What a handsome Komodo dragon!
Except... what's going on with those labial scales? Something about them looks off. And that row of spikes down the back, Komodos don't have that. And their nostrils aren't so round. Hm. I think I know what's happening here.
I did not realize that this was AI at first glance; I thought it was just a really heavy editing style. I saw the missing central toe, yeah, but Komodos will do that to each other sometimes. The other forefoot, the one with four toes, is positioned in such a way that the fifth toe could be hidden, and the lifted hind foot could be similar.
But if you know Komodo dragon anatomy like I do, the inaccuracies pop up pretty quickly. Still, it's not like it was during the early days of GenAI; what I'm seeing here is that the GenAI image algos are getting better at discerning what part of a picture is a Komodo dragon. Since the inception, GenAI has really struggled to make accurate reptiles. In the beginning, everything was an iguana... even the Komodo dragons.
Over time, the models have been refined, but there are still some pretty obvious anatomy differences- the slit pupils, the mouth shape, the overall definition of the snout...
And they often struggle with the tongue. This isn't what the inside of a lizard's mouth looks like!
There's a lot of talk about how GenAI is bad for the brain- but it seems like most of it is actually about writing. And I think we should be talking more about images, too. Not even just about the stolen training data or the erosion of opportunities for artists, but... what is such easy access to these generated images doing to our ability to perceive what's real versus what isn't?
Every single one of the images I pulled is from a highly popular stock photo site. In case you don't know what a stock photo is, it's a photograph (you can also have stock illustrations and stock footage) that's been licensed to use in different applications. These pictures aren't taken for a specific client; anybody who pays can use them within the terms of the image's license.
But all of these images- they're not photos. They're inaccurate illustrations. I recognize them for what they are because I spend a lot of time looking at lizards, but what if you've got someone writing a quick news story, or designing a science worksheet, or throwing together a museum brochure or a zoo sign? If they don’t know what a Komodo dragon is supposed to look like, they’ll use whatever looks convincing.
Images trigger something deep within us- you know that saying "A picture is worth a thousand words?" We're wired to trust what we see. But the problem here is that genAI doesn't create an image of the thing; it creates what its internal logic says is associated with the subject of the prompt. It all comes down to probability; generative AI makes images by looking at its training data and creating output based on what the data is associated with.
(For more info on how AI "sees" what it does, check out the LENS project, which you can read more about here.)
We don't see things the same way the computers do, and we're willing to trust images more than words. How many stock photos do you think you see each day? It's probably more than you think; after all, the average American sees around 5,000 ads per day. And while those photos are marked as AI generated on the stock sites, they aren't marked as AI generated once someone has licensed them. And if the stock site doesn't have what you need? No problem, just use the AI image generator to fake that photo yourself!
We already have seen political deepfakes and AI generated images used to spread misinformation. Did you see the image of an ICE agent arresting a Doordash worker? That was an AI fake, part of a larger hoax. Some of us are already learning to respond with increased skepticism to important images, because people have an agenda to fake those. But what about the less important images, the background images, the completely mundane images? GenAI seems to be quietly coming for them, and it's something we should be paying attention to, because if we're exposed, constantly and quietly, to generated images and are trained to believe it's photography, we'll be more accepting of the bigger lies when we see them.
I don't really know what the solution is here, other than for people to be aware of the stock image issue, and to stop using stock sites that allow generated images, like Adobe Stock. We can't put the generative AI genie back in the bottle, but we can at least be aware of the damage it's causing. And maybe part of the solution is to look for alternative stock and reference options. Maybe we'll start to see more photographers licensing their images directly, or putting together specialized repositories of images based around a theme or topic that they specialize in. The downside there is that it's less convenient than the stock model where there's thousands and thousands of images on every conceivable topic to choose from. I don't know what genAI is going to do to the traditional stock model, but I'm concerned about what the end results might be and what those results might do to our ability to perceive reality.
