Two Boundaries Beyond the Model
Copyright questions do not end at the U.S. fair-use analysis. They also depend on the jurisdiction governing the training use and on whether the final output contains human-authored expression. This chapter compares a few international approaches by their recurring design choices—lawful access, analytical purpose, expressive enjoyment, opt-out signals, and compensation—then separates permission to use a source work from protection for a human contribution to an AI-assisted output.
The distinction prevents a particularly costly category error. A foreign text-and-data-mining exception may permit a defined kind of copying under defined conditions; it does not necessarily make every output lawful. Likewise, a person may own copyright in an illustration, arrangement, or modification that appears in an AI-assisted work without having permission to train a model on someone else’s material. The two questions meet in one system, but they ask for different facts.
Training exceptions are designed around purpose and access
There is no single international rule for generative AI training. The Copyright Office’s training report describes a patchwork of exceptions, guidance, consultations, and developing case law. The most useful comparison is not a country-by-country inventory. It is to ask what each system does with five variables: whether the user had lawful access, whether the use is analytical or expressive, whether an owner can reserve the work, whether copies can be shared, and whether compensation is required.
The European Union provides a prominent example. Its 2019 Digital Single Market Directive directs member states to provide exceptions for certain text and data mining, or TDM, activities. One provision is limited to research organizations and cultural-heritage institutions conducting scientific research on works to which they have lawful access. A broader provision applies to any actor, but still conditions the exception on lawful access and on respecting a rightsholder’s reservation of rights, including a machine-readable reservation for online material.
The EU’s Artificial Intelligence Act connects those rules to general-purpose AI model providers. As described by the Office, providers must establish policies for complying with Union copyright law and identify and comply with applicable opt-outs under the broader TDM exception. The framework therefore does not simply say “training is allowed.” It combines an exception with access conditions, permissions signaling, and provider duties. The report also notes controversy over how these rules apply to generative AI and uncertainty about implementation across member states.
Singapore’s approach, as summarized in the report, also conditions its TDM exception on lawful access and limits the use to computational data analysis. Copies may be supplied to others only for purposes such as verifying results or conducting collaborative research. Those conditions draw a line between analyzing a work and distributing it as content. The legal permission is consequently connected to what the user does with the copy after the analysis, not only to how the copy was made.
Japan’s exception permits certain uses for AI development and data analysis, but it excludes uses aimed at personally enjoying the thoughts or sentiments expressed in the work. Japanese guidance explains “enjoyment” through examples such as reading literature, appreciating music, or executing computer programs. Generating material similar to the original may count as enjoyment, and a use that is partly for enjoyment may fall outside the exception. The law also contains a safeguard against unreasonably prejudicing the copyright owner’s interests in light of the work, its purpose, or the circumstances of exploitation.