Introduction
The copyright issues surrounding music-generating AI have entered another, more concrete phase. The Atlantic investigated multiple music datasets shared within the AI developer community and reported that they contain tens of millions of songs. These datasets are said to include tracks by globally known artists such as Bad Bunny, Nirvana, Taylor Swift, Billie Eilish, and the Beatles.
Questions have long been raised about whether music-generating AI has been trained on copyright-protected songs. Major music companies have already filed lawsuits against Suno and Udio, alleging copyright infringement, and the issue of training data behind AI-generated music is now being contested in court.
What makes this report important is that it goes beyond the impressionistic claim that “AI may be imitating famous songs.” Rather, it shows that the scale and contents of datasets accessible to AI developers are beginning to become visible. Competition in music-generating AI is no longer only about the quality of the songs it produces. It now faces a more fundamental question: what materials did the AI use to grow?
The Core Issue Is Not “Similar Songs,” but “Training Sources”
Discussions about music-generating AI often focus on whether an output song resembles an existing song. Certainly, if a generated track closely resembles a particular artist’s voice or musical style, the issue is easy for users to understand. However, the essence of The Atlantic’s report lies not only in similarity of output.
The more fundamental question is what kinds of songs were used at the training stage, and under what kind of rights clearance. Even if the generated result does not reproduce a particular song as it is, if large quantities of recordings were copied and analyzed for training, the question remains whether that act itself can be permitted without authorization from the rights holders.
Music is not mere data. It is the accumulation of many creative acts and investments, including lyric writing, composition, arrangement, performance, singing, recording, mixing, and mastering. Even if AI developers explain that they used “publicly available materials,” that does not immediately mean those materials may be freely used to train commercial AI systems.
The Limits of the Idea That “A List of Links Is Safe”
One interesting point in The Atlantic’s report is that some of the datasets at issue are distributed not as audio files themselves, but as lists of links to songs on YouTube or Spotify. At first glance, this may appear to carry a lower legal risk because the files are not being directly distributed.
However, for AI developers to actually use those links for training, they must obtain the audio from the linked sources. If they use automated tools in that process and bypass platform mechanisms such as login systems, advertisements, play counts, revenue sharing, and subscription incentives, the issue extends beyond copyright alone. It would also bypass multiple layers of order, including platform terms of service, revenue distribution to rights holders, and the visibility of creators.
Here, we see a problem of “indirectness” that is characteristic of data use in the AI era. No one is directly distributing the audio files. The dataset consists only of links. The actual acquisition is carried out by a separate tool. The boundary between research use and commercial use is also unclear. As responsibility becomes increasingly dispersed in this way, it becomes harder for rights holders to track where, by whom, and how their music has been used.
Why Music-Generating AI Creates Greater Friction
The problem of training data for generative AI is occurring in many fields, including text, images, video, and code. Among these, music is an area where friction is particularly likely to intensify.
First, music is a field in which market substitution is intuitively easy to understand. If AI can generate background music, advertising music, demo tracks, and vocal songs in a matter of seconds, it will directly compete with the work of human composers, performers, singers, producers, and studio professionals. If AI grows by using existing musical culture as its material and, as a result, puts pressure on the human creative market, it is only natural that rights holders and artists would object.
Second, music is strongly connected to personal elements such as “voice” and “style.” A singing style, tone, sense of rhythm, or mixing texture associated with a particular artist is not merely an arrangement of sounds; it is a brand built over many years. If AI learns those characteristics and becomes capable of generating large quantities of songs with a similar atmosphere, the effects extend beyond copyright to publicity rights, reputation, and relationships with fans.
Third, the music industry already has complex licensing practices. Systems for rights clearance have been built up for each type of use, including sound recordings, musical works, publishing rights, performances, sampling, and distribution. The addition of a new type of use called “AI training” has created areas that cannot be fully organized under existing contractual frameworks.
The Risks Hidden in the Term “Open Dataset”
Open datasets and research datasets have played a major role in AI development. Using data that anyone can access improves the reproducibility of research and accelerates technological progress. That value itself should not be denied.
However, “being openly available” and “being freely usable for commercial purposes” are separate matters. The fact that something can be found on the internet, listened to for free, shared by researchers, or used in past academic papers does not mean that rights clearance has been completed. In particular, when audio made available on the assumption of personal or non-commercial use is used to train commercial AI models, that use may greatly exceed the originally contemplated scope.
This report brings to light the reality that materials with different rights statuses are mixed together within what the AI industry has broadly referred to as “datasets.” Before competing over the performance of AI models, companies are being asked whether they can explain the origins of the data that supports that performance.
Transparency Will Become a Competitive Advantage
Until now, many AI companies have treated the details of their training data as trade secrets. The argument that disclosing what data was used could put them at a competitive disadvantage is understandable. However, when there is a possibility that large quantities of copyright-protected content have been used, maintaining complete non-disclosure is becoming difficult both socially and legally.
What will be required of music-generating AI going forward is not merely the ability to generate high-quality songs. It will also need accountability in data procurement: what data was used, what scope of licenses has been obtained, whether rights holders have a mechanism to request deletion or exclusion, and what assurances can be provided for commercial use of generated works.
This is not merely a cost. In the long term, it will become a source of competitiveness. AI services with unclear rights clearance will be difficult to adopt for uses that are sensitive to rights risk, such as corporate advertising, film, games, broadcasting, and in-store background music. Conversely, AI systems that can clearly explain the licenses and permitted uses of their training data are likely to be chosen for business use, even if they are somewhat more expensive.
The Changes Brought by a Searchable Database
The searchable database published by The Atlantic has important implications for artists and rights holders. Until now, it was almost impossible to confirm from the outside whether one’s work had been used by AI. Even when there were suspicions, gathering evidence was difficult.
When the contents of datasets become visible in searchable form, the center of the debate changes. It moves from “it may have been used” to “it is included in this dataset.” Of course, inclusion in a dataset is not the same as proof that a particular AI company actually trained on that song. Even so, greater transparency gives rights holders a starting point for inquiries, negotiations, litigation, and license design.
This is also a change AI developers cannot ignore. The issue of training data, which could previously be left ambiguous, is entering an era in which it can be verified by third parties. Going forward, it will be essential to record rights status, terms of use, and responses to exclusion requests from the stage of collecting datasets.
Implications for the Japanese Music Industry
This issue is not limited to the United States. Japanese music, anime songs, game music, idol songs, Vocaloid culture, and doujin music are all accessible on platforms around the world. It would not be surprising if Japanese songs were included in overseas datasets.
For Japanese companies and creators, the important question is not a simple binary choice between rejecting AI and accepting it. What matters is to make concrete decisions about the conditions under which use will be permitted, where authorization should be required, how rights holders will participate, and how they will receive compensation.
AI has great potential for composition support, demo production, adaptation, remixing, and sound effect generation. However, to expand that potential in a healthy way, it is necessary to build a market based on rights clearance rather than moving forward while leaving the procurement of training data ambiguous. The music industry, too, is now at a stage where it should prepare practical options such as AI training licenses, artist-level opt-in systems, revenue sharing, and use restrictions, rather than relying solely on blanket refusal.
What Is Needed Before Discussing “The Future of Creativity”
Supporters of generative AI argue that AI makes new forms of creativity possible. There is a certain persuasiveness to that view. In fact, AI can lower the barriers to music production and become a tool that enables individuals to create high-quality audio in a short period of time.
However, if we are going to discuss the future of creativity, we cannot ignore where the materials supporting that future came from. If past artists’ recordings, performances, voices, arrangements, and sound design are absorbed in large quantities without explanation or compensation, it may look less like the democratization of creativity and more like the unilateral reuse of existing creative assets.
What AI music needs for its development is not a confrontation between technological innovation and rights protection. What is needed is transparency in training data, choice for rights holders, compensation according to use, and rules that allow users to use generated works with confidence.
Conclusion
The Atlantic’s investigation has shifted the issue of music-generating AI away from the question of “AI has become able to create human-like songs” and back to the question of “whose music did AI grow up listening to?” This is a problem of creative infrastructure that cannot be seen through performance evaluations of AI alone.
Going forward, music-generating AI will become even more advanced and will generate songs that sound more natural and are easier to use commercially. For that very reason, we need to ask what kinds of data that technology is based on.
There is no doubt that AI will change the future of music. However, whether that future leaves creators behind or becomes one in which creators and technology coexist depends on the rules and practices being built right now. This report can be seen as an event that made that turning point visible.