Text-to-image AI models are almost perfect for creating instant dopamine hits in your brain. By entering any text prompt you wish, they can generate an image based on that theme within a few seconds. And while many users are using AI to generate more creative selfies, these models have become quite good at replicating images in the style of real (aka human) artists.
For example, last year, an AI-generated art piece won an art competition at the Colorado State Fair. Most recently, Adobe announced it would permit users to sell AI-generated artwork on Adobe Stock — as long as they are labeled.
Deep learning models like Stable Diffusion and DALL-E, two popular text-to-image generators, are trained on massive datasets which are scraped from thousands of domains. The AI pulls images from websites typically used by artists — such as WordPress, Blogspot, Pinterest, Shutterstock, and Getty Images — to best mimic different artistic styles. But as these models are trained to become increasingly more sophisticated, new concerns are emerging in the creative community.
“This is a hyper-modern reincarnation of copyright issues. In the past, people complained against image-hosting websites because their art was taken, distributed, and sold without their consent,” said Abraham Liddell, a postdoctoral researcher at DSI. “AI art generators are making the same old mistakes again but at a much larger and faster pace than ever before. Artists should have the right to decide whether or not they want their art to be included in these AI training models. But while developers have been rushing to launch their incredible text-to-image-AI models, little to no regard is being given to these ethical issues,” added Liddell.
So far, AI developers and tech companies alike have justified using images that artists have published all over the internet. They have argued that, at least in the US, these images fall under the fair use doctrine — which promotes freedom of expression by allowing the unlicensed use of copyright-protected work. Meanwhile, regulatory and policy efforts are lagging behind in figuring out ways to govern the ethical use of AI technology.
Text-to-image models are groundbreaking technology. How can the data science community better address these concerns and find a path forward? According to DSI experts, one promising tool for making proper attribution and accountability a reality is blockchain. “This comes down to the issue of establishing provenance,” said Bruce M. Kogut, Sanford C. Bernstein & Co. Professor of Leadership and Ethics at Columbia Business School and Professor in the Department of Sociology. “That means an artist has an ownership stake in whatever is being done to their work. Blockchain is an exciting technology that can help artists/creators establish provenance.”
Establishing artistic provenance through blockchain
As a type of decentralized database (or ledger), blockchain can collect and create an immutable record of information across a network of computers.
“Blockchain can allow artists to create a cryptographic log of their work. An artist can then refer to a specific ledger entry to prove ownership of their work,” explained Susan McGregor, an Associate Research Scholar at DSI and co-chair of DSI’s Data, Media and Society Center.
While promising, an ongoing challenge is that even if an artist can prove ownership of an image, their work could still end up becoming training material for a Generative Adversarial Networks (GAN) training model. And many of these training models are AI black boxes, making it difficult to determine which images have been used in them.
“It will be an uphill battle for an artist to prove that an AI system used their work inappropriately to generate lookalike work,” said McGregor. “To hold AI developers accountable, there needs to be an effective membership test to determine if an AI art piece is similar enough to an artist’s original work to be considered a “copy.” Although a blockchain solution could prove someone owns the original image, proving a similar image violates their rights is much harder.”
In addition to existing intellectual property laws, creators might be able to protect their work by opting for a watermarking process that could prevent GANs from training on those images, McGregor added. Or at the very least, watermarks could make it hard for AI models to effectively mimic artists’ content. Recently, researchers at the University of Chicago released a system to help artists do precisely that: the Glaze Project allows artists to cloak the images they upload so that AI systems can’t emulate their artistic style as effectively. Combined with a cryptographically authenticated log of the artists’ original works, such systems can strengthen the protections for artists in the age of AI reproductions.
Fighting against fake news with blockchain
Beyond artistic provenance, blockchain could also potentially help in tackling the safety issues surrounding free open-source text-to-image AI models. After all, nefarious elements can also use these models to generate fake news or disinformation at scale with limited resources.
McGregor argued that the widespread use of cryptographic signatures on images by artists and photographers alike could help address this problem as well. If any image produced by a human being has a cryptographic signature that cannot be tampered with — similar to barcodes on food items in a supermarket — when users encounter an image without any provenance information it could be seen as lacking credibility and be deemed as inauthentic.
“It could be a promising norm,” explained McGregor. “What this boils down to is questions surrounding governance and interoperability of that cryptographic ledger. There needs to be an organization that can supervise and govern the process for such a broadly distributed system to work well.”
To tackle these problems, the creative community, including artists and journalists, will need to come together with technologists to fully implement these solutions at scale. An interdisciplinary approach will be crucial. “For this to start happening, there needs to be a market solution where someone develops an affordable and accessible blockchain tool that artists can use. It is doable from a technical perspective,” concludes McGregor.
Contributing Writer: Anuradha Varanasi