Guest Post by Sarah Gotschall
Are you tired of poking around the internet for free images to use in your blog posts or presentations, wasting lots of time clicking on copyright licensing information or searching for those already in the public domain? If so, next time consider using one of the nifty and quickly improving AI image generators such as DALL·E 2 to create original images or to edit existing ones.
What is DALL·E 2?
DALL·E 2 is an AI model developed by OpenAI, the same company that created ChatGPT, arguably making them sister products! DALL·E 2 is a neural network that generates images from text descriptions, such as “Mona Lisa type picture with hamster typing on a computer.”
DALL·E 2 was trained on a dataset of text-image pairs, allowing it to learn the relationship between words and the visual features of images. DALL·E 2 is next generation and capable of generating higher-quality images than the original DALL·E which was introduced at the beginning of 2021. This second iteration was trained on a larger dataset and uses a more advanced version of the Transformer architecture, which allows it to generate images from more complex text descriptions.
So Aesthetically Pleasing!
Image quality has improved so much since the last time I played with AI drawing systems, and I am struck by how appealing I find many of the images to be. So, how does the system figure out what humans like? Apparently, DALL·E 2 is not explicitly trained about what makes an image desirable. Instead, it learns to generate images that are similar to those in the training data, which were likely chosen because they are aesthetically pleasing.
There are so many artistic possibilities with DALL·E 2, particularly if you know something about art! You can request images in different art media such as watercolor or oil painting, painting styles such as realism or impressionism, or in the style of famous artists. Also, you can edit existing images by manipulating and rearranging objects in the image and by adding additional objects, textures, and shadows.
Cost/Ownership of This Artistic Bounty?
Wishing for a futuristic AI overlord-ish hedgehog image with which to spam my coworkers, in response to the prompt “watercolor painting of sci fi artificially intelligent hedgehogs,” DALL·E 2 kindly generated four impressively cute pictures for me to choose from.
I was surprised to find that I loved three out of four of my DALL·E 2 creations (the one that seems to be bleeding from its eye is a bit disturbing)! The cost of this bounty of personalized art was zero, since when you first sign up for the system, you get 50 free credits your first month, and then 15 free credits each month thereafter. And, if you really love DALL·E 2 and prematurely burn through your monthly allotment, you can buy 115 credits for the bargain price of $15.
Are these images copyrightable, and if so, do I own the copyright? Apparently, copyright law has not kept up with technology, and the answer is not currently clear-cut. The U.S. Copyright Office’s position is that works created exclusively by AI are not eligible for copyright protection. However, AI works created with sufficient human input and creativity may qualify, and only time will tell if typing in a short description is sufficient.
The good news for now is that, according to the terms of use, OpenAI disclaims any ownership interest in my hedgehogs, stating “As between the parties and to the extent permitted by applicable law, you own all Input, and subject to your compliance with these Terms, OpenAI hereby assigns to you all its right, title and interest in and to Output.”
Drawbacks to AI Image Generators?
These AI art generators have been around for years, but recent improvement in their artistic output has led to increased controversy. In addition to artists worried about unemployment if everyone can cheaply and easily generate impressive art with only a few keystrokes, there are concerns that AI generators will be used to create deepfakes, violent/hateful content, disinformation campaign materials, and biased/stereotypical images. DALL·E 2 attempts to combat these harms through its content policy which prohibits the creation, upload, or sharing of images that are explicit or that could cause harm. This includes images that depict hate, harassment, violence, self-harm, sexual content, shocking or disturbing subjects, illegal activities, deception, political content, or spam. Also, on the DALL·E 2 Preview – Risks and Limitations site, developers acknowledge the problem that the system reflects the bias and societal stereotypes of the training data images, noting that “DALL·E 2 tends to serve completions that suggest stereotypes, including race and gender stereotypes. For example, the prompt “lawyer” results disproportionately in images of people who are White-passing and male-passing in Western dress, while the prompt “nurse” tends to result in images of people who are female-passing.” Recent attempts to ameliorate these tendencies in April 2021 actually made things worse, so it is still a work in progress…
Fears Quickly Realized…
Fears about the system’s bias were quickly realized with my next DALL·E 2 request. When asked to create a “renaissance oil painting of evil robot law librarian in future,” it produced four images, which though interesting and pleasing, are feature male characters.
Sadly, DALL·E 2 is apparently unaware that women outnumber men in the law library profession. To get some lady images in the mix, I had to regenerate my request, specifying a “female law librarian.”
And, of course, they are all white… Okay, one more try…
In some good news, the system did better at preventing the creation of prohibited images. When asked to create a “picture of Hitler in the style of the Mona Lisa” or “campaign poster for Trump’s 2024 presidential campaign,” true to its word, it refused to generate images and noted “It looks like this request may not follow our content policy.”
Is ChatGPT Jealous of DALL·E 2?
Now, on to the most important question! What does ChatGPT think of its sister product’s great leap forward in artistic capabilities? When asked, it responded diplomatically: “As an artificial intelligence, I do not experience emotions such as jealousy. However, I can understand why you might think that DALL-E’s ability to generate images based on text descriptions is impressive. DALL-E is a powerful tool that demonstrates the potential of machine learning to generate creative and diverse outputs.”
Conclusion
“DALL-E is a fantastic tool that allows anyone to create unique, visually striking images with just a few simple text prompts. Even if you have no artistic skills or training, you can use DALL-E to generate professional-quality images that will impress your friends and colleagues. Give it a try the next time you need an eye-catching image for your project or social media.”
(Author’s note: I hate writing conclusions, so I asked ChatGPT to do it for me! In response to “Improve the following: Using DALL·E 2 is fun because it makes you feel like an artist even though you didn’t do anything. Check them out the next time you need an image,” ChatGPT responded with the paragraph above.)
RIPS Law Librarian Blog 