Results of the "Can you tell which images are AI generated?" survey
Results of the "Can you tell which images are AI generated?" survey
Previous posts: https://programming.dev/post/3974121 and https://programming.dev/post/3974080
Original survey link: https://forms.gle/7Bu3Tyi5fufmY8Vc8
Thanks for all the answers, here are the results for the survey in case you were wondering how you did!
Edit: People working in CS or a related field have a 9.59 avg score while the people that aren’t have a 9.61 avg.
People that have used AI image generators before got a 9.70 avg, while people that haven’t have a 9.39 avg score.
Edit 2: The data has slightly changed! Over 1,000 people have submitted results since posting this image, check the dataset to see live results. Be aware that many people saw the image and comments before submitting, so they've gotten spoiled on some results, which may be leading to a higher average recently: https://docs.google.com/spreadsheets/d/1MkuZG2MiGj-77PGkuCAM3Btb1_Lb4TFEx8tTZKiOoYI
So if the average is roughly 10/20, that's about the same as responding randomly each time, does that mean humans are completely unable to distinguish AI images?
In theory, yes. In practice, not necessarily.
I found that the images were not very representative of typical AI art styles I've seen in the wild. So not only would that render preexisting learned queues incorrect, it could actually turn them into obstacles to guessing correctly pushing the score down lower than random guessing (especially if the images in this test are not randomly chosen, but are instead actively chosen to dissimulate typical AI images).
Maybe you didn’t recognize the AI images in the wild and assumed they were human made. It’s a survival bias; the bad AI pictures are easy to figure out, but we might be surrounded by them and would not even know.
Same as green screens in movies. It’s so prevalent we don’t see them, but we like to complain a lot about bad green screens. Every time you see a busy street there’s a 90+ % chance it’s a green screen. People just don’t recognize those.
If you look at the ratios of each picture, you’ll notice that there are roughly two categories: hard and easy pictures. Based on information like this, OP could fine tune a more comprehensive questionnaire to include some photos that are clearly in between. I think it would be interesting to use this data to figure out what could make a picture easy or hard to identify correctly.
My guess is that a picture is easy if it has fingers or logical structures such as text, railways, buildings etc. while illustrations and drawings could be harder to identify correctly. Also, some natural structures such as coral, leaves and rocks could be difficult to identify correctly. When an AI makes mistakes in those areas, humans won’t notice them very easily.
The number of easy and hard pictures was roughly equal, which brings the mean and median values close to 10/20. If you want to bring that value up or down, just change the number of hard to identify pictures.
It depends on if these were hand picked as the most convincing. If they were, this can’t be used a representative sample.
Personally, I’m not surprised. I thought a 3D dancing baby was real.
From this particular set, at least. Notice how people were better at guessing some particular images.
Stylized and painterly images seem particularly hard to differentiate.
One thing I'm not sure if it skews anything, but technically ai images are curated more than anything, you take a few prompts, throw it into a black box and spit out a couple, refine, throw it back in, and repeat. So I don't know if its fair to say people are getting fooled by ai generated images rather than ai curated, which I feel like is an important distinction, these images were chosen because they look realistic
Well, it does say "AI Generated", which is what they are.
All of the images in the survey were either generated by AI and then curated by humans, or they were generated by humans and then curated by humans.
I imagine that you could also train an AI to select which images to present to a group of test subjects. Then, you could do a survey that has AI generated images that were curated by an AI, and compare them to human generated images that were curated by an AI.
All of the images in the survey were either generated by AI and then curated by humans, or they were generated by humans and then curated by humans.
Unless they explained that to the participants, it defeats the point of the question.
When you ask if it's "artist or AI", you're implying there was no artist input in the latter.
The question should have been "Did the artist use generative AI tools in this work or did they not"?
But they were generated by AI. It's a fair definition
Technically you're right but the thing about AI image generators is that they make it really easy to mass-produce results. Each one I used in the survey took me only a few minutes, if that. Some images like the cat ones came out great in the first try. If someone wants to curate AI images, it takes little effort.
I think if you consider how people will use it in real life, where they would generate a bunch of images and then choose the one that looks best, this is a fair comparison. That being said, one advantage of this kind of survey is that it involves a lot of random, one-off images. Trying to create an entire gallery of images with a consistent style and no mistakes, or trying to generate something that follows a design spec is going to be much harder than generating a bunch of random images and asking whether or not they're AI.
I think getting a good image from the AI generators is akin to people putting in effort and refining their art rather than putting a bunch of shapes on the page and calling it done
Did you not check for a correlation between profession and accuracy of guesses?
I have. Disappointingly there isn't much difference, the people working in CS have a 9.59 avg while the people that aren't have a 9.61 avg.
There is a difference in people that have used AI gen before. People that have got a 9.70 avg, while people that haven't have a 9.39 avg score. I'll update the post to add this.
Sampling from Lemmy is going to severely skew the respondent population towards more technical people, even if their official profession is not technical.
I still don’t believe the avocado comic is one-shot AI-generated. Composited from multiple outputs, sure. But I have not once seen generative AI produce an image that includes properly rendered text like this.
Bing image creator uses the new DALL-E model which does hands and text pretty good.
generated this first try with the prompt a cartoon avocado holding a sign that says 'help me'
Image generation tech has gone crazy over the past year and a half or so. At the speed it's improving I wouldn't rule out the possibility.
Here's a paper from this year discussing text generation within images (it's very possible these methods aren't SOTA anymore -- that's how fast this field is moving): https://openaccess.thecvf.com/content/WACV2023/html/Rodriguez_OCR-VQGAN_Taming_Text-Within-Image_Generation_WACV_2023_paper.html
Yeah I'm sceptical too, what tool and prompt was used to produce this?
Its Dalle 3 its not that difficult to generate something like that using dalle 3 here's some shreks I generated as a showcase Shrek 1 inage
All of these are just generated nothing else
I found the avocado comic the easiest to tell, since the missing eyebrow was so insanely out of place.
Its not that difficult to generate something like that using dalle 3 here's some shreks I generated as a showcase Shrek 1 inage
All of these are just generated nothing else
Something I'd be interested in is restricting the "Are you in computer science?" question to AI related fields, rather than the whole of CS, which is about as broad a field as social science. Neural networks are a tiny sliver of a tiny sliver
Especially depending on the nation or district a person lives in, where CS can have even broader implications like everything from IT Support to Engineering.
I got a 17/20, which is awesome!
I’m angry because I could’ve gotten an 18/20 if I’d paid attention to the thispersondoesnotexists’ glasses, which in hindsight, are clearly all messed up.
I did guess that one human-created image was made by AI, “The End of the Journey”. I guessed that way because the horses had unspecific legs and no tails. And also, the back door of the cart they were pulling also looked funky. The sky looked weirdly detailed near the top of the image, and suddenly less detailed near the middle. And it had birds at the very corner of the image, which was weird. I did notice the cart has a step-up stool thing attached to the door, which is something an AI likely wouldn’t include. But I was unsure of that. In the end, I chose wrong.
It seems the best strategy really is to look at the image and ask two questions:
- what intricate details of this image are weird or strange?
- does this image have ideas indicate thought was put into them?
About the second bullet point, it was immediately clear to me the strawberry cat thing was human-made, because the waffle cone it was sitting in was shaped like a fish. That’s not really something an AI would understand is clever.
One the tomato and avocado one, the avocado was missing an eyebrow. And one of the leaves of the stem of the tomato didn’t connect correctly to the rest. Plus their shadows were identical and did not match the shadows they would’ve made had a human drawn them. If a human did the shadows, it would either be 2 perfect simplified circles, or include the avocado’s arm. The AI included the feet but not the arm. It was odd.
The anime sword guy’s armor suddenly diverged in style when compared to the left and right of the sword. It’s especially apparent in his skirt and the shoulder pads.
The sketch of the girl sitting on the bench also had a mistake: one of the back legs of the bench didn’t make sense. Her shoes were also very indistinct.
I’ve not had a lot of practice staring at AI images, so this result is cool!
does this image have ideas indicate thought was put into them?
I got fooled by the bright mountain one. I assumed it was just generic art vomit a la Kinkade
Yes, "the end of the journey' got me too. So AI-looking. It's this one:
https://www.deviantart.com/davidevart/art/The-End-Of-The-Journey-952592148
About the second bullet point, it was immediately clear to me the strawberry cat thing was human-made, because the waffle cone it was sitting in was shaped like a fish. That’s not really something an AI would understand is clever.
It’s a Taiyaki cone, something that already exists. Wouldn’t be too hard to get AI to replicate it, probably.
I personally thought the stuff hanging on the side was oddly placed and got fooled by it.
Wow, what a result. Slight right skew but almost normally distributed around the exact expected value for pure guessing.
Assuming there were 10 examples in each class anyway.
It would be really cool to follow up by giving some sort of training on how to tell, if indeed such training exists, then retest to see if people get better.
Imo, 3,17,18 were obviously AI imo (based on what I've seen from AI art generators in the past*). But whatever original art those are based on, I'd probably also flag as obviously AI. The rest I was basically guessing at random. Especially the sketches.
*I never used AI generators myself, but I've seen others do it on stream. Curious how many others like me are raising the average for the "people that haven't used AI image generators" before.
I was legitimately surprised by the man on a bench being human-made. His ankle is so thin! The woman in a bar/restaurant also surprised me because of her tiny finger.
One thing I'd be interested in is getting a self assessment from each person regarding how good they believe themselves to have been at picking out the fakes.
I already see online comments constantly claiming that they can "totally tell" when an image is AI or a comment was chatGPT, but I suspect that confirmation bias plays a big part than most people suspect in how much they trust a source (the classic "if I agree with it, it's true, if I don't, then it's a bot/shill/idiot")
Right? A self-assessed skill which is never tested is a funny thing anyways. It boils down to "I believe I'm good at it because I believe my belief is correct". Which in itself is shady, but then there are also incentives that people rather believe to be good, and those who don't probably rather don't speak up that much. Personally, I believe people lack the competence to make statements like these with any significant meaning.
And this is why AI detector software is probably impossible.
Just about everything we make computers do is something we're also capable of; slower, yes, and probably less accurately or with some other downside, but we can do it. We at least know how. We can't program software or train neutral networks to do something that we have no idea how to do.
If this problem is ever solved, it's probably going to require a whole new form of software engineering.
I don't know... My computer can do crazy math like 13+64 and other impossible calculations like that.
And this is why AI detector software is probably impossible.
What exactly is "this"?
Just about everything we make computers do is something we’re also capable of; slower, yes, and probably less accurately or with some other downside, but we can do it. We at least know how.
There are things computers can do better than humans, like memorizing, or precision (also both combined). For all the rest, while I agree in theory we could be on par, in practice it matters a lot that things happen in reality. There often is only a finite window to analyze and react and if you're slower, it's as good as if you knew nothing. Being good / being able to do something often means doing it in time.
We can’t program software or train neutral networks to do something that we have no idea how to do.
Machine learning does that. We don't know how all these layers and neurons work, we could not build the network from scratch. We cannot engineer/build/create the correct weights, but we can approach them in training.
Also look at Generative Adversarial Networks (GANs). The adversarial part is literally to train a network to detect bad AI generated output, and tweak the generative part based on that error to produce better output, rinse and repeat. Note this by definition includes a (specific) AI detector software, it requires it to work.
What exactly is "this"?
The results of this survey showing that humans are no better than a coin flip.
while I agree in theory we could be on par, in practice it matters a lot that things happen in reality.
I didn't say "on par." I said we know how. I didn't say we were capable, but we know how it would be done. With AI detection, we have no idea how it would be done.
Machine learning does that.
No it doesn't. It speedruns the tedious parts of writing algorithms, but we still need to be able to compose the problem and tell the network what an acceptable solution would be.
Also look at Generative Adversarial Networks (GANs). [...] this by definition includes a (specific) AI detector software, it requires it to work.
Several startups, existing tech giants, AI companies, and university research departments have tried. There are literally millions on the line. All they've managed to do is get students incorrectly suspended from school, misidentify the US Constitution as AI output, and get a network really good at identifying training data and absolutely useless at identifying real world data.
Note that I said that this is probably impossible, only because we've never done it before and the experiments undertaken so far by some of the most brilliant people in the world have yielded useless results. I could be wrong. But the evidence so far seems to indicate otherwise.
Wow, I got a 12/20. I thought I would get less. I'm scared for the future of artists
Why? A lot of artists have adopted AI and use it as just another tool.
As with other AI-enhanced jobs, that probably still means less jobs in the long run.
Now one artist can make more art in the same time, or produce different styles which previously had required different artists.
Having used stable diffusion quite a bit, I suspect the data set here is using only the most difficult to distinguish photos. Most results are nowhere near as convincing as these. Notice the lack of hands. Still, this establishes that AI is capable of creating art that most people can't tell apart from human made art, albeit with some trial and error and a lot of duds.
Idk if I'd agree that cherry picking images has any negative impact on the validity of the results - when people are creating an AI generated image, particularly if they intend to deceive, they'll keep generating images until they get one that's convincing
At least when I use SD, I generally generate 3-5 images for each prompt, often regenerating several times with small tweaks to the prompt until I get something I'm satisfied with.
Whether or not humans can recognize the worst efforts of these AI image generators is more or less irrelevant, because only the laziest deceivers will be using the really obviously wonky images, rather than cherry picking
These images were fun, but we can't draw any conclusions from it. They were clearly chosen to be hard to distinguish. It's like picking 20 images of androgynous looking people and then asking everyone to identify them as women or men. The fact that success rate will be near 50% says nothing about the general skill of identifying gender.
I have it on very good authority from some very confident people that all ai art is garbage and easy to identify. So this is an excellent dataset to validate my priors.
Curious which man made image was most likely to be classified as ai generated
and
About 20% got those correct as human-made.
Sketches are especially hard to tell apart because even humans put in extra lines and add embellishments here and there. I'm not surprised more than 70% of participants weren't able to tell that one was generated.
Thank you so much for sharing the results. Very interesting to see the outcome after participating in the survey.
Out of interest: do you know how many participants came from Lemmy compared to other platforms?
I guess I should feel good with my 12/20 since it's better than average.
12/20 is not a good result. There's a 25% chance of getting the same score (or better) by just guessing. The comments section is a good place for all the lucky guessers (one out of 4 test takers) to congregate.
What I learnt is that I’m bad at that task.
It'd be interesting to also see the two non-AI images the most people thought were.
I did 12 and I thought it was a pretty bad result. Apparently I was wrong!
I tried to answer by instinct, without careful examination, to make it more realistic
15 here and when I reviewed the answers I realized most were sheer luck.
The fact that it's a clean gaussian shows that it's mostly luck anyway.
I got 11/20 and there were a couple of guesses in there that I got right and wrong. Funny how there are some man-made ones that seem like AI. I think it's the blurry/fuzziness maybe?
I would second the comment from the original post which said that they didn't pay much attention in the latter half.
Oof. I got about 65% on the images I hadn’t seen in the post. I must be pretty close to being replaceable by an adversarial network.
You put 'tell' twice in the title 😅
God DAMN it
I can't tell if that would be more of a human or an AI mistake 🧐
Interesting. So you've given us a 50/ 50 chance. Usually you've given us the art that was used and then the AI has attempted it's own version? Did you train the ai using the art ?
Are you allowed to do that ? Is the art in the public sphere?
I have no knowledge or understanding of ai.
No, the AI didn't try to copy the other art that was included. I also don't train the model myself, I just tell it to create an image similar to another one. For example the fourth picture I told it to create a rough sketch of a person sitting on a bench using an ink pen, then I went online and looked for a human-made one that's of a similar style.
Can people tell tell
That avocado and tomato post took me out, that and the Legos. Very impressive.
The most obvious ai one for me was the last cat picture, somehow it just screamed ai
To me 4 was pretty obvious as a lot of the lines made no sense, such as the ones right below her face.
Also, is there a correlation between image generator and accuracy?
Is there any chance that the human artists experimented with AI and posted the result along with their usual work? :)
Huh, I felt the 12/20 was a bit low but I guess not so much. As someone that has never used an image generator (or an LLM for that matter, chatGPT not even once baby) nor has actually worked at tech (though I have been learning programming on my own) and doesn't even know how to draw... I guess I didn't do too bad.
8/20 and I use it everyday. I got all humans correct, but the landscape and stylistic images is what I got wrong. Ai is getting good.
AI just has a niche. Certain things, like highly stylized images, it's great at because they're generic, and genetic/common is what AI does best.
No shot the bottom left was ai gen without human help. AI has so much trouble delivering words and text.
Yes shot. Check out DALL-E 3.
There are newer AI tools that can do text accurately. Usually it's with text provided in the prompt though, so it's arguably not AI generated, just AI placed.
The unreadable AI text you're familiar with is done without the AI really knowing what text is, and they're just making something that looks vaguely like it. It's the same way they normally handle any kind of object or shape. Newer tools are built specifically for text, so it actually is readable and makes sense.
Luckily we've already begun developing tools to help detect AI Images with higher accuracy than a human curator.
Unluckily, they're also detecting human-made images as AI images.
I said "with higher accuracy than a human curator." You didn't really build upon that, no offence. You also didn't upvote despite literally repeating something that I said. You just like to take up space in people's inboxes? I'm trying not to be an asshole about it but I feel legitimate confusion about the purpose of your reply.
This isn't possible as of now, at least not reliably. Yes, you can tailor a model to one specific generative model, but because we have no reliable outlier detection (to train the "AI made detector"), a generative model can always be trained with the detector model incorporated in the training process. The generative model (or a new model only designed to perturb output of the "original" generative model) would then learn to create outliers to the outlier detector, effectively fooling the detector. An outlier is everything that pretends to be "normal" but isn't.
In short: as of now we have no way to effectively and reliably defend against adversarial examples. This implies, that we have no way to effectively and reliably detect AI generated content.
Please correct me if I'm wrong, I might be mixing up some things.
