7 Issues With AI Art

With the emergence of powerful AI image generators like Midjourney and Stable Diffusion, numerous questions about the technology have arisen. From legal and ethical problems of AI using other peoples’ work to technical difficulties, there are many issues with AI art.

AI art faces many issues because it takes work that could’ve been done by a paid artist, uses image data without permission for machine learning, and AI images can’t be copyrighted. There are also some technical limitations. Namely, AI tools aren’t very configurable and struggle with small details.

The rest of this article will go over all the various problems with AI art. We’ll first review some societal problems of AI art and then dive into the technical side.

1. AI Image Generators Take Work Away From Artists

One of the most prevalent issues with AI art, and AI tools in general, is that they’ll replace humans. BBC reports that AI may replace up to 300 million jobs, largely in creative sectors like drawing and writing.

Namely, freelance designers and illustrators are already losing out on potential income due to AI. Netflix’s controversial use of AI art in an anime made headlines, for one.

All these potential jobs could’ve gone to artists who have to pay bills, rent, and buy food. Essentially, digital artists spent years learning drawing tools and how to draw, only to be replaced by AI.

This is more of an ethical issue rather than a technical one. AI art is already “good enough” for small-scale businesses, bloggers, lawyers, and other smaller potential clients.

After all, AI art generators like Midjourney only cost about $10 to make 200 images. Good luck finding an artist willing to make 2 images for $10.

This is why many companies, publishers, journals, and others have taken a firm stance against AI art. They respect their artists and are trying to push back against the wave of mediocre AI images.

This will be an uphill battle for artists. But with the right legislation and copyright laws in place, it may result in a victory or a draw.

2. AI Is Trained Using Data Without Permission

After stealing jobs, the biggest issue of AI art is that virtually all tools have been trained on copyrighted images.

For instance, Stable Diffusion used billions of image-text pairs from Common Crawl’s publicly available datasets. If it’s on the internet, it’s in their dataset.

Data crawling is considered fair use, and artists can’t really do anything about it. However, using it to train AI is morally and legally ambiguous.

And even though AI doesn’t splice different images together to create new ones, it’s still inherently wrong. For example, if you’re an established artist with a specific style, AI can “steal” it and apply it to whatever it wants.

You can learn more about how AI uses copyrighted data from this interactive Washington Post article.

In a nutshell, AI takes a publicly available image (copyrighted or not), breaks it down into noise, and then rebuilds it. It compares the original image to the new one and “learns” from it. It changes some parameters within the neural network and repeats the process.

It also uses textual descriptions from images to learn how to recognize patterns. That’s why it can replicate iconic art styles, from Da Vinci and Van Gogh to contemporary artists like Damien Hirst. And that’s why artists whose work has been stolen are so furious about it.

3. The Images Can’t Be Copyrighted

I previously explained that AI was trained using copyrighted images. But that’s not the end of it.

Another enormous limitation of AI-generated art is that it legally cannot be copyrighted. But that comes with an asterisk.

Namely, artists can copyright AI-generated images if they can prove that they changed them significantly.

This essentially means that any AI-generated images are free to copy and be used for commercial purposes. For example, if you’re a business that uses AI images, everybody else can just download those images and use them. They’re not stealing anything because you don’t own them, and they could print those images on T-shirts and sell them if they want.

Of course, this has some major implications. Any reputable company that cares about its brand will want to avoid AI art for this reason alone. This is one of the few things unlikely to change to favor AI art.

4. AI Isn’t Good With Body Parts

Let’s now delve into some of the more technical issues with AI images.

For one, AI sucks with body parts. It never seems to know how many fingers humans have, what a normal handshake looks like, or how many teeth should be visible in a smile.

For research, I asked Stable Diffusion to give me an image of a girl with beautiful blue eyes. What I got were convincing-looking faces but with hideous, distorted blue eyes that stared into my soul.

It also struggles with less prominent details, such as muscles. It may not be noticeable at first glance, though.

But if you look closely, you’ll often notice that the muscles don’t look anatomically correct. For instance, abs often have way too many rows, and arms tend to look pretty weird as well.

Other common issues include glass frames, ears, and feet, and it sometimes even blends body parts together in uncanny ways.

5. AI Makes Mistakes With Smaller Details

Tying into the previous point, AI doesn’t handle any sort of minor details particularly well.

Sure, it can make a button on a shirt look like a button at first glance. However, if you zoom in, you might notice that there aren’t any holes for the thread. Sometimes, they’re not even buttons at all but some random circles.

Most notably, the background in AI-generated images never looks right. If you’re trying to tell whether a picture was made using AI, just look at the details in the background. You’ll quickly notice that they often don’t make any sense. One of the funniest errors I saw was a bald waiter with hair that followed him.

Also, everything AI generates is always too perfect. There is no speck of dust on a linen sheet, random gray hairs, or a single scar.

These issues are caused because of the poor-quality data that AI was trained on and computational limitations. As computers get more powerful and human-built databases increase, these problems will likely vanish.

6. Most AI Parameters Can’t Be Configured Easily

Although we designed neural networks and have a vague understanding of how they function, we can’t understand the parameters. Understanding what each parameter does is impossible, as most modern AIs have trillions.

So, the way engineers train AI is by essentially going back and forth with the AI. Several complex machine learning algorithms are used for this, such as:

Gradient descent, stochastic gradient descent (SGD), and Minibatch SGD
Adaptive gradient (AdaGrad), AdaDelta, Root Mean Squared Prop (RMSProp)
Momentum-based gradient
Nesterov accelerated gradient (NAG)

All of these methods are computationally demanding and extremely time-consuming.

7. AI Is Biased

The last issue is likely the worst and most controversial, and it again concerns what data is used to train AI. AI has a bias toward white males. Some AI image generators use limited datasets for women due to explicit content. Other data may simply be statistically smaller, such as images of Native Americans or the Inuit.

These incomplete datasets and biases that seep into AI can be inadvertently discriminatory.

To be more exact, AI has a strong tendency to ignore different ethnicities and genders in many business-oriented prompts. If you ask an AI tool to generate an image of a CEO, it rarely draws women or non-whites.

Final Thoughts

AI art is impressive, but it also has a lot of major issues. AI art is facing a lot of backlash from artists because it’s stealing their jobs simply because it’s cheaper. Aside from ethical and legal issues regarding AI art, there are also problems with the quality of the images.

AI doesn’t handle backgrounds and smaller details well, and it also tends to leave out women and non-binary people, as well as different ethnicities.