AI-created faces now look so real, humans can't spot the difference

Humans can no longer reliably tell the difference between a real human face and an image of a face generated by artificial intelligence, according to a pair of researchers. Two boffins – Sophie Nightingale from the Department of Psychology at the UK's Lancaster University and Hany Farid from Berkley's Electrical Engineering …

[Evil Auditor]

...and if they don't, just don't create the tech...

Yeah, like this has ever worked.

_{I for one would like to get a fake face for myself which looks pretty much like me. Specifically for use during early morning video calls. I believe it would be beneficial for all participants: me not needing to pretend that I'm awake and the others not needing to pretend that they don't notice me pretending to be awake.}

[Dinanziame]

fake face for myself which looks pretty much like me

It's called a picture (or a selfie if you're a millennial). Just set it as your background image for the VC and go back to bed.

[ThatOne]

> Yeah, like this has ever worked.

Came here to say the same thing. "Cool" will always trump over all other considerations, and "it might be easily abused and dangerous, but it's so interesting to create" is a proven bullet-proof excuse to invent "solutions in search of a problem" which will make criminals happy and rich. It's the adult version of playing with matches, with the added benefit that it's someone else's house who will burn down, not yours, so why hesitate.

[Anonymous Coward]

Not even close to new. Back when films and photos used actual film, post production was done in darkrooms to retouch photos and adjust such things as color or contrast in both photos and film.

The prime motivators, then as now, are vanity (e.g. looking good in virtual meetings) and money (e.g Hollywood avoiding reshoots, unaging old actors, and doing away with actors entirely).

[chivo243]

S or R?

Real and ? I'm confused on a Monday morning.

[nagyeger]

Re: S or R?

Simulated? But what do the percentages mean? It probably says in the text, but it's monday AM here too.

[NATTtrash]

Re: S or R?

Simulated? But what do the percentages mean?

It doesn't say it in the text (at this moment). Similarly puzzled I checked twice.

The caption of this image in the original study however says:

The most (Top and Upper Middle) and least (Bottom and Lower Middle) accurately classified real (R) and synthetic (S) faces.

So the percentage is the number of people that rated the picture correctly as R or S.

Having a think about it though: what is it with the human species that we always end up making everything around us become/ generate a reflection of ourselves? I mean, there Shirley must be more useful challenges for that invested time/ tech/ money/ energy?

[Charles 9]

Re: S or R?

Nope, because nothing is more precious to us than our identity. It's damn instinct, ecen.

[General Purpose]

Re: S or R?

Five paragraphs further on, there's "The final images used in the study included a diverse set of 400 real and 400 synthesized faces representing Black, South Asian, East Asian and White faces."

Yup, I failed the Register's real-reading test the first time too.

[steviebuk]

Fascinating

This is a bit of AI that fascinates me. Seeing all the fake faces the AI creates yet they look like a real live human. Sent one image to a group of friends asking if they knew the person. Obviously everyone said no, then the real that the face doesn't exist, just AI created. Appears I was the only one that found this interesting.

[Rafael #872397]

Re: Fascinating

So.. you're the only robot (sorry, synthetic person) among all your friends?

[Anonymous Coward]

Re: Fascinating

I find it interesting. This is the first step to an AI that can create a talking, moving, acting face and body.

In just a few years, your favourite actor may never grow old and you can have a continuous sequence of "Raiders..." films without the need to "introduce" a replacement character in an attempt to secure the franchise as the original star moves slowly but surely towards the retirement home.

Now all they need is an AI to produced decent plots and dialogue and Hollywood can close shop.

[Kubla Cant]

Re: Fascinating

Now all they need is an AI to produced decent plots and dialogue

Accounts suggest that the way film scripts are written is already a Generative Adversarial Network, except that the components are human. It should be a simple matter to replace them with computers.

Actually, how do we know they haven't already done so?

[ThatOne]

Re: Fascinating

> your favourite actor may never grow old and you can have a continuous sequence of "Raiders..." films

Don't forget the issue preventing it isn't usually the actor(s) aging, but the successive scenarios going progressively down the drain, till you really, really wish they'd stop flogging that corpse already.

[TeeCee]

Re: Fascinating

While the criticism of the most recent one in that series was loud, I'm a bit stumped as to what the problem was.

Maybe somebody needs to explain to me exactly why transdimensional aliens and crystal skulls is more far fetched than sky fairies and magic boxes/stones/novelty mugs.

[A K Stiles]

Re: Fascinating

Shia LaBeouf

[ThatOne]

Re: Fascinating

> Maybe somebody needs to explain to me exactly why transdimensional aliens and crystal skulls is more far fetched than sky fairies and magic boxes/stones/novelty mugs.

Well, it's not in the ingredients, it's in the cooking. As you said yourself, the ingredients have always been kind of corny, and that even made some of the charm of the franchise (taping into old Hollywood adventure movies' atmosphere). The problem is the last movie lacked the charm and the wittiness of the previous ones, leaving only the raw corn.

[Ken G]

Re: Fascinating

You really didn't like Police Academy VIII did you?

[stiine]

Re: Fascinating

Actually, sometime in the near future, your favorite actor is going to get a lump sump payment and then be fired, along with every other actor and actress in Hollywood.

Industrial Light and not Magic will then buy another building full of computers and start releasing comedies, dramas, horror films, etc that won't 'star" any humans. No more catering, no more best boys, no more costuming... No more Emmys and no more political encorsements. Also, Alec Baldwin won't get to shoot anyone else.

[Anonymous Coward]

Re: Fascinating

That will save the studios a lot of money because they will "own" the face and "voice" (quite literally I assure you) and they will be able to choose from a pool of talented but ugly and unknown actors to provide the template upon which the polished images are superimposed. The gossip rags will have to pay a fee to the studios to use the images for their fake gossip news.

[JamesTGrant]

Re: Fascinating

Haven’t read a post which makes me feel so weird in a long time. I think you are exactly right.

[Filippo]

"The authors suggested those developing such technologies should consider whether the benefits outweigh the risks – and if they don't, just don't create the tech."

This approach doesn't even really work long-term for WMDs, which are insanely difficult and expensive to research. I don't see how it could possibly work for something that only takes readily-available specialist training and off-the-shelf computer tech.

People who are concerned about fake news need to get a grip. It's not a technical problem, it's a political problem. Stop looking for technical solutions, start looking for political solutions.

[Charles 9]

"People who are concerned about fake news need to get a grip. It's not a technical problem, it's a political problem. Stop looking for technical solutions, start looking for political solutions."

Then there's no solution because political problems are icreasingly personal problems: matters of identity. And there are few things more vigorously defended than identity.

[ThatOne]

As Charles 9 said.

In our societies which are increasingly losing their points of reference, individual identity is increasingly defined by ideology, which includes all kinds of religions and more or less strange convictions.

"Fake news" aka lies have always existed, but till recently they lived a shameful shadow existence, while nowadays they are political and societal stars: People aren't ashamed of telling lies anymore, they take a pride in doing so as long as it serves their cause.

So indeed, there is no "political solution" to peoples' identity crisis, except maybe summary execution (potentially using gas, in collective shower rooms?...). Politics is (or at least should be) about keeping a coherent society of different people with differing agendas working, not about telling individual persons what to think. That's way too dangerous a path.

[Doctor Syntax]

Good. Start feeding about a trillion to Clearview to poison their data.

[Anonymous Coward]

Wonderful crowdsourcing project. I'd even open a Facebook account for that!

[quxinot]

Heck, I'd open half a dozen.

[Mike 137]

Detail and perception

The highest resolution images in the data sets are 1024x1024 pixels. The typical facial image as reproduced in the paper, is just under 0.6 of a frame wide, so the facial image relied on is at best a mere 600 or so pixels wide. That's a pretty low resolution for adequate reproduction of small scale characteristics we use subliminally all the time to recognise individuals.

Furthermore, as a general rule, all studies that are based on human judgement suffer from an uncontrolled factor - the attentiveness of said subjects (the same problem that besets the Turing test). Plus, the paper does not disclose the breakdown (e.g. by ethnicity) of the subjects - a factor that could introduce bias (particularly in respect of the 'trustworthiness' judgement).

Consequently, although this is an interesting preliminary study, it would be unwise to assume it represents a universally applicable finding.

[jdiebdhidbsusbvwbsidnsoskebid]

Re: Detail and perception

"a mere 600 or so pixels wide. That's a pretty low resolution for adequate reproduction of small scale characteristics we use subliminally all the time"

We can check that.

Just doing some numbers, if a human face is 0.2 metres wide and represented by 600 pixels, when viewed at a distance r, each pixel subtends an angle of 0.2/(600r) radians.

The angular resolving power of the human eye can be calculated from lambda/d where lambda is the wavelength and d is the diameter of the pupil. At 550nm (about the middle of the visible band) and d of 5m, this comes out to about 1.1x10^-4 radians. Setting this equal to the angular width of a pixel gives 0.2/(600r) = 1.1x10^-4 and solving for r gives r=3.03 metres.

So the images used represent a face seen at a distant of about 3 metres. Could be closer, but 3m seems like a reasonable everyday distance to use. Certainly not wildly inaccurate.

[jmch]

Re: Detail and perception

"d is the diameter of the pupil. At... d of 5m... "

!!!

[jdiebdhidbsusbvwbsidnsoskebid]

Re: Detail and perception

Good spot! Should have been mm not m! What's a factor of 1000 anyway? I think the numbers are still correct, just a typo on the unit.

[jmch]

Re: Detail and perception

Yes it's just the unit. The rest of the numbers look correct, and by the way, excellent analysis!

[Richard 12]

Re: Detail and perception

Tell me Legolas, what do your elf eyes see?

[Precordial thump]

Re: Detail and perception

Nothing. In this bright sunlight, my pupils are so small the photons have to go through single-file.

[johnrobyclayton]

Re: Detail and perception

I have never upvoted before.

This one made me laugh out loud.

QM, AI, LOTR, All in one funny package

[Version 1.0]

Re: Detail and perception

The elf eyes see the world, humans need to join the elves on the boat and leave the AI (All Idiots) world.

[Mike 137]

Re: Detail and perception

BUT:

A digital photograph pixel is some (possibly non-linear) integral of (as a minimum) a 3x3 photo site array, so the actual resolution is substantially lower than the linear pixel density. And even if this were not the case, 3 metres (10 feet) is rather a long distance for a head and shoulders photo.

Consequently, I still maintain that the criteria for these experiments are rather lax - particularly as the paper does not declare either the image resolution or the viewing conditions. I had to look the former up from the data sets themselves, and the latter is not mentioned anywhere. So although interesting in principle, this verges on rather sloppy science.

[TheProf]

Trouble

I have trouble spotting fake faces.

The papers (real or online) are full of 'celebrities' that I don't recognise and I'm sure most of them are synthetic.

The one called Rylan(?) is obviously a fake. That painted on beard and the 'Banana Splits' teeth. Not going to fool anyone.

But for the rest everything is so instagram filtered and tweaked it's difficult to tell which is which.

[Pete 2]

Missing option

> human evaluations of both real photographs and AI-synthesized images

How about faces of those who have had cosmetic surgery?

I reckon those would be even easier to spot than simulated faces. But would they appear more "fake" than photoshopped faces of real (unaltered) people.

[Anonymous Coward]

Re: Missing option

I suppose it's because the AI has been trained on real world photographs, and the majority of people haven't had cosmetic surgery, so that demographic has less influence in how the algorithm works.

One thing I've noticed is that the example AI-synthesized faces that I've seen, while not all necessarily being handsome/good-looking, definitely fall towards the right-hand side of scale for looks. Training data based on images available online, which tends to skew that way? Seeing as I'm ugly as sin, I guess that means I don't need to worry about being mistaken for a creation of AI

[Omnipresent]

it's not your imagination

Time really IS speeding up. The universe (spacetime) is expanding at an ever increasing rate, and the monkey apes are marching faster towards self made destruction with it. California will fall in the ocean, and Texas will be a flood plain. This has all happened before. You are the BORG, and will be destroyed, like all life "artificial or not" when the Universe is ready for a rebirth. You are fodder for the next world, with an ego to match.

[Anonymous Coward]

"The authors suggested those developing such technologies should consider whether the benefits outweigh the risks – and if they don't, just don't create the tech."

We tried that, but somehow still got Terminators. Money rulez, altruism droolz.

[Omnipresent]

I will be laughing as i watch their money burn in the sun, and as they die twice (once in real life, and once digitally). Somehow, we got a lot more done before computers, and had a lot more fun broke.

[Jedit]

We saw this coming

Or we did if we watched The Book of Boba Fett. No spoilers, but that includes a deepfaked character whose voice was even synthesised. And even though you knew what it was, it was still tough to tell. If it's getting that good in TV - not even in a movie - then it's going to spread wide, and soon.

[Brewster's Angle Grinder]

I haven't see the show. But that's the legitimate, bona-fide money-making use case: Hollywood. WMD can only kill. Deepfakes of dead actors can make money, even if people can abuse the tech.

[stiine]

Why?

Who says that you have to use a living actor/acress as a model. Why not use a represenation of an 21 year old Hedy Lamaar. Find a few popular ones who are out-of-copyright and use them.

[Brewster's Angle Grinder]

I'm going to get a tatoo with a QR code of my public key...

Looking at some of the real faces that people failed to identify, does this presage a change to less overt make up and a more naturalistic look? (Although, fashion being fashion, it will probably flip and we'll all have to look like synthesized images. *sigh*)

[ThatOne]

Re: I'm going to get a tatoo with a QR code of my public key...

The criteria of "fakeness" are usually complexity and imperfections, plus realistic (thus imperfect) lighting.

The top row is a good example of this: Bad lighting, wrinkly faces, bad skin. Looks real.

The 3rd row shows people with flat lighting and perfect skin, much like the fake bottom row ones. That's why everybody thought they were fakes too.

[Scott Pedigo]

OK, Now Do the O-Faces

And then the people who score really high are going to have to explain to their partners how they have so much experience recognizing them.

[Paul Hovnanian]

I'm afraid ...

... that following the long term Covid lockdown and isolation, I am losing the ability to recognise human faces, either real or synthetic.

[Solly]

Thankfully the tech isn't ready to do it in real-time

https://www.youtube.com/watch?v=qXLugdeogQU

Oh nevermind.

[Anonymous Coward]

...creates potential problems such as nonconsensual intimate imagery

That's not the same though - real faces imposed on someone else's body.

[General Purpose]

We may be reaching the point where it's harder to process a real image of a face (rotate it, adjust the skin tone and lighting, change the expession and graft it onto another body) than to simply train an AI to synthesise suitably posed and lit faces based on a database of images of one person.

[jdiebdhidbsusbvwbsidnsoskebid]

I don't think they are talking about real faces superimposed over someone else's body. I think they are talking about atrificially created faces and bodies that look just like a particular person - whoever you want.

[Ozan]

I guess we will see AI generated porn next year or so.

[Circadian]

AI porn

With how weird AI gets things, there will be all sorts of unbelievable stuff that no-one could imagine… oh wait, Rule 34. Does this mean that it’s already happened or that even AI cannot possibly keep (it) up?

[jmch]

photo vs movie

Firstly, this tech seems to be limited to photos and not movies. I would think full-realistic AI generated movies would be far more complex, although given the rate of advancement they surely can't be far behind.

Secondly, I would think that as far as nefarious purposes are concerned, simply being able to generate a photo is limited in scope. Based on what I read of the process, the AI is simply generating pictures that can fool another AI (and only incidentally, humans) into classifying a fake photo as real. As with all of these "AI"s it's not really an AI it's a pattern matcher/generator. It doesn't 'know' anything about what it's generating. It's not like someone can tell the AI to generate the face of a specific gender / age / race / body type etc, or specify a particular background, facial expression etc (short of retraining the dataset with only that type of image, which sort of defeats the purpose of having an AI do it for you). Any fraudster who needs to use a photo of a random face, or even a very specific type of face, can already pick them by the thousands on the web.

What is more concerning is the indistinguishability of photoshopped / deep-fake photos and movies.

[heyrick]

Well, I guess I failed...

To me, the top and bottom rows look real, and the middle rows look fake (faces a little wonky and odd). But, yeah, half of them are actually real and half of the ones I picked are actually fake.

But, then again, it's getting increasingly harder to tell who is real and who is fake In Real Life, never mind a photo!

[Bartholomew]

This Person Does Not Exist

Not one mention of "https://thispersondoesnotexist.com/" a site that generates images of people who do not exist.

The way I tell if it is real or fake is by ignoring the human and concentrating on the background! It is either heavily blurred or just does not make any sense if you really look at it. (e.g. https://imgur.com/a/cBRicKs )

[sreynolds]

I think that it has been around for a lot longer....

I mean I always thought that Zuck looked a bit fake, but him moving to the metaverse and all just goes to show that he isn't real.

[Anonymous Coward]

All the synths have perfect skin and no wrinkles.

[Maryland, USA]

Darn.

I was ready to propose to Miss 88% (S). I still might.