Is that image AI? Here are 14 telltale signs to look for.

Rob Laughter
10 min readMay 2, 2024

--

One of the most common fears I hear with regard to AI generated images is that image generation will become so good that we won’t be able to distinguish between real images and synthetic images.

First, I hate to break it to you, but we are already there, and we’ve been there for some time. I’ve personally seen friends sharing AI generated images believing that they were real photos, who seemed downright betrayed when I told them that they weren’t authentic.

Second, AI image generation isn’t going anywhere, and it’s only going to get more convincing from here. And not just images. Video and audio, too.

Reactions to this reality setting in are wide ranging—from demands to ban all AI to just assuming that everything on the Internet is AI—but one thing is for certain. We’re living in a time when convincing—but completely fake—images can be generated out of the ether with a few strokes on a keyboard, and we need to be more discerning about the content we engage with in the world around us than ever before.

How I learned to spot a fake

When I was a young lad, I worked as a cashier in a fast food restaurant. This was an era when most people paid with cash, so I’d handle hundreds of bills in a single shift.

One morning, a customer handed me a bill that felt “off.” The paper was too smooth. A little too thin. The ink didn’t have the contrast of all of the other bills I had handled that day.

At the cash register, I had a counterfeit detector pen—one you’d use to mark a bill, and the ink would change color if it were a fake—but I didn’t have to use it. Because I had handled so many examples of the real thing, I could tell it was a fake from the moment it touched my fingers.

In the same way, I’m in a unique position to spot AI generated images. As a photographer, I’ve edited hundreds of thousands of my own photos, peering at the subtle details of images down to the pixel. And as an AI enthusiast, I’ve generated at least 100,000 images using every mainstream image model under the sun—and a lot of obscure ones, too.

That exposure has given me a fine sensitivity between what a real image looks like and what a fake image looks like.

How you can spot AI generated images

But what if you—unlike me—have a life, and you haven’t seen a bajillion photos and AI generated images? How can you tell them apart?

At first glance, an image may be super convincing, but it turns out that there are common traits that many AI generated images share, and if you learn to look for them, you’ll be much more likely to spot AI generated images.

So buckle up; I’m going to help you differentiate between the real deal and the fakes. Here are 14 things to look for if you suspect that an image has been generated by AI.

1. Bad hands

Either that guy was born with an extra digit, or this image was generated with AI.

A common and long-standing refrain with AI generated images has been, “It can’t do hands!” To quote the great poet Homer Simpson, it’s funny because it’s true.

While image models have gotten better at generating coherent images of hands, this observation holds. Because there there is a virtually infinite number of positions in which a hand and its digits can be posed, fine details such as hands are tough for image generators to get right.

AI can generate images where hands are the main focus just fine. But look closely at images where hands aren’t prominent and you’ll often see subtle incoherencies—extra digits, fingers blending together, strange hand positions, etc.

2. Malformed text

Repeated and malformed text on the left, imitation text on the right.

In the same way that AI image generators struggle with hands, they are also notoriously bad at generating coherent text. That’s because they’re image generators, not text generators. Their primary job is to arrange pixels into images that look good together, not to write War and Peace.

As with hands, image models are getting better at creating text when the text is the focal point of the image, but merely imitates text in the fine details. If you suspect that an image has been generated with AI, look closely at places where text would appear in an authentic image.

3. Skin and hair textures

AI generated skin textures can have a plasticine quality, or have repeated unnatural textures.

Image models are often trained to recognize the big picture details first, which means that some of the finer details get overlooked. (You’re probably starting to detect a pattern here.)

That includes details such as skin textures. AI generated skin textures can look flat, “photoshopped,” or plasticine in quality. Certain features such as freckles or wrinkles may be exaggerated or generated in a repetitive pattern, kind of like they were clone stamped in Photoshop.

4. Lifeless, unfocused, or misshapen eyes

The eyes here have a lifeless quality to them, and the iris and pupil are misshapen.

Again, models continue to get better here day by day, but another detail that AI image models tend to struggle with are the eyes. There are a few telltale signs here that might tip you off to an AI generated image.

The first clue is that the eyes often have a lifeless or unfocused quality to them. Granted, my actual eyes probably have that sort of quality at the end of a long day, so it’s not a dead giveaway.

Another telltale sign is misshapen irises or pupils. This could be inconsistent sizing, blurred irises, or other unnatural traits that feel out of place.

5. Repeating patterns and textures

The fur here should be more chaotic. The repeating pattern and texture is a sign the image was generated with AI.

Another common trait that you’ll find in AI generated images is repeated patterns where you would normally expect randomness or chaos. You’ll often see this in hair and fur, foliage, wood, and other similar textures that should have a chaotic quality up close, but have the appearance of repetition at a distance.

In these cases, the image model is just approximating the idea of the texture, without the nuance and sophistication that you would find in the same texture in the real world.

6. Repeating patterns and textures (again)

The orderly lines of a roller coaster start to break down in the distance.

Patterns and textures deserve a second mention here, but in reverse. Previously, we looked at patterns where you’d expect disorder. Now, we’re looking at disorder where you would expect patterns.

These are areas where you would expect to see repeating patterns, but they aren’t there.

Things like parallel or perpendicular lines, tiles, and other such patterns where you’d expect order will often be misaligned or deformed in AI generated images.

7. Subject blending

Look at that sweet family OF CREEPY CLONES!

Subject blending happens when the image model applies the same traits to multiple subjects in an image. You’ll get this a lot in faces, clothing, and colors, particularly when the image generator is prompted with generic terms— such as “a family” in this case.

If your image feels like a clone army, it’s probably generated with AI.

8. Broken and incoherent lines

Totally incoherent lines found in the chains of a swing set.

To be fair, I struggle with this one when I’m sketching, so I can’t be too critical when an image model gets it wrong. Fortunately for us, this telltale sign helps us recognize images when they’re generated with AI.

If you look closely at AI images, you’ll often see that if a line—trim on a wall, a window sill, a pipe, a tree— is partially obscured by an object in the foreground, it will often be “broken” when it emerges from the other side.

9. Light rays, reflections, refraction, and shadows

These reflections look good on first glance, but don’t make any sense when you take a second to think about them.

As we’ve been exploring, image generators are good at mimicking reality, but they’re not necessarily good at simulating reality. That is, it can make details such as reflections and shadows look plausible, but it can’t simulate what they should actually look like if the light was obeying the basic laws of physics.

In the image above, you should be seeing more of the underside of the car reflected in the wet pavement, but instead you see the glare from the top of the car, as if the image of the car was just flipped over. Similarly, with the car in the distance, its headlights are reflecting in the pavement—while it’s driving in the opposite direction.

Shadows and refraction (the way light bends as it passes through a material) also often break the laws of physics as AI mimics, but doesn’t simulate, those optical qualities.

10. Inconsistency in fine details

This image looks good at first, but quickly fails the photographic Turing test as you examine it.

AI images can look stunning at first glance, but as I read in a Reddit post last week, some AI images “get worse the longer you look at it.” That is, the more you notice, the more obvious the AI quirks become. A clock has four hands. A teapot has two spouts and no handle. A man is flipping a burger with a wooden spoon.

If you suspect an image is AI, look for the subtleties that are out of place.

11. Incorrect or inconsistent proportions

That guy is cooking 2 pound burgers. I’d comment on the three grills, but that’s not out of the ordinary for a man.

AI can generate all sorts of incredibly realistic imagery, but on closer inspection, you may notice that proportions are just a bit off. On a person, limbs might be just a bit too long, or facial proportions might be close, but the eyes are too large. The relative sizes of objects within the scene may not make sense, as the AI does its best to reconstruct them, but doesn’t understand the relationships between them.

12. Perspective and orientation

Not quite…

Image models need to understand not only what an object (such as the Space Shuttle) looks like—they need to understand how that object is oriented in space relative to the viewer. When it’s time to generate an image, the model needs to coherently arrange all of the features of that object relative to the scene around it.

This is fine for simple objects, or objects that have very clear ups and downs, fronts and backs. It’s trickier for objects that can look radically different from different angles and perspectives, such as airplanes, wielded items (bats, clubs), or objects with straight lines (handles, railings, ropes).

Frankly, I’ve always had a hard time drawing wings on airplanes and space shuttles, so I’m not surprised that image models struggle here, too.

13. Incoherent relationships between objects

That’s an unconventional technique for pouring a glass of whiskey.

In a similar manner as above, AI image models can often understand individual objects very well, but when it comes to the relationships between those objects, they begin to struggle a bit.

In AI generated images, you might find the right objects in the scene, but you’ll find them in places they shouldn’t be, in arrangements that defy the laws of physics, and with other objects that shouldn’t go together.

14. Items that are out of place

Pens in pots with dirt. Pens in full cups of coffee. And how much coffee does this person need?!

If you don’t prompt for something when generating an AI image, the model will often fill in the details for you. And when it does, you can get some strange items in stranger places.

Like everything else in this guide, things will look fine on first glance, but the more you look, the more obvious it becomes that the image was created by a machine that is mimicking reality rather than understanding it. Look for objects that wouldn’t typically belong in a scene, like the pens in random places in the image above.

Developing your AI image detection skills

As I mentioned in my preamble, we are now officially living in a time when using discernment is more important than ever.

As the early Christian author, Lactantius, is attributed as having said:

“The first point of wisdom is to discern that which is false; the second, to know that which is true.”

We learn to recognize what we are familiar with, and the more you use AI image generators, the more you’ll learn to recognize when they’ve been used.

I hope this guide serves as a starting point to help raise an alarm when you see images that display some of these telltale signs of being generated with AI. But none of them are foolproof, and image models are only going to get better from here.

Be wise. Use discernment. Think twice before you share.

This post is brought to you by the A.I. Collaborative. Learn and practice generative A.I. skills in a community of like-minded peers. If you want to see more examples of realistic A.I. generated images, check out my side project, PixHub, an “Unsplash” for A.I. imagery.

--

--

Rob Laughter

Rob is a creative professional exploring the intersection of technology and creativity. His current muse is generative A.I.