The scourge of AI slop

Every equine artist I know collects reference pictures. My own collection numbers in the hundreds of thousands. I started collecting images in the late 1970s, cutting out magazine pictures and "relocating" my favorite horse libraries one xerox at a time. 

But the internet was a game-changer. That was when my reference collection went from several boxes of papers to a box of external hard drives. 

In the past, the biggest issue with using the internet to collect reference pictures was misidentified breeds or colors. That changed with generative AI. Now, realistic-looking photos can be fake. That is what the image at the top of this post is. Here it is with the original. 

That's the Arabian stallion Marwan Al Shaqab. Because he's a well-known horse, and that's one of the more recognizable images of him, quite a few people pointed out that this was an artificially generated image. 

That probably won't stop artists from painting models with patterns based on the image, or judging from encountering documentation showing that "pinto Arabians can look like this." Interesting photos spread through the model horse community. Trying to get the word out that one of them is misleading can be an uphill battle. 

Just ask me how many times I've talked about the horse on the left, and how often I still see his picture on documentation. No, he is not a buckskin silver. He is a cream-diluted horse turning grey. The horse on the right is a silver-diluted buckskin. 

Silver and cream interact, making the black areas of the horse look warm and brown (right) rather than cool and silver (left).

At least Alen's Smokey Bandit—the buckskin grey Tennessee Walking Horse—is a real horse. 

 At the moment, there are still things that tell someone when an image isn't a real horse. Probably the most consistent error is headgear. The programs do not recognize that there are two different groups of equipment—halters and bridles—so it tends to merge the two things. Even more telling, though, is that it cannot seem to distinguish bits from other metal hardware. Lead clips are likewise a mystery to it.

If the unrealistic appaloosa pattern wasn't instantly suspicious, the "snaffle bit as halter ring" is a good giveaway that this image is fake.

Generative AI struggles with pinto faces, and often substitutes "white grey with markings" in areas that would have pink skin.

But as AI-generated images get more sophisticated, artists—and anyone trying to evaluate realism in art —are going to face challenges. It will not be long before very convincing images, and even videos, will be out there. Eventually, the consistent "tells" will improve. 

Evaluating images

As AI-generated images become increasingly difficult to identify visually, how can artists avoid using misleading references? 

Here are my rules of thumb:

1. Professional-looking photos are usually marked by a photographer. Look for a signature or logo. Be on the lookout for partially removed watermarks; they are more common in AI images than you might expect. (See the first image on this post for an example.) 
   
   
2. Generally speaking, you want to see a clear connection between the horse pictured and the account posting the image. That might mean the poster is the owner, breeder, breed association, event, or photographer.
   
   
3. Look for motivation for sharing. Is the horse being offered for sale or promoted at stud? Is the owner sharing a recent success? Are they asking for input about color, ancestry, or breeding? Posts using AI images are often clearly trolling for engagement. (Responding to them is a great way to get the scammy "Forgive me, but I noticed how clever/pretty/interesting you were and just had to reach out..." replies.)
   
   
   
4. Look for a name. Owners and breeders usually identify their horses. Sometimes you'll need to find the original post to see this text. It may not be visible when viewing photos using the "media" tab.
   
   
5. Unless you are looking at historical photos, be somewhat cautious if you only find one image of a particular horse. 

Captions from AI images (left) often include emojis and hashtags. Captions for real horses (right) usually have identifying information.

Reliable images

Here is an example of an unusual pattern reference that checks all five "likely real" boxes. 

I should note that while this breed, the Freiberger, is known to have a dominant white mutation (W1), the expression of that pattern is typically all-white or near-white. There has also been an unusual splashed white Freiberger with a sabino-like pattern, though that horse was sterile.

This type of patterning, which falls within what I call the "patchy sabino" phenotype, is not something I have seen in W1 Freibergers—or any Freibergers to date.

However, the images are watermarked by the photographer (#1). The group where this first appeared, Freibergerpferde suchen (Seeking Freiberger Horses), is run by a non-profit that helps breeders find buyers (#2). The motivation for the post is, obviously, to sell the foal (#3). The foal's name, sire, and damsire are all listed in the description (#4). Finally, there are multiple images of what is clearly the same colt (#5). Those facts make this a much more reliable reference for painters than the image at the top of this post.

Collect from reliable sources

General interest horse groups on social media—and particularly Facebook—are some of the worst offenders for providing steady streams of fake images. If the group has the adjective "Beautiful" followed by a horse-related noun, or an adjective followed by "Stallions", or "Love" anywhere in the title, it's probably best to treat images there with a lot of skepticism. 

Facebook can still be a good source for reference images if you know where to look. What I look for are the following types of groups:

• Breed associations
• Breed appreciation sites (this one requires caution, especially for color breeds)
• Breeder pages
• Equine photographers
• Breed-specific color groups
• Breed history groups

I am also a big believer in looking beyond social media sites. Some of the best resources for reference images are auction catalogs, sales videos, and stallion directories, all of which are typically found on independent websites. 

A list of reliable image sources

Social media sites are a lot more likely to push AI-generated content into user's feeds. There are good resources for photos, but it takes some work to find them. With that in mind, I am assembling a list of reliable sources from the categories I've listed above. Once that is finished, the list will be added to the blog's "pages" list with the Horse Breed Reference List.