Missing fingers, garbled text, and unlikely architecture are just the beginning: Here’s how to detect AI image fakery.
In today’s world, AI images can be created by anyone with access to a handful of AI engines including OpenAI’s DALL-E, Midjourney, Gencraft, or Stable Diffusion. They’re cropping up on social media and websites all over the place, frequently without any identification clearly explaining that they’re artificially generated.
And in fact, AI images have been used for dozens of hoaxes. They’ve tricked people into thinking Trump has been arrested in a huge public spectacle, or that the Pope has developed a radical new fashion sense.
Perhaps you think you can always spot an AI image. Or perhaps you’re scared that you never will. Whatever the case, we’re here to help. The truth is that AI-generated images can’t fully replace real life photographs — at least, not quite yet.
1: Weird Fingers
AI images are famously bad at hands. They’ll include a sixth finger, leave off a thumb, or add an extra joint. Some hands might have veiny palms, or some fingers might blend together.
To AI engines, hands are a fairly small part of an entire human, and don’t show up as consistently in images as a human face does. With more limited data, getting the ratio and number of digits correct is tough for an AI. But to humans, those errors jump out intuitively. This makes them a great way to spot a fake.
AI images have improved their hands recently, with one Midjourney software update making headlines for its realistic hands in mid-March 2023. Still, a sixth finger remains a common giveaway.
2: Blurry or Abnormal Backgrounds
Many AI image generators will blur the background entirely, ensuring that any bizarre details don’t jump out quite as readily. As a result, a fully blurred background is already a little suspect. But you should look even deeper: Chances are pretty good that an AI image’s background will be messed up.
Architecture is one giveaway. Steps may be misaligned, walls might be oddly curved, and ceilings might be sloped with no rhyme or reason. An office chair may be off-puttingly large or a coffee table might have too few legs. Light fixtures are another common issue, so check them to see if you can spot any structural impossibilities or differences within a row of hanging lamps that are ostensibly the same.
In the case of a huge crowd, the blurry background might be hiding truly large mistakes, like missing limbs or three-eyed people.
3: Terrible Hair
Individual hairs are a big problem for AI. An image engine might generate sharp detail, gauzy whisps, blurred sections, and radical changes in texture — all on the same head. By checking for texture inconsistencies, you can flag AI images.
Looking for hair impossibilities may also help. Hair might hover too high off of a head or might weirdly loop around to connect itself to a head on both ends like a Dr. Seuss character. In the most extreme cases, hair can even transform into a scarf or piece of clothing. But in a lot of cases, the hair won’t be obviously wrong: It’ll just be a little too thick, or a little more blurred than you’d expect.
4: Sharp, Overly “Rendered” Appearance
AI images’ problems with texture don’t stop at hair. Many entire images come with a glossy, unrealistic sheen to them, reminicent of how a randered video game character can never fully replicate film.
This stands in contrast to the blurriness many AI images include, although images often include both: A blurry background, a crisp and smooth foreground, or a flucuating combination of the two. This odd appearance is most obvious when it comes to cartoon-like artworks created by artificial intelligence, but it has plenty into ostensibly photorealistic creations as well.
One day, we may be able to find the words to describe this unique “rendered” appearance beyond just “AI-looking.” But until then, it’s one way to spot a fake.
5: Accessory Mistakes
Shirt collars, necklaces, earrings, scarves, purse straps, and shirt buttons. In real life, all these little add-ons are the right size, make sense, and obey the laws of physics. In AI images, anything goes.
Check for jewelry that’s warped or one earring that isn’t the same size as another. A ring might not wrap around a finger, or a necklace might hang too high on a neck. And once again, blurs may magically appear to steer your eye away from a tough-to-create detail like a watch face.
Hand-held objects are a problem area for AI, partially due to its issues with hands: Check for elongated coffee mugs or hovering pens. Object perminance is an issue, as well: The top of a walking cane might be visible over a leg, while the entire bottom half fails to reappear under the leg.
6: Garbled Text
AI images are designed to make up new combinations of existing content. But when it comes to a written languages, we don’t really have a use for new “combinations” of letters, give or take a few terms like “rizz.”
Artificial images that try to create their own storefronts, bedroom posters, or street signs are far more likely to wind up looking like an alien language than anything a human would recognize. Check for any text hidden in a background, and you might uncover the final clue you need to determine that an image is a hoax.
7: Stereotypes
Artificial intelligence works off of existing data and ideas. It remixes. As a result, it replicates baises or factual errors that exist in that data. And unfortunately, there’s plenty of bias out there in the world. There’s racism, sexism, classism, fatphobia, and ablism — and that’s just to name five that the TikTok algorithm has been credibly accused of.
Ask an AI image generator to give you a “doctor” and it’ll produce a white man in a lab coat and stethoscope. You’ll have to give it more specifics in order to generate an example that reflects the diversity of the real world, and even then half the time you’ll just wind up with a more specific stereotype.
Prejudices aside, AI images even tend to reproduce common poses or lighting conditions, since their datasets have the most examples of these.
8: Flubbed Details
One theme you may have noticed by now? AI images tend to mess up the little things. We’ve mentioned architecture mistakes in backgrounds, jewelry on the wrong fingers, or fingers on the wrong hands, but these types of mistakes can ultimately turn up on anything that’s detailed enough.
Other common areas that AI still can’t handle replicating as perfectly as Mother Nature include:
- Teeth
- Eyes
- Ears
- Tools, like scissors or wrenches
- Knick-knacks and decorations
Treat an image like a spot-the-changes page, and you’ll probably find something that gives the game away.
9: Failing an AI Detection Tool
Developing an eye for detecting AI images is an incredibly useful superpower: It boosts your odds of avoiding misinformation in the moment, and it makes an artificial images jump out at you that much faster. But as generative AI tools continue to evolve, you may not be able to trust your own eyes as well as you once did. That’s where AI detection tools come in.
Downloading an app or browser extension allows you to judge the veracity of an image with a single click. One option is “Hive AI Detector,” a Chrome extension that will issue a score that ranks the odds of an image being real or not. It may tell you that one image is “85.9%” likely to be AI-generated, for example. It even suggests which AI engine likely created the image, and which areas of the image are the most clearly artificial.
You can also try a reverse image search tool: TinEye is one of the best.
One word of caution: AI detectors can be fallible, which is why we ranked them last on this list. Building up your own understanding of what to look for, from blurry background details to overly sleek and shiny renders, just might be the best approach to staying aware of what’s real and what’s not. And it’s definitely a lot quicker.
