AI Art: Artists nightmare, or amazing new tool?
We’ve all seen the impressive images pouring down our social media feeds. The images were tagged with words as AI Art, Midjourney, Dall-E 2. But shortly after the image explosion, the artist community responded. A large protest on art portfolio website Art Station, and many disgruntled reactions on social media and videos on youtube.
Art Station Feed showing users protesting against AI Generated Art. (December 2022) Source: Artstation.com
Many of the artist responses claim that the machine learning software ‘steals’ from artists, that mangled parts of signatures can still be seen in the pictures, and that the software will destroy any chance of a future career in art.
In this article, I will try to explain how Image Generating AI’s work, what the current ethical problems are, and I will express my opinion on the validity of some of the arguments that are being made against AI Image Generating Software.
Image created using Midjourney AI. Prompt: a robot holding a paint brush and painting a landscape on a canvas
As this is a rapidly developing and relatively new technology, I might alter my opinion in the future.
I should also note that I am not an AI expert. I am a digital artist and AI enthusiast. If I made any mistakes in this article, or you would like to update me with information, I am happy to be educated, so feel free to contact me.
How does it work?
Before we can talk about the ethics of AI art, we first have to know how it works.
Scraping images for datasets
The internet is filled with images, most of these images contain something called a meta description. This is a description of what the image represents that is added manually or automatically to the image depending on the internet platform.
There are companies that scrape millions of these images and their linked meta descriptions, and put them in a dataset that can be used for machine learning. One of the most popular datasets that is used for Image Generating AI’s is the LAION-5B Dataset. This dataset contains more than 5 billion images and their correlating meta descriptions.
It is important to note that these datasets are usually put together by different companies than the ones that provide the Image Generating AI software to the public. It is also important to note that some Image Generating AI companies are very open about the datasets that they have used, while others prefer to keep that information private.
The dataset is then entered into a computer that uses software to train on the images and their description. Most of the Image Generating AI’s train with a “diffusion model”. This means that noise that looks like TV Static, is added to an image and it will keep adding noise as long as it can still recognize the subject that is described in the meta description.
For example, the computer takes an image that has the meta description ‘cat’. It recognizes from its previous training that the subject in the image is a cat. It adds noise, until the cat is barely visible anymore.
The next step is removing random noise from the image, until the image is clear again. By doing this over and over again, the computer gets incredibly skilled at making a noisy image clear. It gets so skilled that it can create clear image results from an image that contains pure noise.
This means that once the computer has finished training on millions of images and their meta description, you can tell the computer what to look for in random noise patterns. Similar to how a person might be able to see a cloud that resembles an animal.
Image created using Midjourney AI. Prompt: Cat in a shark onesie
Now that the computer has finished its training, the original dataset is discarded and the computer only works with its trained knowledge.
We can now request an image from the Image Generating AI in the form of a line of text. This line of text is often called a ‘prompt’.
Now, if we ask the Image Generating AI for ‘a cat wearing a shark onesie’, It will start with a random noise image, and it will try to find patterns that fit the knowledge it has of the words ‘cat’, ‘shark’, ‘onesie’. It will go back and forth with adding and removing noise, until the result resembles something it can recognize.
Important to note is that the image that is created is a completely new image. It didn’t copy and paste parts of images and put them together in some form of photo collage. It used its knowledge of what a cat, a shark, and a onesie is, to create something new.
The Ethical Problems with AI Generated Art
There are definitely problems with Image Generating AI’s, but maybe they are not what you think. There has been a lot of misinformation based on misunderstandings of how the Image Generating AI’s work. Before I will share my view on what the biggest ethical problems with AI art are, I will first address some of the arguments against AI art.
Argument 1: AI steals from artists by copy and pasting parts of their images and creating a new one
This argument is false.
As described, when the computer learns from a dataset it learns what the subjects are by looking at their meta descriptions. When you request an image through a prompt such as ‘cat’. It doesn’t copy parts of other cat images and paste them together to create something new. During the learning process it has studied millions of pictures of cats, and now uses that knowledge to create a new cat from a noise pattern.
That doesn’t mean that it can’t copy styles however. The more images in the dataset had the description of a certain artist's name, the bigger the chance that the AI is able to output something in a similar style. It doesn’t copy and paste their style on an image, but it has learned the artist’s color schemes, brush technique, and so on, to create a completely new image from a noise pattern.
You can’t copyright styles. Styles are being copied all the time by humans. H.R. Giger studied the art of Ernst Fuchs, and their works are quite alike.
This doesn’t make it okay to pretend you are someone else, or sell works as if you hand painted them, while you actually used an Image Generating AI to create them.
Because of the ease of use, and popularity of this new software, it makes it easy for people with bad intentions to capitalize on that. AI Image Generating companies should definitely help to combat the people with these bad intentions, but in the end the person is responsible, and not the tool.
I’m cropping these for privacy reasons/because I’m not trying to call out any one individual. These are all Lensa portraits where the mangled remains of an artist’s signature is still visible. That’s the remains of the signature of one of the multiple artists it stole from.— Lauryn Ipsum (@LaurynIpsum) December 6, 2022
A 🧵 https://t.co/0lS4WHmQfW pic.twitter.com/7GfDXZ22s1
Famous tweet about signatures being visible in AI generated art.
Argument 2: I can see mangled signatures at the bottom corners of AI art, this must mean that parts of images must be stolen from artists!
This again is false.
If you request an image from the Image Generating AI with a prompt such as ‘Painted by’. ‘Oil Painting’, ‘Portrait by’, or ‘In the style of’. You can be sure that the images that the computer has trained on, probably had a signature in one of the corners. Because of this training, the Image Generator now thinks that when you request a painting by, the painting should have something scribbled in the corner. That is why they look mangled, and hard to read.
One exception is on overfitted images. For example, there might be many images in the database that show Starry Night by van Gogh. Because there are thousands of this same image in the database, the AI will know how to generate it very well. The AI might know this painting so well, that if you try to generate something with those words in the prompt, you might even see a resemblance of Van Gogh's signature somewhere.
This only happens when there are tens of thousands of images by the same artist in the training database, and won’t affect smaller creators at all. Most artists that have been trained on are underfitted. That means, even though their artwork might appear in the training database, the Image Generating can’t properly replicate their style of art when you request it in a prompt.
Image created using Midjourney AI. Prompt: someone reading a huge page of terms and services
Argument 3: The database contains my artwork without my consent. That's Stealing!
Now we are getting to an argument that needs a little more debate. Crawling websites and their content isn’t illegal. It is exactly what search engines such as Bing and Google do, to add their content to their search results.
Most art portfolio sites such as Art Station and Deviant Art, and social media such as Instagram and Twitter have terms of service. In these terms of service, they specify that content on their platforms are allowed to be crawled.
So the companies collecting the data for their dataset and the image generating AI companies aren’t technically breaking any current laws. However nobody expected that their images would be used in such a way. Although it is legal, it seems unethical, and the companies seem to realize that.
That might be one of the reasons why the data collecting companies, and the Image Generating AI companies are separate companies. For example the company that collected the 5 billion images in a set called LAION-5B, is non profit and claims it collects the images for research purposes. But some of the Image Generating AI companies, fund the data collecting companies, and once their computer has been trained on the data, are free to bring over that computer and use it for commercial purposes. These kinds of company structures can be found all throughout the corporate world, but in this particular case, one can assume that they use this structure to prevent liability.
So even though it isn’t technically stealing, I think there should be clear guidelines or even laws on how datasets can be collected if their purpose is for machine learning. I also think these datasets should be curated. It should be easy for artists to request their images to be removed from the databases, and if they want to be included in the databases, they should be reasonably compensated.
Argument 5: The technology is getting so advanced that it will cost artist’s jobs!
This is more of a philosophical question. But I'll share my opinion about this.
I can’t predict the future, but I definitely see that Image Generating AI’s will become part of a design or art company's workflow.
New technology has always been changing the workplace. When 3D animated movies became famous, people thought that traditionally animated movies would disappear, but beautiful traditionally animated movies are still released every year.. When photography became popular, people thought that it would ruin art and nobody would paint anymore, but an art career is currently more popular than ever.
Workplaces will evolve with new technology, some jobs will go, new ones will arrive. Traditional art will never lose its value. People buy paintings from their favorite artists because they relate to them, because they can communicate with them, and because they love to see their work process,
Image created using Midjourney AI. Prompt: a living banana with a face, taped to a white wall with ducttape, framed with a wooden frame
Argument 6: AI Art isn’t real art! The computer made it, not the one typing the prompt!
This again is more of a philosophical statement.
I think art can’t be judged by the amount of time or skill that it took. Why someone would pay $120,000 for a banana taped to a wall is beyond me, but someone thought it was art and that it was worth the money. There are plenty of examples of human made art that took little to no effort.
I think the biggest problem with calling an AI Image ‘art’, is that it can resemble something that a human being would have cost many hours and years of training to create. Therefore I think it is only fair that when you publish an AI Image as art, you clearly state that it was made with the help of Image Generating AI's.
I also agree that even if the prompter has a very unique idea, the machine still does the heavy lifting. So if you use Image Generating AI to create art, I suggest that you don’t call yourself the artist, but use another term, such as prompter or prompt crafter.
The lesser known issues concerning Image Generating AI’s
The main problem is, as discussed before, the way that training data is collected. But there are more issues that are often overlooked.
Racial bias in image generation. Image created using Midjourney AI. Prompt: Pretty 30 year old female
Overfitting and underfitting
Because the datasets basically consist of a slice of the internet's images as a whole, the images generated by the machine resemble that. That is why if you use the word ‘CEO’ in a prompt, you mostly get images of white males in suits, simply because that is how the majority of the internet's images display CEO’s. This is an example of overfitting. When a machine learning model has become so attuned to a certain type of data, that it loses its use if you try to use the data in any other way. This causes the Image Generating AI’s to be racial biased, and amplify the unrealistic beauty standards that can be found on the internet.
This is one more reason why I think that curated and ethically gathered training data is THE most important change that we need for Image Generating AI’s.
Open Source Problems
Another issue is the misuse of open source software. Most of the Image Generating AI’s such as Midjourney and Dall-E 2 have filters that prevent the creation of adult content and gore.
Stable Diffusion has a similar filter, but it is open source software. This means that the code of the software is publicly available to be used and altered by other software developers.
One such developer started a Kickstarter with the name of Unstable Diffusion. Their aim is to create an Image Generating AI that is unrestricted and has no filters.
While of course the responsibility lies with the creator, and not the tool, making a tool that makes it so incredibly easy to create nefarious content is dangerous and should be regulated.
Machine Learning (AI) tools are already being used in many fantastic ways such as self-driving vehicles, fraud detection, and healthcare. I can only imagine how much this technology will impact our future.
Image Generating AI’s have been in development for the last twenty years. But only now that the results are similar or better than what a human can create, artists are realizing that they will be confronted with it in one way or the other.
The development is both exciting and frightening, but the cat is out of the bag, and the technology will continue to evolve. All we can do is to make sure that it develops in an ethical way.
It is okay to use the tools, but understand how they work and what their limitations are.
- Don’t make copies of someone else's artwork and present them as your own.
- Even though your results may look like traditional artwork, make sure that you explain that results were created with the help of an Image Generating AI.
- Educate others on the subject and try to prevent the spread of misinformation
Thanks for reading!
Jan Loos (December 2022)
Have I been trained - Check if your work has been used in a 5.8 billion image size dataset.
Washington Post Article - Interactive article that explains how Image Generating AI's work.