The Fascinating development of AI: From ChatGPT and DALL-E to Deepfakes Part 2
Looking back on the decades long history that got us here
This is part 2 of the history of ChatGPT, DALL-E, and Deepfakes. If you missed part 1 you can read it here.📌
On January 5, 2021 OpenAI introduced Dall-E to the world. In the blog post they described their creation as…
A 12-billion parameter version of GPT-3 trained to generate images from text descriptions, using a dataset of text–image pairs.
The paper which describes the process in more detail highlights the insane amount of data these models require.
We created a dataset of a similar scale to JFT-300M by collecting 250 million text-image pairs from the internet. As described in Section 2.3, this dataset incorporates Conceptual Captions, the text-image pairs from Wikipedia, and a filtered subset of YFCC100M
250M images! Assuming three megabytes per image that’s a whopping 750 terabytes of pictures. That’s quite a lot of data. The training has paid off, as DALL-E has become one of the most popular tools for generating AI art. But with the Cambrian explosion of AI art a lot of questions around copyright, ethics, and the automation of human creativity have arisen. But before we get into that…
A little History 📚
While the recent examples of AI art mark a breakthrough in the field, the quest to generate art with computers has been measured in decades. In 1968, the Museum of Contemporary Art in London hosted the Cybernetic Serendipity exhibit, showcasing groundbreaking art, and music generated by computers. In 2007, The Painting Fool won the British Computing Society Machine Intelligence Award for its ability to generate portraits based on the emotions of human faces. In 2015, Google released Deep Dream, which used Convolution Neural Networks (CNNs) to create psychedelic images from photos.
Since then, there has been a surge in the number of websites and apps capitalizing on these technologies. Examples include Lensa AI, an app that allows users to create unique profile pictures by uploading photos of themselves, and Midjourney, which allows AI art to be generated through text prompts to a Discord bot. Depending on when you start counting, people have been working on this for between 20 and 50 years.
But with GPU compute becoming cheaper and more accessible the advances seem to be coming faster and faster. DALL-E 2 was released a little over a year after DALL-E which huge improvements to the model.
DALL·E 2, generates more realistic and accurate images with 4x greater resolution.
But this begs the question. What tools count in the debate about AI art? From Photoshop to Blender, and Inkscape to Krita there are plenty of tools that can be used to create art on the computer. Some of them even use machine learning to perform the transformations of images. According to the Harvard Journal of Law & Technology There are three key points.
Machine learning is required to develop the algorithm.
The algorithm must be trained on existing works. Not “from scratch”
The outputs are novel and surprising
The first two are pretty self explanatory, but the third one requires some explanation. Novel in this case means that the generated work is not an existing work, and surprising means that it doesn’t just take the inputs of the work and manipulate it in a way that can be understood, like rotating every pixel 90°.
The Abominable Intelligence 😯
But with these new technologies come a host of questions. “Are the people who create clever prompts to generate the impressive art artists?”, “How do we evaluate the quality of AI art?”, and “Will AI art displace human artists as the models improve?”. The last question weighs heavy on the minds of many aspiring and professional artists.
ArtStation a popular websites for artists to professionally display their work, sent this email to all of its users on 12/16/2022
we have introduced a “NoAI” tag. When you tag your projects using the “NoAI” tag, the project will automatically be assigned an HTML “NoAI” meta tag. This will mark the project so that AI systems know you explicitly disallow the use of the project and its contained content by AI systems.
We have also updated the Terms of Service to reflect that it is prohibited to collect, aggregate, mine, scrape, or otherwise use any content uploaded to ArtStation for the purposes of testing, inputting, or integrating such content with AI or other algorithmic methods where any content has been tagged, labeled, or otherwise marked “NoAI”.
The new NoAI tag gives artists control over how their art is used. But some artists have lamented that ArtStation did not take a stronger stance against AI art. In a show of solidarity against AI art, many artists on ArtStation uploaded pictures of the word “AI” stamped out with a red crossed circle. A quick search for No AI shows the result. So why are artists protesting AI art instead of embracing it? The reasons are varied.
A lot of the models used to create AI art have been trained on images that the researchers did not have the rights too. While it can’t be proved that DALL-E was trained this why, one of it’s sources The YFCC100M dataset does contain images that have some rights reserved, including excluding the image from being used for commercial use. They do mention in the training section using a “filtered subset of YFCC100M” so it’s entirely possible this subset filtered out those licenses. I’ll give them the benefit of the doubt, but other platforms that offer similar service don’t have as stringent filtering, requiring artists to manually submit DMCA requests to resolve copyright issues.
Another reason is that art and other creative professions are traditionally underpaid and undervalued. From the VFX artists to Music Composers, brutal deadlines, unreasonable expectations, and a lack of understanding of the creative process create hostile environments for these professionals. In a recent example art made using AI won the top prize at the Colorado State Fair. These competitions aren’t just about winning prize ribbons, they also increase name recognition, and builder resumes. Many artists struggle to make minimum wage and competing against AI makes that even more challenging.
Finally artists have been seeing increasing levels of harassment in social media spaces. Some artists claim they are being tagged in pictures of AI art created in their style saying they will be replaced by this next generation of AI art. Others are seeing art that was generated with prompts like “Generate an image in the style of <Insert Artist>” being sold as prints. For many artists the tools and techniques to generate AI art have far outpaced any regulations surrounding copyright, and rules in competitions.
Can AI art be a good thing? 🤖
Is it possible for AI art to be used to improve the artistic process? It remains to be seen. But some possibilities are… Helping artists workshop an idea. Many artists create mood and idea boards using PureRef or Pinterest and AI art could speed up that process. AI art could also be used to generate new insights into art. Creating complex prompts with potentially conflicting themes could spark inspiration for new artwork. On the flip side those without artistic ability could generate images using AI that could help the artist more closely realize the intended vision, allowing for a deeper more collaborative effort.
There are still many unanswered questions about how AI will fit into human creativity, and what its impact will be on the artists community. Regulations and rules are catching up but progress is slow compared to the speed at which these models are updated. I believe we reached a critical junction in the explosion of AI related software, but I’ll save that for another article 😉
Stay Tuned 📻 for Pt3
Thanks for making it to the end of my article. If you enjoyed considering liking and subscribing! I will be releasing part 3 soon, but in the mean time why not check out another one of the articles I’ve written?
Artists say AI image generators are copying their style to make thousands of new images — and it's completely out of their control