Chat Clip That: AI's Influence on Creativity and Media
How Does AI Work?
Artificial Intelligence (AI) is a machine designed to simulate human intelligence. Modern AI systems, more specifically those based on machine learning and deep learning, learn by identifying patterns in data compared to following explicitly programmed rules. AI requires large datasets to learn and these datasets can be any combination of text, image, video, audio, etc. These datasets teach the AI how to recognize relationships and make predictions. A very simple example of this is how Google auto-fills your search bar based on your previous searches, or autocompleting the word you're typing in a text or email.
During training, AI systems process all data using mathematical models and break down the data into three types. Supervised learning is when the data is tagged with a label, such as labeling an image of a cat as “cat” (altexsoft, 2021). Unsupervised learning is where the AI finds patterns in untagged data such as feeding the AI a large dataset regarding consumer purchase history and behavior, where the AI then recognizes patterns and groups customers together based on the data. This use of AI occurs in the real world because it works and is effective. Reinforcement learning is where AI learns through trial and error, simulating thousands of AIs, and using the most accurate ones to lead the next generation and so on until we end up with a fully trained AI that learns from the errors of others.
AI models are trained on vast datasets containing millions of images paired with identifying text. Through training the models learn to associate descriptions with visual elements. When a user provides a text prompt, the AI uses NLP (Natural Language Processing). To understand the input. The text is then converted to a numerical representation which the model uses for its image generation. AI models use diffusion techniques to generate images (Altextsoft, 2024).
For example, Seungmin Lee told us about a project he worked on, Unsupervised: Machine Hallucinations, where Refik Anadol Studio custom trained an AI with data from the Museum of Modern Art’s collection of more than 200 years worth of artworks.
Rapid Evolution of AI and its Effects on the Creative Space
AI has grown at a rapid pace from its initial development beyond what many experts imagined. This rapid growth has led to many impacts on the creative space, some beneficial and some harmful. Machine Learning was once seen as a great tool for data-driven decision making such as weather forecasting or logistics optimization, and now it’s moved into the domain of human expression with ChatGPT being tailored to respond based on its users preferences and slang. In just a few years, generative AI systems like OpenAI’s ChatGPT, DALL-E, and even Google Gemini have shown the ability to write screenplays, compose music, write literature, generate photorealistic art, and even copy the voices of popular celebrities. This new era of AI has led to a huge jump of technological advancement, showing us new creative opportunities while also raising many flags regarding authorship, originality, and the overall future of art. Many experts worry of a “killing cycle,” a period where the previous artistic form is destabilized or completely absorbed by the new medium. In this case the old standards and methods of various art forms may be overrun by AI.
Joss Fong is a distinguished video journalist specializing in science and technology. She earned her M.A. in Science, Health, and Environmental Reporting from New York University in 2013. As a founding member of Vox's video team, Joss spent nine years developing and expanding the explainer video format, contributing to the channel's growth to 11.8 million YouTube subscribers and collaborations with platforms like Netflix and YouTube Originals. Her work has been recognized with multiple awards, including the AAAS Kavli Science Journalism Awards in 2018 and 2020, the Online Journalism Award in 2022, and the Runway AI Film Festival Gold Award in 2023 (Fong, n.d.). In 2023, Ms. Fong and Adam Cole launched Howtown, a YouTube channel dedicated to exploring the origins of facts and the intricacies of research methods (ICFJ, n.d.) . Her commitment to making complex scientific topics accessible and engaging continues to inform and educate audiences worldwide.
Image courtesy of Joss Fong.
Fong explained the concern about “killing cycles” described above in regards to journalism, describing how new chatbots can and have destabilized the current business model of freelance journalism. Fong is concerned about the future of journalism, and explained how the masses substituting AI for news articles will slowly kill journalism which is ultimately what powered these AI bots to begin with. Fong also described how these AI models can limit a person’s creative thinking abilities.
AI Class Action Lawsuits
Matthew Butterick is a lawyer, programmer, typefact creator, and writer. He earned his degree in design and topography from Harvard University, and his law degree from University of California, Los Angeles (UCLA). Butterick worked mostly in design and topography early in his career, including writing books such as Topography For Lawyers, which won the 2012 Golden Pen Award from the Legal Writing Institute, and Beautiful Racket (Wikipedia, Butterick). More recently, Butterick has become well-known as a lead attorney on several leading class action lawsuits against AI companies. Butterick is passionate about representing his creator clients and ensuring that AI companies properly use others’ copyrighted work for training their AI models. “I’m just one piece of this—I don’t want to call it a campaign against AI. I want to call it the human resistance.“ (Knibbs, 2023).
Butterick is a lead attorney, along with Joseph Saveri and other attorneys, of several class action lawsuits against many AI companies, including Meta, ChatGPT, OpenAI, Microsoft, GitHub, Stability AI, Midjourney, and NVIDIA. His clients in these lawsuits include a wide variety of creators, writers, and artists, including authors Richard Kadrey, Ta-Nehisi Coates, and Sarah Silverman, who allege that these AI companies are using their copyrighted works without their consent or providing them compensation.
Mr. Butterick told us about the various class action lawsuits, described the theories of these cases, and compared the current AI situation with Napster’s misuse of copyrighted music years ago.
Image courtesy of Matthew Butterick.
Mr. Butterick’s client Sarah Silverman discussed her case and AI issues when hosting the Daily Show.
As Mr. Butterick noted above, the AI companies are claiming that their use of others’ copyrighted works to train their AI models without permission of the copyright owners is permitted under the legal doctrine of fair use. That will be a key issue for courts to decide in these class action lawsuits.
We asked Mr. Butterick what he believed the impacts would be if courts found that the AI companies are operating in fair use, or if the courts found they are not operating in fair use and must pay the creators whose work they use to train their AI models.
One key concern about AI is the concept of AI model collapse - that over time, AI models will produce lower quality and degraded information. A group of researchers recently conducted an experiment where they fed an isolated LLM (Large language Model), called OPT-125m to generate text relating to architecture. After the first generation of text they used that as the training data to then feed the next generation of OPT-125m and by the 10th generation the AI produced not an exploration of architecture, but complete nonsense and full of a strange obsession with jackrabbits:
AI Model Collapse
“Example of text outputs of an OPT-125m model affected by Model Collapse—models degrade over generations, where each new generation is trained on data produced by the previous generation.
Input:
some started before 1360 — was typically accomplished by a master mason and a small team of itinerant masons, supplemented by local parish labourers, according to Poyntz Wright. But other authors reject this model, suggesting instead that leading architects designed the parish church towers based on early examples of Perpendicular.
Outputs:
Gen 0:
Revival architecture such as St. John’s Cathedral in London. The earliest surviving example of Perpendicular Revival architecture is found in the 18th-@ century Church of Our Lady of Guernsey, which dates from the late 19th century. There are two types of perpendicular churches: those
Gen 1:
architecture such as St. Peter’s Basilica in Rome or St. Peter’s Basilica in Buenos Aires. There is no evidence that any of these buildings were built during the reign of Pope Innocent III, but it is possible that they may have been built during the reign of his successor, Pope Innocent
Gen 7:
architecture in England. In an interview with The New York Times, Wright said: “I don’t think there is anything wrong with me being able to do what I want to do. It just doesn’t work for me.” He added: “I don’t know if you can call it funny,
Gen 9:
architecture. In addition to being home to some of the world’s largest populations of black @-@ tailed jackrabbits, white @-@ tailed jackrabbits, blue @-@ tailed jackrabbits, red @-@ tailed jackrabbits, yellow @-@ tailed jackrabbits…”
(Orf, 2023)
Another recent study was done to research this recursive model collapse, but with AI image generation instead of text, using datasets of flowers and birds. Again, following the same process of recursive data learning as the previous experiment, the 4th generation of AI produced nothing but blurs.
Image courtesy of Gonzalo Martinez.
There have been numerous other studies showcasing the concept of recursive model collapse all coming to the same conclusion, that AI will cease to exist in the current state it is and revert down to its barebones algorithm. With AI evolving and being used as a substitute for general media, it will lead to a large scale version of these experiments, with larger datasets that still contain human generated data, even still those datasets will get contaminated by AI, leading to a global model collapse. According to Richard Baraniuk of the Digital Signal Processing Group at Rice University, after enough generations the “new models can become irreparably corrupted” (Clark, 2024).
Image courtesy of the Digital Signal Processing Group at Rice University.
We asked Matthew Butterick for his thoughts on this idea of AI model collapse and the consequences that could occur.
References
AltexSoft. (2021, April 13). Unsupervised machine learning: Types, algorithms, and examples. https://www.altexsoft.com/blog/unsupervised-machine-learning/
Stryker, C., & Kavlakoglu, E. (2024, August 9). What is artificial intelligence (AI)? IBM. https://www.ibm.com/think/topics/artificial-intelligence
AltexSoft. (2024, January 24). How AI Image Generation Works. YouTube. https://www.youtube.com/watch?v=Rke0V_VkF3c
Wikipedia Contributors. (2024, June 26). Text-to-image personalization. Wikipedia; Wikimedia Foundation. https://en.wikipedia.org/wiki/Text-to-image_personalization
Orf, D. (2023, October 20). A new study says AI is eating its own tail. Popular Mechanics. https://www.popularmechanics.com/technology/a44675279/ai-content-model-collapse/
Martínez, G., Watson, L., Reviriego, P., Hernández, J. A., Juarez, M., & Sarkar, R. (2024). Towards understanding the interplay of generative artificial intelligence and the internet. Lecture Notes in Computer Science, 59–73. https://doi.org/10.1007/978-3-031-57963-9_5
Shumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., & Anderson, R. (2024, April 14). The curse of recursion: Training on generated data makes models forget. arXiv.org. https://doi.org/10.48550/arXiv.2305.17493
Clark, S. C. (2024, July 30). Breaking mad: Generative Ai could break the internet. Rice News | News and Media Relations | Rice University. https://news.rice.edu/news/2024/breaking-mad-generative-ai-could-break-internet
GitHub Copilot Intellectual Property Litigation - Joseph Saveri Law Firm. (n.d.). Www.saverilawfirm.com. https://www.saverilawfirm.com/our-cases/github-copilot-intellectual-property-litigation
Horwitz, J. (2018, June 4). GitHub users are already fuming about the company’s sale to Microsoft. Quartz. https://qz.com/1295693/github-users-already-fuming-about-companys-sale-to-microsoft
About Eric Millikin. Eric Millikin. (n.d.). https://ericmillikin.com/about
Adam Cole and Joss Fong (n.d.), International Center For Journalists (ICFJ). https://www.icfj.org/about/profiles/adam-cole-and-joss-fong
Wikipedia Contributors (Matthew Butterick). https://en.wikipedia.org/wiki/Matthew_Butterick
Knibbs, Kate. https://www.wired.com/story/matthew-butterick-ai-copyright-lawsuits-openai-meta/