Generative AI is a form of artificial intelligence that can be used to create new images, video, audio, code, synthetic data, or text. Generative AI tools are built on supervised and unsupervised algorithms that constantly learn as they create more and more output. The term “generative” means that the tools learn to create new data, like text or images, instead of just learning to recognize and label something. For example, a generative AI tool can create an image of a face, a step beyond just recognizing the face and labeling it as one. Some popular tools right now include OpenAI’s mega-viral hits DALL-E 2 and ChatGPT, LensaAI’s image generator, and even Google’s AI assistant, Bard.
Generative AI models are built using a machine learning model called a generative model. When the model learns from a large data set of examples, say a gallery of faces, it then uses its new knowledge to generate new data points that become part of the training data set. Their original data set is at the core of generative models, allowing them to begin learning. This initial training data set has raised the alarm as generative AI tools become increasingly popular, and the question of ownership over the training data is becoming more and more complicated. The question of just how far AI will go to replace the people who produce the types of data sets the models train is also on many artists, producers, and actors’ minds.
But, another consideration is the question of who owns the data that generative AI models are trained on. When the data set is filled with content like images or songs, those pieces belong to the people who created them or those who own the copyright and trademark. Right? Well, the question of ownership is more complicated than you might think. To interrogate who owns the data for generative models, we will look at three examples of the question of ownership in the context of generative AI tools that create novel audio and art from text prompts.
One of the most compelling uses of generative AI is to use AI to replicate the voices of actors and actresses. For example, when James Earl Jones, the voice of Darth Vader in the Star Wars movies, decided to retire, he passed his role onto an AI-powered voice replica. Jones’s voice was synthesized by the Ukrainian company, Respeecher, which managed to replicate his voice for Disney’s “Obi-Wan Kenobi” series while Russia invaded the nation. The decision caused waves in the acting community as some wondered whether the ownership of their own voice belonged to them in perpetuity, or to the company that owns the character they played.
In addition to Respeecher making headlines in the wake of James Earl Jones’s retirement, Microsoft released VALL-E, an AI model that can generate speech from just three-second audio samples. The text-to-speech generator was trained on 60,000 hours of English language speech from Meta’s audio library, LibriLight. The generator mimics the desired speaker as it vocalizes the desired text output. One possible use for text-to-speech generators was pioneered by Apple when they released audiobooks narrated by an AI-generated voice. Apple described the books, released on its Books App, as narrated by a digital voice based on a human narrator’s voice. The narration is semi-convincing. While some reacted to the news with concern, worrying that AI could take over the jobs of narrators, some felt that publishers would always gravitate towards human voices that are not only variable but convey more emotion.
Google recently published its new music generator called MusicLM, which generates music from text prompts using AI. The generator creates music in any genre from a user prompt. MusicLM’s vast data set of 280,000 hours of music sets it apart from other music generators and allows it to create music with surprising depth. Some of the songs are even enjoyable. But, most pieces are still far from the complexity and variability possible with human-generated music. The compositions sound strange to the human ear, and the AI’s vocal stylings are sometimes impossible to understand. Google decided not to release the generator to the public over concerns that the songs generated by MusicLM copied songs in its training data. According to the whitepaper, about one percent of the music produced by MusicLM was copied directly from songs in its training set.
Google’s decision not to publish the tool eased the concerns of some industry professionals who are worried about the future of AI-generated music. The Recording Industry Association of America flagged AI music mixers as a threat to copyright holders and music makers. The RIAA listed several AI music mixers, writing that online services that use artificial intelligence to extract or copy any element of a song are guilty of pirating copyrighted materials. The key concern is that the training set used by the generator consists of copyrighted music, and the AI is trained on the data without obtaining permission from the owners to include their work. While the future of AI taking over the music world might be far off, concerns about the inclusion of copyrighted music in potential generators are pressing and present.
The image generator is perhaps the most advanced generative AI tool on the market. Image generators like OpenAI’s DALL-E 2 shot to fame as people tossed insane prompts at the generator and received high-quality images in return. Unfortunately, image generators quickly caught the attention of artists for the wrong reasons. The training set for image generators includes artwork by artists who did not allow their art to be used to train an AI algorithm that some worry will replace the need to hire artists. Adobe recently reassured artists who use its platforms that it will not use content generated on Adobe to train generative AI tools. Adobe CEO Scott Belsky said that Adobe views generative AI as a hyper-effective “creative” assistant for users of its platforms. So, if Adobe does want to incorporate an AI tool, it won’t be trained on pieces created by its users without their permission.
One of the most visible battles over image generators is the one between Getty Images and Stability AI, which released an image generator called Stable Diffusion that uses text prompts to generate images with AI. As excitement built around Stable Diffusion, so did concerns about its training data. Getty Images responded to the generator by suing Stability AI for infringing on its copyright and trademark protections by training its generator on a data set that included images from Getty. In addition, Getty alleges that Stability AI used 12 million images in its training set without permission from Getty and without compensating the image library.
Legal experts say that Getty has a better case against Stability AI than individual artists may have if they chose to sue Stability for including their work in their training set without permission or compensation. Getty’s case is particularly strong because the company has previously licensed its images to other AI art generators, highlighting the fact that Stability AI scraped the images with Getty’s database without permission. Stability AI will need to argue that scrapping the images constitutes a case of “fair use”, which protects the use of copyrighted materials without a license. Stability AI will also need to prove that the image generator “transforms” the images used in its dataset instead of memorizing and replicating them. Researchers have noted that Stable Diffusion has a habit of memorizing and replicating images from its training set, reaching a 0.03% memorization rate.
Who owns generative AI, the jury is still out? It is the people who create the images, music or voices the models are trained on? Is it the platform that the data is created on, like Adobe? Or is it the companies who create these large scale models capable of answering a text prompt with a piece of art? When products are prone to memorization or plagiarism, like music and image generators are, the question of ownership becomes more important. Is the future of the training sets needed to build generative tools one where each piece of data has been uploaded with permission from the owner? If it is, then the owners of generative AI are the people who build the pieces which become its data set. The case against Stability AI will have major implications in the industry if the scope of training sets are narrowed to only open source licenses. As generative AI usage grows, it is important to understand the ownership of the IP that feeds the tools we play with.
.svg)
.svg)
