Skip Navigation
46 comments
    • AI is still dodgy when it comes to hands, though this isn't the sure tell it used to be. Count fingers first, but also look for fingertips that seem to split/merge, weird fingernails, and unnaturally bent fingers.
    • Try to read any text in the image. Missing/extra letters are common, as are nonsense words and misplaced punctuation. The smaller and more verbose the text, the more likely the AI will fuck up.
    • Check in particular for weird/smudgy text in the corners of the image. Sometimes the AI copies artist signatures. It never copies them correctly.
    • Coat/shirt buttons. Check the buttons one at a time first, and if none of them seem wrong individually, then compare them to the others to see if they are consistent.
    • Look at any visible characters' eyes, and in particular the irises. AI tends to generate weird fragmented/lined irises.
    • If the image has any surfaces that are reflecting nearby objects (not just mirrors, but things like polished wood), check for discrepancies between objects and the reflections of those objects, or for reflections that are missing entirely.

    This whole arms race of trying to look for new "tells" as AI images get harder to distinguish from genuine ones reminds me of this Terminator quote:

    The 600 series had rubber skin. We spotted them easy, but these are new. They look human — sweat, bad breath, everything. Very hard to spot.

    • That's a helpful checklist! Just tried on this image from a recent post about an AI-generated game video:

      • Eyes and the face look surprisingly okay, except for that one deep line that goes right across his nose. Thought at first that he's wearing glasses, but apparently he's not.
      • Reflections are inconsistent, there's a car reflecting on the ice when there is no car that could cast that reflection
      • I guess whoever published that picture was smart enough to pick a frame that doesn't show hands... No buttons on the jacket, either
      • I'm not sure what a car dashboard is supposed to look like, but this does not look right
      • Weird detail: WTF is that red thingy on the house supposed to be
      • This particular one gets watermarked by the AI, but I'm trying to learn how to spot it without that

      Good quote, indeed they're getting hard to spot, it made a very realistic face for the guy... I guess all I can do is practice to at least spot the most obvious ones. Surely they'll be used a lot for misinformation in the coming decade, I want to be prepared.

      • Another tell-tale sign of AI is that things that should be aligned are not. If you look at the rear view mirror, you can see how the small switch at the bottom is placed off-center.

        AI usually sucks at doing parallel lines and repetitive patterns. It is weird to have a grille where the glove compartment should be but if you look closer at that grille it is complete nonsense. It is the same thing with the A column that is an inconsistent mess.

      • Next one... I think here, the main tell would be how the bridge disappears into nothingness? The thing the other people in this thread mentioned about little details being wrong. And the way how the metal rods that make up the bridge are unnaturally blurry/squiggly/whatever.

        The reflection in the lake might be wrong, too, but I don't really know how that's supposed to look. Guy's face is a little blurry, but I could be convinced that that's bad camera settings if not for the bridge. He's wearing a clothing that doesn't have buttons. Right hand looks normal, but left hand is unnaturally tiny, even if well-formed.

      • That guys's hood is tattered in a very specific, odd way. The fabric also has an oddly smooth texture.

      • Nice catch! The line across the bridge of the nose is a definite tell. You mentioned glasses, and I think the AI pulled from images of some people with glasses and some without, so it generated the bridge but rest of them.

      • The face also lit wrongly (it’s winter, it will be whiter), the edges of cloth are wrong, the shoulder is giant actually

      • I think that red thingy on the house is their attempt at putting a basketball hoop on the garage.

  • Looking at detail in general, especially stuff that's complex like wrinkles, knots, frills and patterns around edges. AI is really bad at small details like that and keeping them stable, especially when compared with each other. Perspective can be really off too sometimes. Other stuff to look for: hair strands, the height of trees in the background, clouds, clothing trim, reflections, etc.

    I heard if you use Photoshop (or something similar) you can use that to show JPG artifacting in machine-sludge images too even if they're lossless files, though I didn't fully grasp that fully.

  • AI pictures are often too smooth. Like you can't see pixels in photos or there's no brushstrokes for digital paintings. Here's a digital painting done professionally:

    Notice on the waist how you can see where the artist started and stopped with their colors, despite the rest of the image having really good blending.

    Here's the same character done by Stable Diffusion:

    Zero brushstrokes. The lighting isn't stylized. The background is too in-focus.

    You'll notice this with photos of people, too. There won't be any skin blemishes or easily seen pixels if you zoom in.

    • Yeah, here the moon is red tinted, but the reflection from the silver colored armor is neutral to blue. That's either an inexperienced artist or AI. Also, the shadows can't seem to agree on where the blueish light is coming from. The head's shadow is sorta where the moon would cast it, but then the front of the chest armor is reflecting light where it should be cast in shadow from the moon.

      Also, more obviously, the sword is just floating there next to their third knee (they're standing, presumably on two legs, but their cloak appears to be just floating there in the bottom left).

    • Note how none of the objects are objects. The fasteners on the shoulders are smeary blobs, not representations of anything. The gem on the breastplate, on close examination, isn't a gem. It's a red blob. What is that thing sticking out of the left arm? It's nothing, it's just pixels stuck together by a machine.

  • i think the easiest thing is inconsistencies. easiest way to spot that would be if something is drawn multiple times, it's drawn differently. i've seen eye highlights be drawn completely differently in each eye, for example. check elements that are repeated a lot in the background in places that aren't attached.

    • elements that are repeated a lot in the background in places that aren't attached.

      Would the background "forest" in this "AI"-generated image be an example of this? The one that I circled in green. It looks oddly repetitive, I don't think natural forests are this regular...

      • Ah my advice was mostly for digital drawing stuff as that's what I look at most often, oops

        I think in this I'm drawn to the buildings and snow. they kind of look like they're bleeding into each other and aren't as rigid as they should be. but yea, the trees definetly look off. to get that sorta blur you'd have to blur the background or have a camera's focal length set to... whatever i don't know shit about cameras. that happens in photography of landscapes a lot, but doesn't make sense to be right next to something that isn't blurred at all.

      • That image is all kinds of fucked up. Like there are serious alien geometries in it.

  • A lot of people who use AI Generated pictures are lazy, so most of them use very simple prompts which results in most of the ones having humans who look way too fake, sometimes they look like they are made out of plastic and very cartoony

  • It's getting less reliable, but look at the details of complex objects. The computational plagiarism engines can't think, and can't do abstract concepts. They fuck up hands because they don't have a concept of "hand" as a category of objects with fingers on the end of arms that grasp things. They're not drawing hands, they're putting pixels where those pixels are statistically probable.

    So, look at complex objects - jewelry, buckles, electronics, guns, anything complex. The plagiarism engine doesn't "'know" what those things are. It doesn't know what buttons are or that buckles fasten things together. It's just putting pixels in statistically probable positions. Hence those objects tend to be smears of color and light and dark without actual details. That smearing is still present in many images even as the machines become more sophisticated. The boundaries between things are indistinct and wrong because the machine doesn't "know" what things, objects, concepts, are.

  • abstract/minimalist/any heavily stylized image where strokes (pencil or brushes) aren't expected - i cannot tell tbh (if its not cgi looking that is). So where gaussian blur is not slapped on top and nonsense might be expected -

    photo-realistic in general does still have blur slapped on top of it, and weird anatomy/skin issues which require effort to fix (but are fixable already, people just do slop). Also issues with background are fixed with compositing, but as it is slop - people accept blurred background resulting in hyper-focused subject/blurred background which is the easiest tell for now.

    Also, for cheap slop humans: if they are small in the image, they don't resolve right in the details (weird ears/eyes/lips/hands), anatomical proportions might be wrong (legs/elbows/neck/body angle/length); skin is flat-out weird frequently (no hairs/too smooth/too wrinkly/just weird), phantom ligaments or muscles where there shouldn't be any (collar/arms)

46 comments