posted by: Nova K. Retro-Future Columnist

In the United States, the debate about generative AI is no longer merely a matter of technical achievement. It has shifted toward a quieter, almost ritualistic question: what exactly does a model do with the works it absorbs? Between training on books, images, or recordings and producing text or responses that resemble existing works without reproducing them verbatim, the legal boundary is tightening.[1][7][8] This is not a compliance detail; it has become a central point where AI’s economic form is being determined.

The final report of the U.S. Copyright Office dedicated to training generative models, published in May 2025, has given this debate clearer structure.[1][12] It reminds us that fair use is not a blanket permission, and that the analysis depends on context: purpose of use, nature of the work, amount copied, and potential market impact.[1][4][11][12] Public summaries of this report emphasize a particularly sensitive point: when the final use competes with a work’s original function, the transformative argument becomes more fragile.[1][3][12] In other words, transformation is not just a change of form; the economic role of the object must also shift.

This nuance is crucial because many AI advocates have long presented training as the digital equivalent of human reading.[3][8] However, the compiled sources here show that this comparison has its limits.[3][12] Under U.S. law, being transformative alone is not enough if the market for the original works is threatened.[4][11][12] The report and several legal analyses published on this topic converge on a simple, almost austere idea: the “it’s like learning” argument no longer closes the case. It instead opens questions regarding substitute uses and licensing.

The most notable decision in this sequence, in June 2025, comes from the case Thomson Reuters v. Ross Intelligence.[2][5][10] A federal court ruled that using protected content to train an AI system aimed at producing competing results could constitute infringement rather than fair use.[2][5][10] Available analyses highlight that the case involved a legal research tool rather than a generative model in the strict sense, but its implications go beyond that scope.[5][10] The message is clear: when a system learns from protected works to better serve the same market, relying on fair use protection becomes harder to justify.

The Anthropic case, meanwhile, reminds us that jurisprudence does not advance in a straight line.[2][6][9] Another federal ruling issued in June 2025 found that training a model on books could fall under fair use under certain conditions, while distinguishing that from how copies were obtained and from separate issues related to pirated copies.[2][6][9] This coexistence of decisions is far from anecdotal. It outlines a landscape where AI is neither banned nor absolved, but assessed according to the source of the data, the nature of the final product, and proximity to an existing market.[2][6][9][12] Law does not yet decide the entire future; it carves out zones of risk.

This is precisely where the issue moves beyond pure technicalities and enters the political economy of models. If training on protected works requires more licenses, AI companies will have to integrate this cost into their margins, timelines, and product choices.[1][4][12] This advantages players capable of negotiating at scale, documenting their corpora, and presenting robust compliance chains.[1][4][12] For creators, the stakes are less abstract than they seem: it is about whether works become free raw materials of the model era or regain measurable contractual value.[1][3][12]

Yet a significant shadow remains, and it must be kept open. Available documents do not yet allow a general conclusion that all AI training on protected content is legal or illegal.[1][4][12] The cited cases distinguish books, legal databases, copies obtained legally or illegally, as well as internal uses and competing commercial uses.[2][5][6][9] What matters next are upcoming rulings refining the notion of substitute markets, the role of collective licenses, and the fate of models producing outputs too close to original works.

In the background, this evolution also touches on the culture of the web and digital tools. A search engine, a documentary database, or a generative assistant do not operate with the same social expectations.[5][8][11] The first indexes; the second serves; the third synthesizes and sometimes replaces.[5][8][11] The more the interface becomes silent, the louder the question of the content it has absorbed becomes in law. This feeling is familiar in Tokyo as elsewhere: as the screen grows smoother, the invisible chain of what feeds it suddenly becomes more important for

Ultimately, the real question is not whether AI “cites” works as a human would. It is to determine whether it uses them to produce something else or to occupy the space those works already held. As long as this distinction remains unclear, fair use will remain shifting ground, not an automatic refuge.[1][4][11][12] The next turning point will likely come from how courts handle the combination of licensed data and competing markets; this will shape the lasting legal memory of generative AI.[2][5][6][9] And it’s this thread, more than passing trends, that will require close attention going forward.