Topic: How to actually tell if an image is AI?

Posted under Art Talk

Was looking around found a forum that linked to a deleted post, looked at the original for the context and below saw this image: https://x.com/BeJackko/status/2024618972105076841

So I went looking for the original image and found this: https://x.com/MexicoUd/status/2024579869753889195

And below that image are memes implying the image is AI. Not to mention that no artist is linked, and the account that posted it is not an artist account.

With or even without that context, how does one actually tell if an image is AI anyway?

Considering that any information on how to tell if an image is AI is also advice for AI users on how to make their images more convincing I don't expect you to get very detailed answers.

Ultimately, you want to ask yourself if these are mistakes an actual artist would make. Keep in mind
I am no expert at this

Line consistency - Does the thickness, colour and boldness of the lines stay consistent and make sense?
Anatomy - Perhaps less of a tell in this day and age. But still, there might be anatomy that looks weird and off.
Accurate portrayal of details - Similar to line consistency. you want to see if the angle of body parts make sense, if things like emblems are mirrored properly on different sides of the body/ armour/clothing, if everything is properly rotated if a character turns around, stuff like that.
Output - you might find the output of AI accounts to exceed the majority of artists and be very unrealistic. Like say, 15 uploads in 1 or 2 weeks. Daily or near-daily uploads of full illustrations don't happen for actual artists.
Colouring style - AI colouring tends to look unnatural. Even now, it does seem to look plasticy or weirdly off
Bizzare quirks - Things like colouring outside the lines or having weird endpoints for lines. Do any lines melt into one another(usually smaller details like teeth or jewelry on a finger)? Stuff like this.

Mind you, none of these are ironclad rules. Genuine artists might do these and higher-effort scammers will be better at covering these up. You still gotta investigate individual pieces/accounts and use common sense. I suggest looking at a bunch of AI images if all this seems too hard. You should notice patterns if you look at enough of them

wanni said:
Considering that any information on how to tell if an image is AI is also advice for AI users on how to make their images more convincing I don't expect you to get very detailed answers.

AI users already have their own communities. Those looking to deceive others are already discussing these kind of things on places like reddit. Sometimes under the guise of "it shouldn't matter whether it's AI or not so I'm going to deceive people anyways." Besides, these things are sort of common knowledge if you know how to look.

peskeon said:
Was looking around found a forum that linked to a deleted post, looked at the original for the context and below saw this image: https://x.com/BeJackko/status/2024618972105076841

So I went looking for the original image and found this: https://x.com/MexicoUd/status/2024579869753889195

And below that image are memes implying the image is AI. Not to mention that no artist is linked, and the account that posted it is not an artist account.

With or even without that context, how does one actually tell if an image is AI anyway?

If we’re talking specifically about this image on Twitter, you can see a major issue with the teeth, differently “drawn” eyelashes on the male character, and blurry, smudged eyes. The characters’ bodies seem to look fine, but you need to pay attention to small details — especially on the face — to determine whether it’s AI or not.

In images with low facial and hair detail, it becomes harder to tell. Sometimes I also fall for AI tricks and think it’s art made by a real person :c

For me, I just look at the date of it's publication, anything prior to 2023-ish is not likely to be made by ai because it's still in it's developing stage. Though it would be bad for re-uploaded stuff for archival purposes or whatever.

My next strategy is to look what's the artist stands, are they anti-ai or not. Though there's a chance that would lie and use it or "prank" us in the future without knowing.

If all else fails, maybe look up any up-to-date ai detection tools? (It's not 100% full proof tbh)

whatismyname1234 said:
For me, I just look at the date of it's publication, anything prior to 2023-ish is not likely to be made by ai because it's still in it's developing stage. Though it would be bad for re-uploaded stuff for archival purposes or whatever.

Generating a video doesn't really compare to generating an image, image generations were happening in 2021/2022, and while many weren't convincing you could get lucky

whatismyname1234 said:
If all else fails, maybe look up any up-to-date ai detection tools? (It's not 100% full proof tbh)

You are better off never touching those, unless the image has actual metadata embedded into it listing the model or prompts (which the site would detect on its own), those things aren't often accurate at all

whatismyname1234 said:
For me, I just look at the date of it's publication, anything prior to 2023-ish is not likely to be made by ai because it's still in it's developing stage. Though it would be bad for re-uploaded stuff for archival purposes or whatever.

My next strategy is to look what's the artist stands, are they anti-ai or not. Though there's a chance that would lie and use it or "prank" us in the future without knowing.

If all else fails, maybe look up any up-to-date ai detection tools? (It's not 100% full proof tbh)

I have an FAQ on this for a very good reason.

We do not use "AI detector" services: the basis for software-based perceptual AI image detection is fundamentally flawed, and they produce false positives on known human-made images (post #1592031 predates AI image generation by a few years, but AI image detectors think it's suspicious because of its coloring), and false negatives on known AI-generated images (post #5503447 uses a known character model and comes from an artist with AI image usage on their storefronts, but AI image generators think it's nonsuspicious because of its coloring.)

lafcadio :
We do not use "AI detector" services: the basis for software-based perceptual AI image detection is fundamentally flawed, and they produce false positives on known human-made images (post #1592031 predates AI image generation by a few years, but AI image detectors think it's suspicious because of its coloring), and false negatives on known AI-generated images (post #5503447 uses a known character model and comes from an artist with AI image usage on their storefronts, but AI image generators think it's nonsuspicious because of its coloring.

Quality of AI detection tools varies a lot, and when I say a lot, it's really a lot. Most of them are only good at one or two type of pictures, like only between real photo and AI photo. They often don't say it when their tools doesn't include detection for painted pictures or 3D render pictures. 3D render picture is the most common one to not be trained for and give non-sense result.
You (I mean everyone) should test multiple website, and never trust what they said. I have seen one website that tells complete bullshit about 99% accuracy and is catastrophic at everything. The biggest issue in general is them telling how their are good in one test without saying that it is a test that exclude categories of pictures they are not good or made for.

Website that combine test from multiple other websites are not better. Bad + Bad doesn't make good.

The best AI detection tool I know is clearly decopy.ai.
It can distinguish between real/human and AI for : photo, physical painting, numerical painting, 3D render even ones of really good quality, and if the stuff has been retouched.
It say what the category is and give details of why.
But those details are not useful to improve our own capacity because it tells things like if it was obvious when it's impossible for a human eye to tell the difference.
I have not seen it yet (with prove) say that something is AI when it's not.
But it fails at saying that pictures generated by "This Person Does Not Exist " are AI (one of the harder test).
I have also seen the tool say, on really difficult AI generated pictures, that it was AI generated at the top right (true) but textually only explaining why it is not AI generated and thus conclude textually that it is not AI. When the text contradicts the probabilities, the probabilities are better to follow.

mikael_the_d said:
Quality of AI detection tools varies a lot, and when I say a lot, it's really a lot. Most of them are only good at one or two type of pictures, like only between real photo and AI photo. They often don't say it when their tools doesn't include detection for painted pictures or 3D render pictures. 3D render picture is the most common one to not be trained for and give non-sense result.
You (I mean everyone) should test multiple website, and never trust what they said. I have seen one website that tells complete bullshit about 99% accuracy and is catastrophic at everything. The biggest issue in general is them telling how their are good in one test without saying that it is a test that exclude categories of pictures they are not good or made for.

Website that combine test from multiple other websites are not better. Bad + Bad doesn't make good.

The best AI detection tool I know is clearly decopy.ai.
It can distinguish between real/human and AI for : photo, physical painting, numerical painting, 3D render even ones of really good quality, and if the stuff has been retouched.
It say what the category is and give details of why.
But those details are not useful to improve our own capacity because it tells things like if it was obvious when it's impossible for a human eye to tell the difference.
I have not seen it yet (with prove) say that something is AI when it's not.
But it fails at saying that pictures generated by "This Person Does Not Exist " are AI (one of the harder test).
I have also seen the tool say, on really difficult AI generated pictures, that it was AI generated at the top right (true) but textually only explaining why it is not AI generated and thus conclude textually that it is not AI. When the text contradicts the probabilities, the probabilities are better to follow.

I just uploaded the following test images:

  • an AI-generated image sourced from an AI roleplaying site, which Decopy gave 100% odds of being human-made.
  • a programmatically-made pixel art conversion of AI art, which Decopy gave 99% odds of being human-made.
  • post #4443685, which Decopy gave 83% odds of being AI-generated.

No. These services are genuinely useless. Every single one. If you advocate for these I automatically think less of your ability to scrutinize art.

I would remind people about confirmation bias, placebo effect, but most importantly, these tools are basically all inherently useless when analyzing anything that isn't just raw direct output from AI image generation. Whatever detection method you think they are using, do you think it still works if they rescale the output? Convert it to a different file type? Add a filter? Trace it?

Additionally, there's tools for localized AI generation inside a preexisting image; good luck detecting that. Btw, please keep this in mind when someone says they couldn't possibly be using AI because they have sketches; there's a hell of a lot of different ways to use AI beyond full image generation. It can just touch up sections of the image, just add shading & texture, just do linework... These things are extremely versatile, & automated tools are going to have significant trouble as anything beyond an early warning system (assuming they aren't just snake oil through & through, which some of them likely are). With tests like these, there's 4 groups; true positives, true negatives, false positives, & false negatives. These things can favor increasing true positives & accept increasing false positives as well, etc. If you know it's overly sensitive, it can be used to flag things for manual review; outside of that, disregard them entirely, & remember that you'll still get false negatives, even with that tuning.

AI Detectors are fatally flawed. They inherently will always be behind the latest models and techniques. AI is trained off humans, so AI images that have been carefully tuned using whatever techniques (LoRA, Controlnet w/ openpose/depth/etc, IPAdapters, inpainting, regional/noisy/step-varying conditioning, special samplers/sigmas, etc. etc.) will be hard to tell apart from the vast array of ways humans make art - thats their entire purpose, to produce and output that aligns with the (initially) human-made input data. To train an ai detector accurately, you would likely need even MORE initial data than the generator had, and unlike the generator, it needs to care whether those samples are AI, so you would need to label all of them accurately as well, which is hard when all the data is from the web.

Any AI detector that is checking the image itself is very likely going to be inaccurate. If a detector is instead only checking for metadata (which can be stripped out), some kind of special pattern the generator added, or common resolutions, it might be helpful to use as an initial check before you look at it with more scrutiny.

lafcadio said:
I just uploaded the following test images:

  • an AI-generated image sourced from an AI roleplaying site, which Decopy gave 100% odds of being human-made.
  • a programmatically-made pixel art conversion of AI art, which Decopy gave 99% odds of being human-made.
  • post #4443685, which Decopy gave 83% odds of being AI-generated.

No. These services are genuinely useless. Every single one. If you advocate for these I automatically think less of your ability to scrutinize art.

When I look at post #4443685, I think that it visually strongly look AI-generated. It has a strange contrast of between being simple and being precisely worked on. The shape doesn't look natural, with those back legs and the tail shape.
When I look at what the artist make, many thing don't make any sense : post #3319589, post #3355714, post #3330486. I am sure this one use AI.

For pixel art, it's not surprising that it doesn't work. There is almost nothing to work on. The number of pixels is extremely low.

Manitka

Former Staff

mikael_the_d said:
When I look at post #4443685, I think that it visually strongly look AI-generated. It has a strange contrast of between being simple and being precisely worked on. The shape doesn't look natural, with those back legs and the tail shape.
When I look at what the artist make, many thing don't make any sense : post #3319589, post #3355714, post #3330486. I am sure this one use AI.

And this is why we don’t trust users to figure out what’s ai or not. You’re accusing posts from before ai was even half decent as being ai.

mikael_the_d said:
When I look at post #4443685, I think that it visually strongly look AI-generated. It has a strange contrast of between being simple and being precisely worked on. The shape doesn't look natural, with those back legs and the tail shape.
When I look at what the artist make, many thing don't make any sense : post #3319589, post #3355714, post #3330486. I am sure this one use AI.

A. I'll reiterate confirmation bias; you trust the thing, so you're going to be predisposed to its verdict, consciously or otherwise.
B. Don't you think our resident expert validated these examples before presenting them?
C. If you genuinely think otherwise, I'd say so elsewhere.

People seriously need to discard the idea that there's some kind of AI visual style or whatever that you can just consistently pick out; our calls aren't just based on vague vibes.

1. A lot of AI images seem to share a limited selection of styles, which should help narrow things down a bit.

2. The two characters aren't exactly in the same style, like facial proportions and such.

3. Nobody's body looks like that with a shirt on.

mikael_the_d said:
When I look at post #4443685, I think that it visually strongly look AI-generated. It has a strange contrast of between being simple and being precisely worked on. The shape doesn't look natural, with those back legs and the tail shape.
When I look at what the artist make, many thing don't make any sense : post #3319589, post #3355714, post #3330486. I am sure this one use AI.

Mmmm. There is no text generated with post #4443685 and it says it's generated by AI. But when I am censuring the dick with a black box, then the text is generated and it says it is human made.
Porn seems to break the AI detector. Maybe because talking about explicit stuff is not allowed ?

I need to do some tests.

Updated

mikael_the_d said:
Mmmm. There is no text generated with post #4443685 and it says it's generated by AI. But when I am censuring the dick with a black box, then the text is generated and it says it is human made.
Porn seems to break the AI detector. Maybe because talking about explicit stuff is not allowed ?

I need to do some tests.

I reached the limit of use. Porn seem to do quite a mess with the tool, and probably more when it's more detailed. There is very often no text. Sometime it knows it is real, sometime it don't. Sometime there is a little of text but almost nothing. Censure can help a bit, but it just make it closer to 50/50 meaning it don't know anything. I have done more test on real pictures than art.

oh_no_a_skeleton said:
1. A lot of AI images seem to share a limited selection of styles, which should help narrow things down a bit.

Because unoriginal people all ask it to use the same styles, not because it's incapable of others; that's not how these things work.

oh_no_a_skeleton said:
2. The two characters aren't exactly in the same style, like facial proportions and such.

That's not what I meant; I'm referring to

mikael_the_d said:
I think that it visually strongly look AI-generated. It has a strange contrast of between being simple and being precisely worked on. The shape doesn't look natural, with those back legs and the tail shape.

oh_no_a_skeleton said:
3. Nobody's body looks like that with a shirt on.

And that's something that couldn't be made by people? There's a great many mistakes made by both.

Regardless, I'd prefer to not need to repeat myself; under no circumstances are we debating if a post is AI or not on the forums. If you think it's AI, flag it & move on; no public accusations.

mikael_the_d said:
Mmmm. There is no text generated with post #4443685 and it says it's generated by AI. But when I am censuring the dick with a black box, then the text is generated and it says it is human made.
Porn seems to break the AI detector. Maybe because talking about explicit stuff is not allowed ?

I need to do some tests.

...or... y'know... they're wholly unreliable? Like we told you? Multiple times?

mikael_the_d said:
Porn seem to do quite a mess with the tool, and probably more when it's more detailed.

...do you know what a placebo is? Confirmation bias? These ringing any bells? Any relevance to your current state of mind?

You are analyzing with the assumption that it can identify these correctly with any reliability; that's not how the scientific method works specifically because it can reinforce incorrect presumptions.

aacafah said:
...or... y'know... they're wholly unreliable? Like we told you? Multiple times?

...do you know what a placebo is? Confirmation bias? These ringing any bells? Any relevance to your current state of mind?

You are analyzing with the assumption that it can identify these correctly with any reliability; that's not how the scientific method works specifically because it can reinforce incorrect presumptions.

I am doing test with pictures that I know if they are real or not. I have a folder for that. And it's not normal that the text generation doesn't work. I always had text on non-porn pictures.

I know that it can say that some pictures are not AI when they are. But I have been confident against the artist because I haven't known of a false positive yet with this tool. I don't use it much and never on e621 content. I have done as much test with thing I know than try on stuff I receive.

Updated

aacafah said:
...do you know what a placebo is? Confirmation bias?

And yes I know what it is. It's just that I trust my experience more than a simple intuition that it shouldn't work.
And I know there is trash tools, I have seen it, thus I know why people can think it is impossible that it work.
Just this last one seemed better.

mikael_the_d said:
I am doing test with pictures that I know if they are real or not. I have a folder for that. And it's not normal that the text generation doesn't work. I always had text on non-porn pictures.

I know that it can say that some pictures are not AI when they are. But I have been confident against the artist because I haven't known of a false positive yet with this tool. I don't use it much and never on e621 content. I have done as much test with thing I know than try on stuff I receive.

The entire problem that you will never know of any false positives because you don't have a good idea of what makes a positive result to begin with. You don't have any reasonable heuristics for what is or isn't likely to be AI-assisted, all you have is vibes.

This service is a ChatGPT wrapper (this is the entire reason it won't write detailed descriptions for porn!) trying to get marks like you to pay money and spread the word.

Did you ever actually check their site in detail? The entire business model is like this

  • AI Detection
  • AI Content Detector
  • AI Image Detector
  • Humanize AI
  • Plagiarism Checker
  • Reverse Image Search
  • AI Code Detector
  • AI Voice Detector
  • AI Summary
  • YouTube Video Summarizer
  • YouTube Transcript Generator
  • YouTube Subtitle Downloader
  • Video Summarizer
  • Audio Summarizer
  • PDF Summarizer
  • PPT Summarizer
  • Word Summarizer
  • Image Summarizer
  • AI Summarizer
  • Text Summarizer
  • Article Summarizer
  • Study
  • AI Math Solver
  • AI Homework
  • Ask AI
  • AI Flashcard
  • Math Flashcards
  • Spanish Flashcards
  • AI Math Worksheet
  • Writing
  • Grammar Checker
  • AI Translator
  • PDF Translator
  • Image to Text/OCR

post #2386545 gives 95% odds of being AI-generated but its August 2020 date makes this impossible. Not "incredibly unlikely", not "Lafcadio thinks it looks fine", it is literally impossible to generate explicit furry AI images locally in August 2020. Ditto for post #3398391, which displays 100% odds of being AI-generated.

These services are all trash. Every single one. Without exception. The number of ways that AI usage can be incorporated in art is functionally equal to the number of ways that digital art can be made.

mikael_the_d said:
It's just that I trust my experience more than a simple intuition that it shouldn't work.

...your biased experience... more than simple intuitionconcurring logical analyses from multiple software developers... as if that isn't textbook placebo... Ok, whatever. I'll stop wasting my time; just flag it silently if you insist on refusing to listen to Laf - a person who routinely performs this exact analysis, has pages upon pages of documentation on the subject, & has repeatedly discussed methodology & validated their research with people who admitted to using AI - over some random rebranded Magic 8 Ball.

One last question; are you familiar with Dunning-Kruger?

Updated

lafcadio said:
This service is a ChatGPT wrapper (this is the entire reason it won't write detailed descriptions for porn!) trying to get marks like you to pay money and spread the word.

On review? Yup.
...I rescind my statement that calling this a rebranded Magic 8 Ball was a slight over-exaggeration.

aacafah said:
...your biased experience... more than simple intuitionconcurring logical analyses from multiple software developers... as if that isn't textbook placebo...

I haven't read that it is was according from "logical analyses from multiple software developers". I would have listen more closely in that case. I just saw people trying to explain to me why it would not work, without me knowing if those explanation would match reality.

aacafah said:
I'll stop wasting my time; just flag it silently if you insist on refusing to listen to Laf

I have reacted to the picture being said by the tool to be AI and was fooled by my trust in that this one tool was better than the other ones. I wasn't asking to remove it.
I know it still damage the author reputation but I had to respond.

aacafah said:
a person who routinely performs this exact analysis, has pages upon pages of documentation on the subject, & has repeatedly discussed methodology & validated their research with people who admitted to using AI - over some random rebranded Magic 8 Ball.

Ok. Sorry. I will not talk about it again.

It is also possible that the tool was cheating with my test. I didn't removed any of the metadata and most of the time I left the original name. In my test, the vast majority of the non-AI stuff can be trust just by the edition date which is left untouched in the metadata.
Other tools was worst despite of the same data being available.

aacafah said:
One last question; are you familiar with Dunning-Kruger?

I have heard of it multiple time, but I doesn't come back to my mind each time I do things.

Sorry about all of that.

Updated

lafcadio said:
This service is a ChatGPT wrapper (this is the entire reason it won't write detailed descriptions for porn!) trying to get marks like you to pay money and spread the word.

If it wasn't clear to anyone why that's a problem: If I were a techbro running AI generation services and trying to get more business, I would be highly motivated to make people opposed to AI look like fanatical idiots and spread the idea that my products are actually indistinguishable from the real thing. And starting an "AI detection service" that consistently produces inaccurate results is a perfect way to achieve both those goals.

And I can just about guarantee you that all the other ones are either openly or secretly run by gen-AI companies too.

Unless someone is using a straight generator with watermarking you can't tell.

I've built out a work flow that takes an AI image, drops it to a sketch and line drawing, converters it over beziar curves and then g-code which draws it in gimp (along with varying pressure weights). I've got a colorizer that does the same. Your not detecting that, even with your eyes. I'm also not uploading anything to this site, I respect the rules even if I disagree with them.

Point is, AI is past the point where you can detected it for drawings, if someone wants to hide that it's AI they can.

I know staff won't share their exact methods, but I'm curious as to how confident they are when removing posts. There's one instance I can think of which I find incredibly unlikely to be AI. It's pretty baffling and a little frustrating to see it get taken down with no further information.

easygoing-tigershark said:
I know staff won't share their exact methods, but I'm curious as to how confident they are when removing posts. There's one instance I can think of which I find incredibly unlikely to be AI. It's pretty baffling and a little frustrating to see it get taken down with no further information.

Based on your sole deleted favorite: Chantiyii's current Twitter profile includes an AI art disclosure, and a disclosure that they're a minor. These are two separate reasons that their art will not be displayed on e621.

lafcadio said:
Based on your sole deleted favorite: Chantiyii's current Twitter profile includes an AI art disclosure, and a disclosure that they're a minor. These are two separate reasons that their art will not be displayed on e621.

I was referring to a different post, but thanks for pointing that out, I didn't even notice it got deleted! I won't mention the post I'm thinking of here (to avoid calling out the artist), but do you think you could provide some more details if I sent you a DMail?

easygoing-tigershark said:
I was referring to a different post, but thanks for pointing that out, I didn't even notice it got deleted! I won't mention the post I'm thinking of here (to avoid calling out the artist), but do you think you could provide some more details if I sent you a DMail?

Maybe.

Unless there are similar levels of obviousness (e.g. public AI disclosures, or specific traces for which we know the original AI image), or you're a commissioner or somebody else who'd have a special claim to some of the posts, I won't be much more specific than "we recognize this style", "there are specific data markers normally found in AI images", etc.

The most detailed responses are basically only saved for when there is zero chance of the artist ever making a successful appeal, and even then certain details might be kept private because they also apply to other artists.