If you look at the left eye of the "lion" on the left, it looks quite odd. That's why I assumed it was fake. However I did initially think the one on the right was fake because of its awkward pose/expression.

@bennnotben7291

Жыл бұрын

Your reasoning for the right one looking fake is exactly why I thought it was real. I thought it was an odd pose that still felt real, something I imagined AI having difficulty with. On the other hand, the left lion looked very symmetrical which you'd think it could handle easily.

@TreesPlease42

Жыл бұрын

The paw gave it away for me. There are just too many variables to look for depending on the image, given the speed that people scroll.

@YouReyKarr

Жыл бұрын

You can also see the terrain stitched together behind the lion because the AI doesn't know what stuff looks like behind objects, it just starts the ground again at some arbitrary height that seems reasonable.

@kingbling7571

Жыл бұрын

the first lion only has one paw while the second has 2 so it is clear as day that the left one is fake lol

The fact she was able to correctly identity about 80% of the images on a dataset curated by a human to deliberately be as difficult as possible is, I think, a pretty clear rejection of the idea that AI pictures are indistinguishable from human-made ones, at least right now.

@Jehty_

Жыл бұрын

Maybe not indistinguishable, but definitely good enough to fool people in real life scenario. When you see a picture in a newspaper article, or Twitter how much time do you actually spend looking at it?

@user-on6uf6om7s

Жыл бұрын

AI images as a whole? Definitely not, there were certainly some complete abominations that got thrown out that would have fooled no one but you can generate a thousand, potentially many thousands in an hour depending on your settings and rig and if just one of those is equal or preferable to the human alternative, a profit-driven business will choose the AI every time. You just need the human to be the filter, at least for now.

This game is so awesome. On my first try I got 92% animals, 92% animals, 67% photography. The only complaint I have is sometimes you get the same image twice in two different games, so it skews your score because you already know the answer.

also in her defense, this wasn't _really_ a human vs AI. It was a human vs AI and another human filtering out all the goofy mistakes the AI made.

@DreadKyller

Жыл бұрын

While I understand you point, that's really not the point. Those who deal in AI generated images are already generating many images and looking for one in a hundred good ones. The question isn't whether AI can always generate images that may fool someone, it's whether they can in general. This was comparing real images and art to the actual use-case people would use AI art for, people aren't just sharing every image the AI generates. This is a test of whether someone would be able to identify actual versus AI images on the internet, and the majority of the time the AI images uploaded are filtered, specifically selected samples.

@Lomky

17 күн бұрын

it's a nuanced but valid distinction. the question isn't "can humans spot (the average) AI generated image" it's "can humans spot a quality AI generated image". Both are valuable!

You could upload once a day or once a decade and I would still be equally excited for each video and call myself a diehard Jabrils fan

@Jabrils

Жыл бұрын

i appreciate you JMV.

"Web Dev is what God used to build Hell" bro I feel you

@Jabrils

Жыл бұрын

😔

The cope can and the taut definition are the details im here for

You could use the data from this site to get a list of images people are really likely to classify as real, when they're not. You could then probably train a classifier conv net (a descriminator) on that, to remove future images that would be "to easy" from the database.

@jannikheidemann3805

Жыл бұрын

Exactly what I thought too. I mean, an adversarial approach against real users is not that far fetched realy.

@kingbling7571

Жыл бұрын

Indeed, you could use an antagonistic approach to this and get the same result in the end, not to difficult with the an ann

When the world needed him most, he returned...

Make an A.I that can detect if an image is A.I generated or not.

I decided to play around with the Sus AI before watching the rest of the video and seeing the results and I'm actually quite surprised with my results. I expected the art category to be difficult while the photography and animals would be easy but I found it to be the exact opposite. The AI generated animals and photography can sometimes be incredibly obvious but other times even after I was told which ones were AI i still couldn't find anything that really gave it away. I found that the art was much easier once I started looking at the fine details and asking myself "Why would the artist do this?". If I couldn't answer that question it was usually fake. I averaged about 65% on both animals and pictures with around 90% on art. Great video, and things are definitely weird now.

@user-on6uf6om7s

Жыл бұрын

That's weird, I ask myself that question all the time when it comes to erotic fan art and as far as I know, most of that is made by humans.

Unfortunately, you are a man of integrity. If I were in your place, I'd tell her the right picture was the fake one and asked her to explain how she can tell it was fake.

I got the lion at the start correct due to the slightly warped eye. But then i tried your game and got fooled by a lion with slightly weird eyes. I'm pretty hopeless outside of slightly different eyes and hands, I have been fooled.

I agree with your friend on the "have never worked with a client before". You show a client an AI generated image, they'll return with something like "I like it, but could you move to the left more and change the color of . Also I feel the lighting in the scene is a too soft, could you make it a bit harder lighting and maybe add some accent lighting from the other side?" Etc. While Inpainting may help some with this, generally at this point in time making these types of changes without changing much of anything else is very, very difficult if not impossible. Given enough time I'm sure it'll become more feasible, but at the moment it really can't replace an actual artist, doesn't mean it's not a bit terrifying. I think the main groups that will be effected the most by this will be stock photography and concept art, as both only need to really be decent and not very exacting and precise.

Nice game, also my guess is one ai voice was used.

The mans is back! So close to 500k.

thanks for the story telling me you uploaded, 90 percent sure youtube was hiding your uploads from my feed

Awesome experiment, you brought the right person in to test it!

@Jabrils

Жыл бұрын

Sabrina is too OP

@thisistherun4015

Жыл бұрын

@@Jabrils definitely thought this was a new Answer and Progress video and was pleasantly surprised with a new creator

Jabrils, I swear I love everything about you and what you do - hell you even busted out a sneaky DKC Boss theme in there, this is as good as my parasocial relationships get. I just think you’re dead wrong about how important it is to be skeptical. Not just of AI, but of the idea that the end game of it is to trick us. Stable Diffusion & friends are great at mashing visual ‘signs’ together and spitting out believable imagery, but see in visual art specifically (and I’m including photography) “believability” doesn’t really matter as much as originality. And while you may think that such AIs that are good at mashing things together into somewhat eye-catching and intricate results, the wide availability of the technology has now only opened the flood gates to a lot of similar looking images that quickly get boring because they just use the same technique: mashing signs together. Artists are not worried, they can always master the tools, subvert them, or go where AIs can’t. On the plus side, these AIs are probably gonna help with visual literacy. That’s also greatly beneficial to artists :)

@user-on6uf6om7s

Жыл бұрын

An AI can't give you taste or an original concept, it can just execute what you ask for to the best of its abilities. The lack of interesting AI work out there isn't primarily the fault of the AI's inherent inability, it just can't stop you from wanting it to make thousands of doe-eyed waifus which is what 90% of the AI crowd seems to gravitate towards.

In general I wouldn’t say that AI is smarter than people yet but it’s definitely better at bullshitting

Very interesting! I spotted the weird eye on the left lion right away, but that sloth one still looks real to me.

tbh, the right eye of the lion on the left is a dead giveaway. The background trees also have quite undefined normals in comparison to the right one

I love your channel! What a fun take on these new generative models.

@Jabrils

Жыл бұрын

thanks mate :D

That was a very dark endcard.

@Jabrils

Жыл бұрын

i fixed it, so now only those who know, know. lmao

Oh no. I can see it already. AI generated CAPTCHA.

I got around 50% correct, so definitely fooled me.

🤣 the domain name 👍🏿

I like what she said. Those are not art, those are visuals. and yes, for someone that studied art, it's easy to spot (i'm a visual artist so yeah, the first one of the drawing was very obvious).

Elon musk also said you can take delivery of your electric truck in 2019...

I heard the fake in the Elon clip but then I forgot about the spot the ai voice game. Guess I gotta watch it again?

How do you know this is really Jabrils? It could be an AI pretending to be Jabrils telling you to play the game so that the AI can improve itself. Also me: *picks the wrong lion*

Please do a follow up video on how everyone does! And let us know which photos trick people the most. Maybe do a version 2 where you take out the ones that never fooled people and put more in, then do a version 3 where you do the same thing again.

She actually lost because she WANTED the sloth in a beanie to be real, not because she thought so. That's so human right there.

Hope you do an update as to which images tripped people up the most, and the overall results?! It'd be so awesome. I got 92% on animals lol But I failed the rest (50/60 something) And yes, I've spent many a years working with and drawing animals, so, that tracks

please create a 3D maze multiplayer teleport game in which is a race to get the most points from collecting various gems which the lowest of value most will respawn right after 20 seconds whereas the highest of value will respawn right after a minute. What makes the game an awesome challenge is figuring out what it the best route not already taken and how to ruin the routes of the other players.

That’s cool but also not why people are pissed off with ai at all. The problem‘s that the data comes from real artists who never agreed to have said data taken, and people are screenshooting all the real art posted by specific people, feeding it into some free generator, and then *charging money* for whatever comes out, even though literally none of the work involved needed any effort by them. It’s like putting frozen pizza in an oven and then selling it.

When the set of AI generated images was produced, was there any bias introduced by the human in charge? Is this a human-filtered set, containing only the best AI images? Or were all the truly awful AI images included, too? Let an AI choose which AI generated images to use, then see how much easier or harder the game becomes.

Machine learning,image generation companies, accurately being sued for copyright infringement, there are multiple lawsuits.

@hippityhop9522

Жыл бұрын

If AI isn't art then half of the modern art museum should be closed because nothing in it is art. You could even expand it to digital art because artist doesn't spend the same amount of time and effort as a actual painter/artist who creates in physical space.

@Jehty_

Жыл бұрын

@@hippityhop9522 digital artists not spending the same amount of time and effort? What tf are you talking about? Please watch some videos about digital artists creating art. And then think about what you just wrote...

@me-ry9ee

Жыл бұрын

@@hippityhop9522 no its just the dataset they are using have copyrighted images of artists and photographers taken without their permission and the companies are twisting the narrative on how the compressed dataset is a transformative act and comes under fair use.

Deleted my KZread account to get new and better recommendations and it was so hard to find your channel! You originally got me started and motivated to program and it was tough to find your channel comparison to the past. Keep it up!

This is all fun and good, but most the actual criticisms over AI generated images are not about how effective it is, but about the legality and ethics of how training data is collected and used.

I totally can’t tell the differences between ai generated vs real pic

havent been recommended Jabrils in over 2 years

Tried out the website and man, I could not tell the difference. I only got 67% in photography and animals but only 47% in art. Huge props to Sabrina for doing so well in that category

The paw and chin of the lion in the left picture gave it away for me. Really unnatural border

"1" AI Voice used

13:23 That deer looks fake as hell with its weird antlers.

The giveaway on the lion was the dirt. The left was just a non-descript dirt 'concept'.

the cope in a can line killed me

1:47 my answer is correct because the fur of that lion is too soft

Interesting video! Thanks

Sound is much better with the dead cat on your toilet brush microphone .

4 AI generated voices

the first lion only has one paw while the second has 2 so it is clear as day that the left one is fake lol

Plot twist: this whole video was created by AI

Honestly after using these programs for a couple months now, it's easy to see the one on the left. I picked it in less then 20 seconds. The eye on the Lion gave it away, I am only 2 mins into your video, so maybe when I watch the rest you will reveal this. But to me that is always how you spot it. AI isn't capable yet of holistically seeing when it has made a mistake. Sometimes it is an extra finger, a leg when there are many people, or in this case, an eye that clearly doesn't match the other eye. In this case sure, it's possible there is a lion with messed up eyes out there, but unlikely, so it was pretty easy to spot the left image. Having said that, it won't be long before this isn't the case, and 2, you can fix these easier and easier now with the new integration into paint programs, etc that allow the artists to paint our regions and ask the algorithm using diffusion techniques to redraw the eye in this case, and it would do so often correcting it's original mistake. Not again because it spotted like humans can, but because it tries to match what is already in the image...

With enough looking at AI art you start to notice things it gets wrong, which, in my experience, are often small things like the way the dirty ground is laid out, or the arrangement or specific patterns in the trees. The left lion was very well put together, and I thought it real at first, but upon looking at the path on the ground, it looked to have a very unusual path for what I'd expect, and a few washed sections. That, in comparison to the clean background of the second image, is what made me think it was the AI generated piece. The AI piece is truly incredible, though. I would've *never* thought it fake without the primer to search for discrepancies, and I barely guessed it to be fake even with the primer (was a really unconfident decision due to the similarity in quality between both pictures). Job displacement sucks, I hope artists find a way to work around the oncoming changes in the industry. I imagine a restructuring in the programming industry will come from AI soon too, with AI generated code vetted by knowledgeable enough people to spot mistakes being used in place of purely low skill programmers in many places. Seems like a massive aid in many places, though the restructuring "growing pains" are always hard to go through as people in the given industries.

Where are you throwing your Chao?!

@Jabrils

Жыл бұрын

( ͡° ͜ʖ ͡°)

Okay, paused at 1:25 and full screened. YT's hyper aggressive compression lately doesn't help. But I'm going to guess the one on the right. Not because of the Lion though, mainly because the clouds just seem... fanciful? maybe idyllic is a better word? The one on the left has a blown out sky, which seems more like what I would expect to see with such a dark and contrasty foreground. But at the same time, the vegetation on the ground in the one on the right is sort of the kind of thing that AI typically has trouble with getting perfectly right...

la-node, that sounds fancy.

How about a third option for Photoshop?

It's 7 months later and AI art is already about 2-3x as good as it was when he made this

You can tell the left lion is fake cos the eyes are a bit weird.

The problem with this test is that it is not really about AI. Because a human chose all the input images it says nothing about the reliability of AI making realistic images, since you discarded all the unrealistic images. You still need a human intelligence to fool the other humans.

@user-on6uf6om7s

Жыл бұрын

Sure, but I don't think a single discerning eye being able to generate thousands of images and only having to do the job of curating the best outputs is going to be much consolation to the concerned artists. You'll still need humans (at least until the AI develops a discerning eye) just far, far fewer than when they actually had to make the images.

Did I hear my name?

1:20 assuming this is not a trap and not both images are of the same type i would say that… the image on the left is a true picture with a wide angle camera, and the picture on the right is generated, multiple things with it are weird, not least the "paw" on the right. edit: RIP

If I play the game, I can't submit because my answer is always: I don't know.

9th grade kids are experiencing this Atari level AI right now. They'll be playing with a 4090 GPU PC level AI when they graduate.

1:18 On the left the floor is weird, as are, as so often, the eyes. I think the picture on the left is AI generated.

the lion on the left had a weird/off eye compared to the rest of the image plus something about the top of the mane and the nose looked off to me too I watch too much national geographic to not at least subconsciously know when a photo of a lion looks correct to me

Banger! 🤯 2:55

YESSS THE MAN IS BACK!

@Jabrils

Жыл бұрын

Let's get this focaccia 🍞

Ayo, what da lion doing?

@jannikheidemann3805

Жыл бұрын

Lying lying down.

AI art changed a lot in these 2months

Sabrina won in my heart.

I paid $5 for a YEAR of hosting. Not many people are using it though, so it hold up.

Listen, content: phenomenal; I love your shit, it’s always so great. Could watch it forever. However, are you ok? You haven’t been eating during your videos… are you being held captive? I can’t do the blink twice bc you’ve got that drip over your eyes but fr you’ve gotta be starving to death by now

I think the lien on the right is generated.

i 100% agree with sabrina right there

beautiful niche ending

@Jabrils

Жыл бұрын

finally, some acknowledgment 😌

The whole video is ai Ganarated voice 🤣

Guessing Left Lion is fake.

I gotta get that shitty bugs shirt haha

i'm not even getting this vid in my feed, it's weird

...Is this an evil Jabrils persona I see emerging. Maybe keep the wacky glasses (The black ones) if you decide to wear a supervillain costume.

Elon blowing something out of proportion? That's rare.

Interesting

I'd say you used more than 2 AI voices, but not more than 5. I chose the lion on the left because that's a dog's face structure. And I got some of the images in the game right because of the depth of field problem: without a 3D reference, AI doesn't know how to blur an image properly, so some things like clothing have either way too sharp focus or way too muddy/blurry, while the rest of the image seems in tune. That's what gave the fake squirrel away, for example. In other images, it's more of a knowledge of how physical objects get put together or how AI tends to draw some things using repetitive structures and fractal clustering (essentially using the JPEG algorithm and similar maths as a behavior for drawing new material, rather than a data handler - AI is generally not aware of compression artifacts as being errors in the presentation of the image). Also I caught you using that new eye position corrector, and it did freak me out just a little. Dunno about mouth replacement but I'm not actually trained in this, I read the comments! I'm just an old lady who's been on the Internet longer than Google has.

I miss when Jabrils didn't have to open his mouth to talk

The one on the left looks too obvious.

yes

Puget: pyoo-jet

why no https links tho ?

1 you'd have to look at a tonne of AI images and learn the mistakes AI makes. 2 if it just happens to be an artwork without any mistakes you're screwed. In that regard, many artists are rejecting a powerful tool. Because great artists know what things should look like, it would be a ridiculously powerful tool for them.

2 voices? i have no idea -_-

That feeling when you try the game too late and it's gone. Maybe you could open source it?

@magemerger2243

3 ай бұрын

Rest In Peace

6 AI Voices!

That's a dog

@tantamounted

Жыл бұрын

That's also why I chose the left lion.

The left lion's left pupil looks weird, so it's fake.

Hey ! Can you do a video on rainworld ?

The right one 1:25