AI’s rapid development is already blurring lines between human and machine creations, copyright law hasn’t caught up

In the past few years, the perhaps most striking innovation developments regarding copyright and the creation of new content have taken place with generative AI (Samuelson, 2023). Recent advances in ML have opened up new possibilities to produce literary, musical, and artistic works, and AI has become a new tool to express ideas for hundreds of millions of users. The applications are many and vary across the creative domains. For example, WaveNet was initially developed to help a ML algorithm create seamless artificial voices and has learned to make music itself by analyzing voice material. Current AI systems can also quickly create customized royalty-free background music, jungles, and even entire albums. Also, some video game environments are no longer created by programmers but by the program itself based on predetermined rules and algorithms (e.g., in No Man’s Sky). In visual art, machines can emulate human thought and decide how to alter an image based on a given algorithm. With AI art generators, such as DALL-E, Stable Diffusion, and Midjourney, users can create artwork by entering a prompt, choosing different styles, and further modifying the generated outputs to express their ideas. As the technology is reaching a point where it is difficult to distinguish works created by humans from works generated with AI (Jakesch et al., 2023), the importance of human influence on creative works is put to the test. At present, AI-generated works are not protected by copy- right in most jurisdictions because of the missing human input.“Human authorship” is seen as the guiding principle in copyright law (Hristov, 2016; Ramalho, 2017). Consequently, these works directly enter the public domain. The ongoing development of AI and its ability to make critical, creative decisions either with or without human intervention gives us reason to rethink our understanding of originality and ownership and evolve copyright law. When no protection is granted, the output generated by AI might be considered fair use of the underlying processing and recombination of training data (Sag, 2024). Since the AI is doing much more than copying bits and pieces of existing works and is creating new works based on the techniques learned from studying the copyrighted prior work, much like an art student learns to paint in particular styles by studying existing artwork in that style. However, output substantially similar to original works might still be copyright infringing, giving rise to the technical issue of “memorization” (Sag, 2023a). The case of a song that uses AI-generated voices of the popular artists “Drake & The Weeknd” released in 2023 further highlights a dilemma of online copyright enforcement for platforms. Under pressure from the record label that released the artists’ original works, streaming plat- forms such as Spotify and YouTube removed the song, citing copyright infringement. Given the legal uncertainty, accepting such infringement claims is tricky for Google, which is YouTube’s par- ent organization and a leader in the development of AI. As discussed in Patel (2023), by agreeing with the record label and treating AI-generated music as copyright infringing, Google endangers its own development of generative AI products (which may have much broader applications) under the fair use argument. On the other hand, by not agreeing with the record label, they risk lawsuits and potentially lose necessary licenses for music on YouTube, one of the biggest content categories. Even in countries where AI-generated work is protected by copyright, for example, in the UK, it is not clear who the author of the output is—several eligible personas come into question, such as the architect, the programmer or the user of the model, who have each contributed necessary input for the final work to be created. Depending on whether copyright protection for AI-generated works is granted has further implications on the question of potential copyright infringement regarding the training data, making it more likely that copying original images for training data is legal because the resulting AI software is used to create new, non-infringing artwork