8, 2019 [5 favorites]

I am not opposed to AI overlords per se, but I do question the skills we are training them. Just perusing this year's headlines, it appears we are teaching them strategy games, social media propaganda, facial recognition, and now, disguising as humans. It's like we want them to go skynet on us.
posted by ryanrs at 3:21 PM on September 8, 2019 [5 favorites]

Don’t forget that when they go on social media we train them to be virulently racist
posted by ejs at 3:25 PM on September 8, 2019

Hell is other people.
posted by chavenet

Hey! That's my line.
posted by Splunge at 3:27 PM on September 8, 2019

This was a fascinating game to play while waiting for my migraine aura to recede.
Which face is real? Both are partly obscured by flashing lines. Good luck!
I did surprisingly well by picking the face with highest contrast in lighting each time. Given that I had no way to tell how the features fit together, it seemed a good method.
Apologies if there are typos in this comment,I won't be able to tell for half an hour or so.
posted by Adridne at 4:08 PM on September 8, 2019 [3 favorites]

I'm terrible at this. Scary technology.
posted by Yowser at 4:09 PM on September 8, 2019

The give aways appear to always be some level of rendering artifacts where the granularity of details are being modeled too abstractly. A lot of people picked up on several of the larger ones, i.e. earlobes and teeth. I also noticed that there was consistently at least one or more abstract “blobs” that were seemingly meant to represent blemishes, but lacked the real world patterns most skin blemishes actually have (moles, acne, rosecia, etc). I also saw several repeated patterns the system appears to be using for stubble and smile lines which started to stand out quickly after getting about 50 correct in a row basing judgement on finding those patterns.

Fun game, would play for hours if my cat didn’t want dinner.
posted by daq at 4:40 PM on September 8, 2019

The babies are usually real. If the person is standing beside a friend, use the friend as the tell - often they are way screwy in fake ones.
posted by Meatbomb at 4:57 PM on September 8, 2019

I knew Tom Cruise was not a real person. You can tell by the teeth.
posted by rikschell at 5:22 PM on September 8, 2019 [15 favorites]

I got a lot better with practice. You can tell a lot by the background, particularly other people. Glasses can be a tell; are they symmetrical? What do they reflect? How do they refract? Do background details carry over from left to right sensibly? And, yes, do ears and/or jewelery match from left to right? Necklaces can be a reliable give-away.

To the credit of the face-sim people, the faces are quite credible across cultures and ethnicity. Their data set was clearly broad and deep.
posted by sjswitzer at 5:31 PM on September 8, 2019 [3 favorites]

Persona non generata?
posted by srboisvert at 5:41 PM on September 8, 2019 [8 favorites]

The teeth situation has gotten a lot better over the last year or so. Not long ago, this kind of thing would only pass the sniff test in the UK, thanks to the gaping hell scapes lurking in every generated mouth...
posted by kaibutsu at 6:05 PM on September 8, 2019 [1 favorite]

I'm averaging less than 50% and not improving. (I also can't follow the plot of an action film if it features two white guys with similar haircuts, so perhaps that's not surprising.) This is neat.
posted by eotvos at 7:06 PM on September 8, 2019 [4 favorites]

I am moderately face-blind, and I did very poorly on this. I guessed the wrong one almost every time, which means I did much worse than random chance would predict, which suggests that I recognize a difference between the real ones and the fake ones but think the fake ones look more real. I'm sure there's some conclusion to be drawn from this somewhere.
posted by Daily Alice at 7:20 PM on September 8, 2019 [5 favorites]

Huh. I'm averaging almost 100% (only got one wrong so far), but I also think that might be because I'm looking for differences. I would love to take this same test, but in a situation in which, somehow, I didn't know that I was looking at real vs fake. Like, yes, I can look for the weird background or hair, or if it's up against a photo with terrible lighting, the terrible lighting is real, but I wonder if this would trip the same mental sensors if I didn't know what I was looking for?
posted by kalimac at 7:25 PM on September 8, 2019 [3 favorites]

Also, I'm pretty sure I can tell Facetune from AI with high accuracy, so thanks, drag queens of the world!
posted by kalimac at 7:30 PM on September 8, 2019 [1 favorite]

100% here. The hair is super weird and their eyes are dead and lifeless looking to me. Even when the hair is good it either looks weird/whispy/glitchy or it's too perfect and appears coarse or strange. Some of them took more than a second but it was pretty much immediate for me.

I've always been very good at spotting digital graphics, FX and manipulation and frankly we should all be worried when that's no longer easy or reliable.

I just want to say I hope I never actually need to use these skills or find them useful in my day to day life but that day has probably been effectively here for a while, and the way this timeline is going who the fuck knows?
posted by loquacious at 7:49 PM on September 8, 2019 [1 favorite]

Yeah, I was able to pick them all out, but I started off by looking at thispersondoesnotexist and you pretty quickly realize there are "tells" in the photos it generates.

They're almost always white or Asian; this is probably not a safe heuristic, because it's just dependent on the data they trained the model on, but it seems like an indicator in this particular scenario. It loves to do wispy hair, but doesn't do it super well—hairs just sort of disappear into vapor instead of either going out of the focal plane or actually ending. It doesn't seem to really get how focus works either, in the images that are supposed to have been taken with a short DOF, the focus plane isn't where it ought to be consistently across the image... but that's also not a safe heuristic because it would be easy to add accurate defocusing as a postprocessing step if you knew someone was going to look for that.

It also seems to have a problem with the corners of people's mouths. A few of the images have weird "tears" that are like the visual embodiment of NIN's Only, a couple are real nightmare fuel / body horror for me. I'm really curious what this comes from in the model.

But overall I was actually pretty impressed / concerned with how good the sims were. In a situation where someone wasn't looking or expecting a fake, and the photo was small, I doubt most of them would get caught. E.g. in a social media profile. Or on a corporate staff-bio page. Someone has to be looking to catch them, which is really quite something.
posted by Kadin2048 at 8:15 PM on September 8, 2019

I did about 25 or so comparisons, got 100% correct. But probably at least 75% of the time I was looking at the background instead of the face itself, and without the much more obvious background cues there were several that I don't think I would have been able to do more than guess. Considering that it would be pretty easy to improve the backgrounds by just passing that part of the image off to a separate algorithm, I'd say this is a fairly strong under-rating of how good these AI-generated faces are.
posted by biogeo at 8:32 PM on September 8, 2019 [1 favorite]

The algorithm does a really shitty job at ears.
posted by mr_roboto at 9:39 PM on September 8, 2019

Yeah, same as biogeo.

I give it about a year (or less) before I can't tell, and already I don't know unless I'm looking for it.
posted by inexorably_forward at 10:15 PM on September 8, 2019 [1 favorite]

I knew Tom Cruise was not a real person. You can tell by the teeth.

Funny story, that.
posted by They sucked his brains out! at 10:49 PM on September 8, 2019 [2 favorites]

Going a step farther and pasting the faked faces into real images deepfake style would probably make things much harder...
posted by kaibutsu at 11:15 PM on September 8, 2019

Easiest way for me was to look at the hair/background interface. That alone got me to 80-90% accuracy. Adding in the rest of the clues/tests got me to ≈99% within a couple of minutes.

But it won't last. It will get harder faster.

Is anything that can be digitised going to become an unreliable witness to history?
posted by Pouteria at 11:17 PM on September 8, 2019

A friend's pic popped up in one of the pairs - I know that's bound to happen occasionally, given the source of their dataset, but it was still a bit weird!
posted by infodiva at 11:54 PM on September 8, 2019 [1 favorite]

A friend's pic popped up in one of the pairs

But which category was your friend's picture in?
posted by Umami Dearest at 12:09 AM on September 9, 2019 [7 favorites]

So, I assume that while we're doing this, we're also training the AI to make better fakes, right?

Jacquilynne, ordinarily I'd suspect something similar, but I used to be in the same department with Jevin West and Carl Bergstrom. IMO, their hearts are as pure as hearts in academia get.

... c'mon, people.

This isn't training AI to create better fakes. This is training AI to do better facial recognition, which is an area that Jevin West has been talking about quite a lot this year. Mostly in terms of creating a conversation on reasonable controls, aimed at ensuring accuracy. Which, hey! Refined data to tell what features do or do not flag as Real Faces sure would help!

Just as a general rule, don't help the police state.
posted by kafziel at 1:52 AM on September 9, 2019 [1 favorite]

I have prosopagnosia (actually more prosopamnesia) and yet I find it very very easy to spot these. Maybe that's why?
My brain doesn't generally do pareidolia so I don't match on faces the same way as other humans.
But for almost all of these I can see which is a real human and which isn't almost instantly with a very nearly 100% success rate.
posted by Just this guy, y'know at 2:59 AM on September 9, 2019 [2 favorites]

I've got a very good hit rate with this, but I'm a little disturbed because I know I've seen the image source site before, and I... think they're getting better? I am more often now choosing based on "this person is clearly real" than "this photo is clearly not". It could just be random chance, but I'm noticing that a lot fewer of the fakes have the teeth problem because they didn't have open mouths. There was only one with an earring, and the other ear in that picture was not visible. The backgrounds are still a huge tell, but a lot of them are more uniformly a single color. In a type of presentation where one would expect a full-face shot with a blank backdrop and with no "real" picture to compare to, I would be missing a lot more of these.
posted by Sequence at 7:49 AM on September 9, 2019

I'm really good at this apparently, because I've been playing on and off since last night and have only gotten one wrong. I started thinking about the tells, and got a poem out of it.

How to Spot a Human

Your lashes and brows lack
a certain airbrushed quality
and your teeth are solid
nearly all the way through.
Your flesh, if it folds
in on itself
does so in a predictable way
rather than in a manner
that makes my stomach squeeze,
like cat fingers, like Tetsuo’s head.
Real life skin, of real people,
rarely does this
(in real life).

If you smile, your gums
do not envelop your teeth, growing
around them.
Your teeth might be chipped
or missing (but they are solid
nearly all the way through).

You have only one neck. Only
one earlobe on each ear. Your hat
or your glasses or your scarf
do not give way to flesh,
becoming a part of you.

If you are a child you are not
the Gerber baby.

Your eyes might be red
from the camera’s flash. The light
from the window might come in
hard. Your skin is dark, or yeasty white.
Your teeth are solid
all the way through and if one thing
is certain, it’s that you’re a little
too much. A strained grin.
Frizzy hair. An unkempt eyebrow
that does not sink back
into flesh. A mascara clump.
A bead of sweat. A tension
in your mouth’s corners. What’s wrong,
human being? Are the thoughts
inside your real, actual head
too much for you today? Are you remembering
that thing your father said to you
when you were nine? The way a stranger
Grabbed your hand and took you somewhere? Are you thinking
about a poem, a bill to be paid--about politics or the black-gray
expanse waiting for you after death?

Hold on to those feelings. They’re the only things that tell us
that you are real.
posted by PhoBWanKenobi at 8:21 AM on September 9, 2019 [14 favorites]

I've started to get into a run of photos that has several Asian women in heavy make-up -- kawaii, harajuku, anime style stuff. It's much more difficult to tell the difference between fake people designed to look real and real people designed to look fake.

My last real person was wearing two different coloured contacts, for example, which looks super fake, but wasn't. I got that one right only because the fake person had a really bad compression artifact where her earring should have been.
posted by jacquilynne at 8:40 AM on September 9, 2019

I didn't do really well at first-- about 75%, maybe? Then I went to the Learn page and read through; they give a list of what to look for/at on the fakes. Which was helpful and shot me up to 100% easy.

BUT....this horrifying : " The StyleGAN algorithm is unable to generate multiple images of the same fake person. Right now, we are unaware of any software that can do that. So if you want to be sure that your tinder crush is a real person, insist on seeing two or more photos. At some point, software will probably catch up. But for now, multiple pictures offer powerful reassurance that the image is not a fake.

Well, it took three months, not a year or two. Egor Zakharov and colleagues at the Samsung AI Center have developed a way to create video of a person moving and talking, based on even a single sample image. Their video demonstration is stunning and well worth a look. Presumably one could supply their algorithm with a single StyleGAN fake face, and it would supply multiple angles and expressions of the same "person". For the time being it might be harder to show the same person in different outfits, settins, etc., but it's clear that we shouldn't be promising any silver bullets against the rapidly evolving technology. "

The italics is mine...I get that online con artists already use other people's images, etc. Now they can create someone just for you!
posted by twentyfeetof tacos at 9:24 AM on September 9, 2019 [1 favorite]

There's a certain muscular tension in the face of the real people, especially if they're not just smiling but expressing something like thoughtfulness or seriousness.

It's also a decent way to tell if someone is hiding something IRL. The tension is out of whack with the intended emotion.
posted by fiercekitten at 3:58 PM on September 9, 2019

I did terribly at this the first time I tried it. But after reading a comment or two where someone pointed out that there is frequently a wonky clone-stamp effect in things like beard stubble or hairlines, it became a lot easier to recognize, and I got nearly all of them right. The fake ones also have some wonky facial lines—sometimes there’ll be wrinkles on a person much too young to have them, or they’ll be going in the wrong direction for where they’re placed (like horizontally stacked wrinkles between the nose and mouth—that doesn’t happen IRL).

But damn. The fakes are good. If I just glanced at a picture in passing, I doubt I would even think to see whether it was a fake.
posted by Autumnheart at 7:44 PM on September 9, 2019 [1 favorite]

I got about 30 in a row correct right off the bat and stopped after that. This is cool because I was having an argument with some guy at work about some photos I thought were "obviously fake" (without being able to put my finger on precisely why) and he was calling me a conspiracy theorist etc etc so now I'll just challenge him to a contest with this to prove that my fake detection skills are superior to his.
posted by L.P. Hatecraft at 2:24 AM on September 10, 2019 [1 favorite]

The face generator is terrible at symmetry. If there's matching earrings, it's likely real. Matching glasses (check the edges of the frames), same. Face generator only generates faces and vague backgrounds: if anything else is clear (an object being held, a postcard, etc.), that should be the real one.
posted by ErisLordFreedom at 3:28 PM on September 10, 2019

« Older Set your timer   |   Anyone who says cross is a cop. Newer »

This thread has been archived and is closed to new comments