can a man with 10,000+ hours on street view tell the difference between what is real and what is ai? good video: • AI-Generated Fakes: Ho... edited by: vidmok
At 7:50 he was talking about the 2nd image and at 8:04 he was talking about the 4th image. Thats why theres always a red arrow that shows what picture hes talking about
@@FaeTheMf poles are used constantly to determine where you are. A lot of places look very, very similar so you see a certain pole you might be able to eliminate somewhere or know which country, or province you're in by that alone or with other clues
@@petrkdn8224 I'd assume that it because foliage being correct relies on the complex 3D structure of the trees being correct, that with branches often being obscured, hidden by bloom, blending with background or getting messed up by image compression due to them being too small. And sometimes you need to get details which are not represented by anything in the image, like wind direction right That and AI does fine texture worse when it's not the focus.
@@petrkdn8224 Yeah any sort of pattern generally throws the AI off. With trees and leaves however, you just need to ask yourself; what’s the logic being that branch and where is it coming from? It usually ends up looking like how fever dreams feel. In this case tho, I couldn’t see anything through my phone screen. Couldn’t see the details in a thumbnail resolution.
Yea I agree with previous replies. AI is terrible at leaves and other coordinated or semi coordinated structures (like candles, fires, lights and shadows) Branches out in the open wind will all lean towards the same direction under the effect of the sun and the wind. Individual leaves will always face the direction that gives them most sunlight, top leaves usually face east while bottom leaves face any opening to the sky they can find. A pile of big leafed plants is always enough to thwart any AI.
I found the difference in how a geoguesser player vs a "normal" person approaches this very interesting. You used mostly your game knowledge to find the right images, while I searched for misstakes in the details like wrong shadows etc.
for me, the difficult part about looking for shadows and distortions is that its hard to tell if its due to the camera/blurring or if its an ai mistake
Personally I could tell because it felt off and looking a bit longer I noticed some details thats seemed weird But crazy that this gut feeling can tell that something is off
yeah for example in the first quiz.. number 3 looks totally off for me.. it looks like they put together bushy greens, too clean accentuated road with too maintained side of the road for that part of the world, and the background sky pops way too much.. but for him it "made sense" somehow..
I think the craziest part about them is if nobody told you that any were AI generated, aside from the obvious "3"s, is that they fool just about anyone.
I went 5/5 but not for the reason of knowing its AI like him I just found distinctive color grading differences between real cameras and FOV it feels like art more than an actual picture its scary how accurate some were tho.
@@ABC-lm9do Yeah, that's another scary thing. When you have to be an expert in a relevant field to have much of a chance when you *know* something AI is present, pretty much no one will have any hope spotting these things in the wild. And it's only gonna get better, with video and eventually synced audio (by eventually I mean like 2 or 3 years tops) that to most people will appear completely seamless.
I went 5/5, I'm not a geoguessr pro, I'm not an expert in any relative field at all, AI just looks uncanny to any human being with functioning eyes(no, I wasn't just guessing)
This kinda reminded me of that one game, exit 8 I believe. Where mind starts playing tricks on you and just start overthinking way too much so that even the most normal things start feeling like they dont belong there
@@lpharmer3496 Thing is, the human brain has vastly more neurons than our image generating AI right now. Them having fewer neurons fundamentally limits their maximum performance even if you train them in a loop like that, i.e. they just don't have the brainpower. Normal adversarial networks like you describe are equal in network size and so they eventually are able to fool the discriminator (the other AI checking if the output is real or fake), but if the discriminator (a human in this case) is much bigger, it may not ever get higher than a certain success rate. True human/superhuman performance might be a long way off (until we can make much bigger neural networks).
Hey editor, I got a pro tip: If you have information on screen like at 3:17, a good rule of thumb is to have it there long enough that you personally can read it twice. I get that it would likely mess with the pacing if you're dedicating time to it, but having it on screen for not even 2 seconds is very easy to outright miss for casual viewers like myself.
I doubt this was the intention of the editor, but flashes of text probably help the algorithm because viewers either need to pause or rewind to read it, which increases the watch time per click.
@@shmooveyea It was on for such a short time that I literally couldn't even pause it in time. If I have to rewind just to not miss kind of important information, it sucks :/
I actually knew this one! If you do rock climbing in UK, it's pretty easy. And something about that rock by the parking also screams UK to me idk why lol but yeah, does not look like UK to a normal person, sky too blue lmao
@@Pavel-yp2je Can you confirm my theory that rock is there to stop cars from pulling out at an angle and prevent crashes with cars coming around the corner?
The last one had some clear giveaways if you spotted them. The trees on the center-left have gaps in them, they are just floating. Also the trees on the right are super squiggly
what bothers me these days with all the AI stuff is those arrogant people who surged who keep telling how "extremely easy" it is to tell real from AI, especially in drawings like No bruh, that's not easy, it's becoming very hard for the common folks to differentiate certain stuff. And then people come with all the technical knowledge about design/art/photography and I be like bro do u know most people don't have said knowledge right lol
I'll be honest I'm surprised you missed the last one; if you look at the center-right of the trees, there's clear artifacting from panoramic/stitched imaging just on the border of it. I don't think AI would have mimicked that sort of artifact.
@@colecube8251top left pic on the last challenge, above where the road ends you can see the sky between the trees on the left and right side of the road there is a horizontal line
@@colecube8251 Look at the top-left image, dead center but on the right edge - there's like a "bone" shape in a branch. Directly to its side, there's an identical copy of that bone shape. If you look closelier, you'll notice that there's an entire smear of clone/copy of that whole area, which is pretty common in panoramic imaging when it's stitching images together. Essentially the whole right edge (like the last 30 columns or so) are just copied from directly to the left of it.
I think with a few small changes you could make this a lot harder. If you limited it to one country (Brazil) then the AI would more consistently generate realistic images. Also if you had a pro work with the person generating them to select the best images you could filter out some of the easier tells of these images.
AI imagery is definitely harder to identify when there's a lot of visual noise and few actual landmarks to compare. This test is one of the first times I've actually had to struggle to tell it apart, even got two of them wrong
@@luka188it's full potential is based on what information we use to train it, most ai already are trained billions if not trillions of input, I doubt we can get more than that tbh, for now the big difference is how well you can train the ai and how much of its potential we can get out
@@luka188it hasn't been for "a few years", it has been literally decades of research and development at this point. It's been only "a few years" since AI image generation got commercially viable, which for many products is close to the final form (though not for AI of course), but still it definitely isn't "bottom".
The scary part will be that when someone is malicious with AI, it won't be in a challenge where you know it's present, and there won't be a reveal to show that you were right and wrong. It will all slowly seed itself unannounced into our images, art, libraries, and legal evidence.
Just want to say this is an awesome concept for a video, would watch more of these for sure. It's a novel idea to have you guessing something other than location. I think the viewer learns things about the thought process of a GeoGuessr that likely wouldn't come up in a normal video.
ngl, you might benefit from an eye tracker. You could just show what you're looking at when you talk about stuff instead of having to explain or edit in the red arrow.
It's really impressive how accurate Rainbolt was, but like he said it's not easy. He's spending a long time analyzing each image and goes in with the knowledge that one of them is fake, the average person is not going to spend more than a minute looking at photos while scrutinizing every detail.
This is so crazy. Once rainbolt pointed out certain features, it made sense, but initially just trying to find things myself I would basically suspect every single image equally. You could definitely make a game where it gets 3 screenshots randomly from street view and generates one ai streetview photo, and you just have to choose which is fake until you lose. It would probably be really easy for experts as unlike in this video, there wouldn't be an ai expert to verify and refine the image, but for everyday people that sounds super cool. In fact now I wish i had the knowledge to host that as an actual site, maybe after i graduate I'll try and code it up but i wouldn't know how to host it without costing me a fortune
The thing that sold me for 1 being real on the last round was because on the right if you look at the edge of the image at the bush you will see a duplication of the bush, which I doubt AI would replicate.
For the first one I noticed that the clouds appeared to be covering the sky which should make the shadows overcast yet the trees at the far point in the road had hard shadows which you would not get on an overcast day.
My method woulda been to zoom WAY in and see which of the noise patterns are weird. I'm not sure whether this method works for all AI image generators, but for many its often quite obvious.
I tried to play along, here were my results (honestly I "locked down" my answer before he revealed his thought process): 1. genuinely didn't know, all looked very good. Though once he pointed out the difference in grass on both sides of the road for pic 2 I saw it. 2. This one was fairly easy actually. Pic 4: The wires and the telephone poles don't match up, and the road has the smoothing effect, a typical fabric of AI art. 3. Like rainbolt says, fairly obvious 3. Smoothing effect, road width inconsistent, weird lines, pavement on the left abruptly stops. 4. Pic 3 was also semi-easy - the wires on the telephone pole didn't continue and the picture is weirdly slanted. I didn't catch Pic 2 though! So 1/2. 5. I was struggling a lot, but ultimately went for 1. Towards the end of the road the greenery "blends" into the road, again a typical artefact of AI art, where distant objects start randomly melting into each other. Also the shadow of the coniferous tree on the left wasn't reflected on the road surface. 4/6. Decent, but defo room for improvement. (the fact that I only got one point below the geoguessr genius is already a giant validation tho lmao)
I think one giveaway for all of them was the incorrect exposure of the sky. especially the last 2 rounds, the imposter images felt like the sky should be bright white to match the exposure of the road/foliage
At 3:20 the thing that gave #4 away for me was the shadow of the snowbank on the right. There is a little dip in the middle of the shadow closest to the camera but there is no dip in the actual snowbank to create it.
for set 4, image top-left: why are there repetitions in the lower-right area of the foliage? there are clearly some branches and leaves which are copy-pasted. are those street view image stitching artefacts?
I actually got 4/5, only getting the one with two wrong. But if I didn't know they were AI then maybe the only one I would've thought was off was number 3
anybody who played with stylegan (including myself). Stylegan2 is about 4 years old by now and can generate faces that are extremely realistic (examples are on thispersondoesnotexist), but the predecessor which is about 5 or 6 years old was pretty good as well.
I think the giveaway in the last set is the leaves in option 4 - All the leaves piled densely on the left, but the trees that had those leaves are on the right - How do the trees close to the camera have no leaves, but randomly a tree or two still have leaves and *all* of the trees in the distance have full dense leaves still?
The odd thing about Nr.3 at 6:30 is how thin the road gets. I think the AI tried to convey the feel of it disappearing in the distance, but with the rocks around, the proportions are just off
In general, I would have struggled to spot these, but when you maybe the 4th image in that last round black and white, I suddenly noticed this weird tagent of shadow that seemed to "continue" the road up into the sky, which I didn't notice in color at ALL. that one where the Trees on the Right gave it away for you too - the thing I noticed there was how the upper branches all formed a line in a way that looked like a climbing vine hanging on wire but without any poles going on.
