Maybe this belongs on /g/ idkCan you tell the difference between an AI generated face and a real one?Test here - https://facetest.psy.unsw.edu.au/aifaces.htmlThe average is 11 out of 20. Post your results. I got 14 out of 20 but I'm going to try again and post results because I think I figured out the trick.
OP here, I got 19 out of 20 on the next try after studying the images in more detail. It's kinda hard.
First try not even really thinking about it.I probably messed up more often by second-guessing than anything.
>>17054628>>17054646yeah, thanks but no thanksI'm not falling in the honey pot and training the AI to make undistinguishable fake faces from real ones
>>17054628>>17054628I also got 14 out of 20. I'm not quite sure what they're out to demonstrate. The actual human faces all have that unreal look of a professional portrait. Possibly digitally enhanced. Probably cherrypicked for that specifically: to look as much as possible like AI slop. Soft "science" strikes again.Can't wait for some tabloid to pick up on this and get Claude to write an article about how the normies can't tell reality from AI anymore and how the government needs to step in and regulate this with Digital ID fingerprints on all your (but not their) gens.
>>17054658>honey potNow I want to retake the test over and over with intentionally bad answer to "anti" train the AI
>>17054665>The actual human faces all have that unreal look of a professional portrait. Possibly digitally enhanced. Probably cherrypicked for that specifically: to look as much as possible like AI slop.Part of doing "good science" is to control for as many variables as possible. If a particular look can be replicated in a genuine photograph, then it is simply a confounding variable that should be controlled for.People who rely solely on the uncanny lighting that is common of AI would get tripped up when genuine photos have that effect or when the AI generates faces with different lighting. So controlling that particular variable by setting the same lighting across all images is the correct thing to do in this circumstance.
>>17054681>Part of doing "good science" is to control for as many variables as possible. That's exactly what you're not doing. How about you try it with a mix of real and fake news photos? I mean raw and detailed ones, not some faux-cinematic, person-focused crap with a blurred background. And if you want to make it real interesting, make it mix of localized and foreign photos. My bet: discerning viewers will recognize fakes set in their own country but not ones depicting (e.g.) war-torn third world shitholes.
>>17054671too late, m8, the AI has tricked your face ID and learned to fake your voice by listening to your callsit highjacked your 4chan account and took over your entire internet persona
>>17054684>That's exactly what you're not doingIt is, actually. The more superficially similar the photos look, the more you zero in on the actual facial recognition aspect, which is the goal of the study.>How about you try it with a mix of real and fake news photos? I mean raw and detailed ones, not some faux-cinematic, person-focused crap with a blurred background.That adds more confounding variables, which is the polar opposite of controlling for them.>And if you want to make it real interesting, make it mix of localized and foreign photos. My bet: discerning viewers will recognize fakes set in their own country but not ones depicting (e.g.) war-torn third world shitholes.That would be a different study entirely. Though probably interesting in its own right.For fun, here's this clip from a year ago with AI "news"https://youtube.com/watch?v=eQU70usQStw
>>17054692>The more superficially similar the photos look, the more you zero in on the actual facial recognition aspect, which is the goal of the study.They look superficially similar because they're cherrypicked to look that way. This isn't measuring anything about "actual facial recognition", which isn't even a well-defined goal.>That adds more confounding variables, which is the polar opposite of controlling for them.Like what? Name them. It's cute watching a social "scientist" trying to reason about science.>That would be a different study entirely.It would be an actual study.
>>17054695>They look superficially similar because they're cherrypicked to look that way.Yes. And that's not a problem with this study. Even if they had taken the photos specifically for this study while explicitly selecting lighting and focus that matches the AI "style," that would be a feature and not a bug.They wanted to test if people could pick out AI generated faces from real ones while ignoring all other details about the image.>Like what? Name themLighting. Background artifacts. Focus. These are not things the researchers were testing for in this study and are therefore confounding variables.
>>170546901. I always change my pitch/tone and accent/language randomly, mid sentence, when talking on the phone to trick the AI voice training2. My 4chan account is protected by a 32 word(yes word, not alphanumeric) password and I enabled 2FA in the settings
>>17054700>>17054700>that's not a problem with this study.Sure, because it's not actually studying anything useful. It's studying how well people can distinguish the least real portraits from the best-case-scenario gens, with a predetermined result: they can't.>They wanted to test if people could pick out AI generated faces from real onesIf so, they failed spectacularly.>Lighting.What stops you from controlling for this? Oh, right: your lack of understanding of what a confounding variable is.>Focus.I explicitly mentioned controlling for this.>Background artifacts.But that's not a confounding variable for my superior study.
>>17054628It’s over
>>17054716You are either incredibly unlucky, dumb, or (most likely) a "super AI-detector" that intentionally chose the wrong answer.
>>17054715The things that make the real photos "unreal" have nothing to do with the subjects' facial features now, do they?It's a well understood finding in psychology that people are uniquely good at recognizing faces in particular. That's the thing that's supposed to make this study interesting. It focuses explicitly on one ability humans are thought to be quite remarkable at. By allowing other variables like background artifacts influence people's selection, this would just be another replication, among many, of the standard "can you tell AI images apart from real ones in general?" studies.Here they wanted to focus on just the faces. If it were the faces themselves that acted as a giveaway for AI generation, then it wouldn't matter that everything else about the image looks uncanny.
>>17054681>controlling variableshuh. doesn't this mean the scientific claim loses generality or is merely forcing outcomes?Also the faces they chose are ridiculous. ABBABABABABBBABABAAALaugh at their picks.
>>17054716Luddites BTFO!!!
>>17054722It "loses generality" in the sense that all the other factors may still impact a final verdict in the real world. This is not "forcing an outcome" since the actual variable being tested for may or may not have the theorized impact. That's what they were trying to find out.
>>17054721Imagine the following situation: an even bigger retard than you tries to "control" for lighting by cherrypicking photos so dim that the faces are barely visible. The lighting is roughly the same across all photos, i.e. controlled for by your logic. The subjects do no better than chance. The Dumbest Scientist in the World then uses this result to conclude something about "facial recognition". In reality, the lighting predicts the entire outcome of the experiment. You're a textbook case of not understanding conditional probability.
>>17054730The problem with your argument is the dim lighting would serve to hide the facial features being tested for. If your ability to detect AI vs real faces were so good that we could tell the difference even in such conditions, then that would be a remarkable finding and would say a lot about human facial recognition. Likewise, the null hypothesis would indicate our facial recognition abilities aren't quite that good. Which is less remarkable but potentially noteworthy.In this case, the facial features are clear as day. So it is a fair test of our facial recognition abilities. Relying on background noise or lighting cues exploits a distinct cognitive function entirely.
>>17054734>>17054734>The problem with your argument is the dim lighting would serve to hide the facial features being tested for.The problem is with your capacity for abstract thought. I was illustrating for you that you can technically eliminate a confounding variable within the context of your shitty experiment, so that it wouldn't predict the subjects' individual answers, meanwhile causing it to predict the outcome of the entire experiment. So it remains an uncontrolled confounding variable, just one level up. The same goes for the chosen photo style in the OP: cherrypick a different one (e.g. high-detail with harsh, unflattering lighting) and you can change the outcome.>Relying on background noise or lighting cues exploits a distinct cognitive function entirely.Yes, it's asking a different question altogether because yours is fundamentally braindamaged. You don't understand facial recognition: it's pretty accommodating as to what counts as a human face at all. It's only incredibly picky when it comes to recognizing a particular face. Facial recognition is not how you would expect someone to recognize a genned face unless it's literally deformed. It's now down to details that are probably handled by a more general faculty and by intentionally choosing fake-looking "real" photos, you've done your best to hinder it.You will never be a scientist.
>>17054734You know what? Here's a simple way to put your mistake: you're conflating the question of what looks like a proper human face with what looks like real life. None of your portraits look real. The reality of a face is in the details photographers struggle to eliminate and image generators struggle to replicate. The experiment in the OP is rigged in favor of unreality. It's asking people to distinguish real-unreal from unreal-unreal. It's nonsense.
>>17054758Because you're clearly too retarded to understand the argument I'm actually making here, let me spell out a specific use case for a study like this:Let's say you're an AI researcher and you're trying to figure out what, specifically, needs to be improved upon to make the images more realistic. Ofc, he wants to know where to focus his efforts.There's already tons of research showing that people can distinguish between AI and real photos so another study saying the same thing is useless to you. Finding specific circumstances where people have a harder time telling the difference can inform you what you don't need to focus on.A study like this indicates that the faces themselves are not a crucial variable. That is genuinely valuable information to know. Identifying conditions where the machine works is important when you're trying to figure out why it's breaking.
>>17054765>you're conflating the question of what looks like a proper human face with what looks like real life.You're the one making that conflation. I made no such claim.The question is whether people could distinguish AI from genuine human photographs. Nothing was said about "real life."
>>17054724Their pictures are entirely bad faith. May as well ask people to identify real burn victims from AI generated ones. How many times have you gazed on a black poojeet or chink?
>>17054770>>17054770>Let's say you're an AI researcher and you're trying to figure out what, specifically, needs to be improved upon to make the images more realistic. Ofc, he wants to know where to focus his efforts.I take it diversity quotas are still very much a thing in your university, huh? Because you're literally 80-IQ. If you wanted to improve your face-generating AI product, you'd need to home in on what makes faces look fake. A study where people can't tell the generated faces apart from the real ones gives you the least amount of information about it. Now, that may be because your product is perfect and such information just doesn't exist. OR! It could be because the researchers' attempts at controlling confounders all boil down to sampling bias as far as you're concerned. Maybe the sample is biased in favor of your products' strengths, which doesn't give you the information you need to improve. You know what you actually want, as this hypothetical AI researcher? A study showing people faces under a wide variety of conditions. If you could get your hands on the raw data about what faces got sussed out, then you could classify them based on exactly the kind of variables a certain brainlet wants to eliminate and look for correlations. That way you could figure out, specifically, what needs to be improved upon.You're so bad at this it's unreal. I hope your poor parents aren't the ones paying for your "education".
>>17054773You're clearly very brown and I'm bored with you.
>>17054791>A study showing people faces under a wide variety of conditions. If you could get your hands on the raw data about what faces got sussed out, then you could classify them based on exactly the kind of variables a certain brainlet wants to eliminate and look for correlations.You're literally advocating for HARKing right now...An exploratory study like that is valid for hypothesis generation. But it has two problems:1. The sample would need to be massive to distinguish any meaningful correlations from noise.2. Any hypothesis you do generate from such a study would still need to be validated by a more targeted study like the one we're talking about right now.I'll try my best to square the circle in your defective brain and say whether this study would be worthy of publication depends largely on how their data is presented in the final paper.If they claimed your strawman version that the results indicate the AI is largely indistinguishable from reality in present form, then it should get thrown back for corrections. If they accurately describe the post processing done on the real photos and limited their presented findings to the specific circumstances in which they were achieved, then it is a perfectly valid contribution.
>>17054837>You're literally advocating for HARKing right now...No, I'm just advocating for your family to be deported back to Bangladesh or whatever hellhole you came from. Not reading the rest.
>>17054846Maybe you should read the rest. Because then you might learn a thing or two about how science is actually conducted instead of LARPing in your mom's basement.
>>17054628
>>17054847>Maybe you should read the restNo. You're mentally inadequate and there's no point reasoning with you - that much is obvious from our exchange so far. FWIW, their page gives this LLM blurb of a rationale:>AI-generated faces are not just an aesthetic curiosity. A synthetic profile photo can be paired with AI-written biographies, messages, and document-style layouts to create a fully convincing false identity — one that is easy to scale and hard to detect.i.e. they're not that interested in the general question. They're trying to prove that social media retards can't tell apart Stacy's digitally manipulated profile pic from a straight up fabrication. Does cherrypicking artificial-looking "real" portraits undermine that purpose? Maybe not. And that's the closest thing to a win you're gonna get here (although it makes your entire framing look retarded).
>>17054861You have literally conceded my core argument and you'd understand that if you had at least the reading comprehension of a high schooler.>it makes your entire framing look retardedMy "framing" was a specific example of a use case. I did not say that was the reason the study was conducted and you are retarded.
>>17054868>You have literally conceded my core argumentYou are literally hallucinating.>My "framing" was a specific example of a use caseYour specific example of a use case was for AI researchers to improve face generation. That was your actual example: using data from a study where people can't tell the difference to iron out cases where people can tell the difference. You are that stupid.
>>17054868And I didn't mean your pants-on-head retarded "use case" when I said "framing", I meant your treating it like it's trying to settle the general question. Zero reading comprehension. You're legitimately stupid. Accept it and move on.
>>17054871>Your specific example of a use case was for AI researchers to improve face generationWrong. Work on your reading comprehension. Your misreading of my argument might sound trivial to you, but it's a very important distinction.
kek i speed ran it too
>>17054873>Wrong. Work on your reading comprehension.>>17054770>let me spell out a specific use case for a study like this:>Let's say you're an AI researcher and you're trying to figure out what, specifically, needs to be improved upon to make the images more realisticYou're either having a psychotic episode or you're a spambot running out of context window. Moving on from this 80 IQ, irredeemable brown thread.
>>17054877>what, specifically, needs to be improved upon to make the images more realisticThis is distinct from this:>AI researchers to improve face generationIf you understood the context in which this data is valuable, as I explicitly laid out, you'd understand that they aren't the same thing.
>>17054707Your 4chan account is tied to your browser fingerprint
>>17054628not too bad, I kind of studied the faces. Too much symmetry is AI for the most part, me thinks
>>17054877>Why yes, I am an arrogant asshole who greatly over estimates my own intelligence and I refuse to take my meds.The thread shitter is back boys! With more retardation and even more ignorance! What will the voices in his head say next? Stay tuned!
>>17054628It used to be obvious, but these days it's getting harder and harder. In part because we are all so used to makeup and photo-shopped images and AI can take advantage of that to disguise uncanny valley effect.
>>17054628I can't even get past the captcha, just an endless loop of "please try again".But then again I wouldn't expect any better from a retarded shithole like australia.
>>17055043>he can't pass the internally visualized captchaKek. Brainlet.
>>17055031>In part because we are all so used to makeup and photo-shopped images and AI can take advantage of that to disguise uncanny valley effect.Yep. But good luck explaining this basic point to some of the complete morons ITT.
