A Discerning Look at the Power of AI Image Generators
The new wave of AI image generators like DALL-E 2 and Midjourney are capable of inspiring wondrous creativity as well as nefarious misuse. In part two of a series exploring the power of these tools to enhance creative expression and exploit AI as an art form, we take a closer look at the dangers and risks of their potential abuse, how they might impact the art trade, and what to do about them.
Fakes and Harmful Images
According to a Dec 2022 Neiman Lab report, “Deep fakes have already been used to create nonconsensual pornography, commit small- and large-scale fraud, and fuel disinformation campaigns.” It notes that powerful image generators could exacerbate these misuses. Companies that have developed AI image generators are aware of their power for misuse and are taking steps toward mitigating the potential abuses.
For instance, Midjourney announced this year it would discontinue the free version of its AI image generator to cut down on the proliferation of high-profile deepfakes being created by the tool (such as images of Donald Trump being arrested or Pope Francis wearing a trendy coat). While some AI-generated images are fantastical and thus obviously deepfakes, a recent test performed by Tidio found that in some survey groups, 87% of respondents mistook an AI-generated image for a real photo of a person, pointing to an alarmingly low level of general visual literacy among the public. Google adopted an even stricter stance and chose not to release its text-to-image technology, mindful of the potential risks. OpenAI initially provided limited access to its technology, implementing guardrails to prevent the generation of explicit or harmful content. Over time, it has expanded accessibility and added features while maintaining certain restrictions. Stability AI, in turn, took a different approach by releasing its Stable Diffusion product without guardrails, relying on the ethical responsibility of users. However, after discovering cases of users generating child abuse images, the company removed the tool’s ability to render images of NSFW content and children.
To adequately prevent the misuse of AI image generators, efforts by the industry to restrict their access or constrain their usage by imposing certain limitations are not enough. The industry needs to explore forensic techniques to distinguish authentic images from fakes. Digital forensic analysis backed by machine learning patterns can help to detect and uncover hidden evidence in digital artworks that can be overlooked if conducted manually.
Licensing, Legal & Ethical Questions
Artwork licensing is another big problem related to AI image generators. The training data behind those generators often include the work of artists who have yet to license their work. As Ron Cheng, Yale Visual Arts Collective board member, puts it, “All of this [AI] art is taken without the consent of these artists, and the laws that exist are not really protecting them.”
For this reason, in January 2023, three artists filed a lawsuit against top companies in the AI art generation space, including Stability AI, Midjourney, and DeviantArt. The artists claimed that the companies were using copyrighted images to train their AI algorithms without their consent.
Andres Guadamuz, a UK academic who focuses on AI and intellectual property law, suggests that there are several key questions from which many uncertainties of the ‘AI vs. copyright’ space unfold. For instance: Can you copyright the output of a generative AI model, and if so, who owns it? Or, if you own the copyright to the input used to train an AI, can you file a legal claim over the model or the content it creates?
Currently, there is no copyright protection for works generated solely by a machine in the US. However, copyright may be possible in cases where the creator can prove there was considerable human input. Dall-E and Midjourney, for example, are giving premium subscribers commercial rights to the images they create.
Establishing copyright is, however, only a first step in preventing the inclusion of artwork in AI training data without artists’ consent. Andres Guadamuz says that copyright registration is, of course, essential for suing someone for copyright infringement. Still, ultimately it is up to a court to decide what is legally enforceable.
For now, as one of the first steps towards legal regulation, EU policymakers deliberated the AI Act, the world’s first legislation to regulate AI. Artist associations are mobilizing to introduce a specific section in the Act dedicated to the creative arts, including safeguards requiring that rightsholders give explicit informed consent before their work is used.
Along with a legal and regulatory framework to deal with the legal and ethical implications, the industry should come up with tools to provide artist attribution and provenance for AI-generated artwork. For instance, a group called Spawning has already launched a tool, Have I Been Trained, which allows artists to see if their images have been used to train AI systems.
Curtailing the Influence of Social Bias
In 2021, researchers Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe analyzed a LAION-400M training dataset like the one used to build Stable Diffusion and found that AI training data is filled with racial stereotypes, pornography, and explicit images. Most other AI image generators, such as Google’s Imagen and OpenAI’s DALL-E 2, are not open but are built likewise, using similar training data, which suggests that this is a sector-wide problem.
How does social bias, baked into the training data, carry through to the outputs of AI image generators? In December 2022, Melissa Heikkilä, a senior reporter at MIT Technology Review, shared her research and insight into the gender-centered bias seen in the AI avatar app Lensa. After trying Lensa, Melissa’s avatar results were “cartoonishly pornified, while [her] male colleagues got to be astronauts, explorers, and inventors.”
Here’s another example: Sasha Luccioni, an AI researcher for Hugging Face, created a tool to spot how the AI bias in text-to-art generators works. Using the Stable Diffusion Explorer as an example, when prompted by the phrase “ambitious CEO,” the tool returned results only for different types of men, while results from the phrase “supportive CEO” included both men and women.
Considering biases around professions’ categories, a fascinating video posted to Reddit entitled “What Midjourney thinks professors look like, based on their department” reveals another layer of insight into inherent biases held by AI image generators.
What is the solution to tackling inherent biases in training data? First and foremost, training sets must be diversified or curated to negate social biases, or algorithms must be written to recognize and account for those biases during the training process. For example, Prisma Labs (Lense app creators) say they have adapted the relationship between certain words and images to reduce biases, while the creators of the LAION dataset have introduced NSFW filters, which use computer vision algorithms to detect and filter out not-safe-for-work images from the web browser.
Does AI-Generated Art Qualify as Real Art?
What is the true nature of art? When thinking about this question, two essential characteristics might come to mind: authenticity and originality. Both arguably come from the unique expression, or imagination, of the person creating the art. We commonly think of art as those forms of expression that come from someone’s emotions, from their essence, and that we can relate to on a human level.
But what happens when an emotionless machine generates art? Machines, or algorithms, have no authenticity or originality. Some traditional collectors and art critics believe that the AI-generated output lacks “soul” and does not, therefore, qualify as art in any sense we have previously understood. While the AI-generated artwork hardly represents intentional art, it is nonetheless defined by the intention of the machine’s human user.
Nisheeth Vishnoi, the A. Bartlett Giamatti professor of computer science and co-founder of the Computation and Society Initiative at Yale University, compared AI-generated art to asking ChatGPT to write a short poem. While it is quite possible the chatbot will write a better poem than many people, this poem would not necessarily be publishable or a candidate for a Nobel Prize in literature, according to Vishnoi.
From the perspective of some artists, the greatest harm posed by AI-generated art is damage to their reputation. “What’s more important is the utter disrespect these AI ‘artists’ promote against the community and art as a craft, which is already extremely undervalued in the modern day,” suggests Kim Lagunas, a student artist.
Carbon Footprint Concerns
New generations of AI models take literally millions of hours of computing time to train, consuming enormous amounts of energy in the process. In fact, AI uses more energy than other forms of computing, and training a single model can consume more electricity than 100 US homes would in an entire year. Constituting a subset of those models, AI image generators suffer from the same energy-intensive, high resource costs, and large carbon footprint.
One of the bigger mysteries in AI is the total accounting for carbon emissions associated with the chips being used. Nvidia, the largest producer of graphics processing units, tackles this issue by completing AI-related tasks faster, making them more efficient overall.
There are other ways to make AI run more efficiently, according to Ben Hertz-Shargel, energy consultant at Wood Mackenzie. Ben Hertz-Shargel believes that developers and data centers could schedule the AI training process at downtime periods when power is cheaper or at a surplus, thereby making their operations more optimized and greener.
Fears of AI Generators Replacing Artists
Some critics of the new wave of AI technologies point to the existential question of what they portend for the already precarious livelihoods of artists and designers. The optimistic view is that they can add to human creativity by automating the mundane side of graphic design work and artistic production, allowing artists more space to focus on their creative talents.
For example, Anne Ploin, a member of a group of specialists investigating how AI will affect creative industries at the Oxford Internet Institute, concludes in her report “AI and the Arts. How Machine Learning Is Changing Artistic Work,” that AI art generators will not replace human artists. “While ML models could help produce surprising variations of existing images, practitioners felt that the artist remained irreplaceable in giving these images artistic context and intention — that is, in making artworks,” Anne states.
Others are more fearful. Some artists, like Paul Chadeisson, have complained about explicit plagiarism using AI models. In April this year, illustrators were taken aback by the official release of the first manga drawn entirely using AI technology, and several of them involved in game designs for Chinese gaming companies are concerned about AI causing a decline in illustration jobs. And the fear of these technologies adversely affecting people’s livelihoods goes far beyond the art trade. OpenAI chief executive Sam Altman shared in his blog that while AI will create new jobs,” I think it’s important to be honest that it’s increasingly going to make some jobs not very relevant.”
In short, will these AI systems eventually replace professional artists entirely? The immediate response is probably not, as there will always be a need for artists to “curate” AI-generated artwork, even with the seemingly unstoppable advancements in AI software. The more likely outcome is that these technologies will support people in their artistic endeavors in new ways and accelerate and facilitate the creation of high-quality artworks by artists.
Final Thoughts
Much like other AI software tools like ChatGPT, AI image generators pose a whole string of legal and ethical challenges and risks of abuse and misuse. While there are many reasons why these tools should be embraced and not rejected outright, their potential abuse and the legal and ethical risks they present must be carefully considered and proactively mitigated using a combination of technical and regulatory measures for their widespread adoption to succeed.
Author: Doron Fagelson,
Vice President of Media and Entertainment Practice at DataArt
Originally published on https://www.dataart.com/blog/
