Week 04: Everyone’s an artist?

In last year’s Detox, when we wrote about generative AI and art, we noted both the uneven quality of image and video generation and the increasing sophistication of image and video generation tools. Since then, the image and video generation has improved markedly and the thorny issues related to AI-generated art have gotten…well, thornier. More than 20 major lawsuits are being pursued against AI companies for copyright infringement. Adobe’s Firefly was promoted as the ethical choice for AI image creation because it only trained on works that Adobe had license to use. It was then found to have used Midjourney images in its training data. Coke came out with an AI-generated commercial that met with fairly negative reviews. The winning photo in an AI photo contest was disqualified for being a real photo. The Concept Art Association’s Go Fund Me raised more than $285,000 to protect artists from AI technologies. $187,500 of that went to pay for a full-time D.C. lobbyist. Meanwhile, the United Nations hosted an AI artwork installation entitled Large Nature Model: Coral. This “testament to the beauty and fragility of our natural world” was “trained on a staggering 100 million coral reef images.” No mention was made of the energy required to train this model. It is an understatement to say we are living in very odd times.

These tools are so new that it can be hard to tell when the failure is in the technology (AI video creation doesn’t work!), a specific tool (Sora doesn’t work!), or the author’s skill in using a specific tool (I don’t know Sora well enough to make it work!). In this case, it’s likely a combination of all three. AI video creation works but there is still a long way to go. We’re also using a pretty basic tool that has only recently been made available–more advanced and specialized tools (see also, higher cost) are likely able to produce higher-quality work. But, even with a basic tool, it’s pretty amazing what you can create. It also raises so many questions.

Ethical considerations

At the heart of many concerns surrounding AI is the issue of intellectual property rights– specifically, the rights of individuals whose work is included in training datasets. Even if courts rule that transforming images, sounds, videos and other creative works into mathematical weights is legally permissible, the ethical question remains: Is it right?

Some AI tools, like Adobe Firefly, claim to use ethically sourced training data. Others, like Mickey-1928, operate in a gray area by fine-tuning an existing model (Stable-Diffusion-xl) on public domain content (stills from the earliest depictions of Mickey Mouse). In response to complexity like this, Fairly Trained is now offering certifications that verify AI models were trained ethically and specifying what that means. Their list of certified models and tools is growing. On the other end of the process, The Dataset Providers Alliance, is creating a system to let content owners control if, and get paid when, their content is used to train AI models.

Many AI models can now mimic specific artistic styles with alarming precision. Using an artist’s own work to generate imitations– especially while corporations profit massively–feels fundamentally unjust. This raises serious concerns about consent, attribution, and fair compensation. Current copyright law was not designed with AI in mind and new precedents will need to be established.

Why do we hate AI-generated art?

That “strong reaction against AI art” is no joke. Studies (like this one) have found that humans consistently rate works of art more negatively when they believe that the works are AI-generated. Interestingly, when participants did not know the origin of the art they often rated AI-generated pieces as favorably or even more favorably than human-made pieces. This bias against AI-generated art appeared strongest when participants evaluated factors like monetary value and artistic talent, but was less pronounced for qualities like emotional impact or aesthetic appeal. In other words, people seemed more hesitant to pay for or acknowledge technical proficiency of AI-generated art, but they seemed more open to appreciating its emotional or aesthetic value.  Unsurprisingly, exposure to AI art may enhance how people view human creativity–when comparing AI and human-created works side by side, people tended to rate human creativity more highly. These reactions echo long-running debates about authenticity and technology in art, from the introduction of photography, to Andy Warhol’s The Factory, to today’s digital editing tools. The question isn’t just whether AI can make art, but how our perception of art changes when we think machines are involved in its creation. Those perceptions of AI-generated art may have serious implications for what role AI-generated art can play in our world. As we saw with the AI-generated Coke advertisement fiasco, if our reactions to AI-generated content are adverse, we may miss out on messages because of how they were created. Lim and Schmälze’s study demonstrated this risk directly–people were more likely to dismiss potentially valuable health information if they knew AI had written it, even when the content was high quality. This suggests our bias against AI-generated content could prevent us from benefiting from otherwise meaningful work.