In 2018 Japan amended its Copyright Act (Act No. 30 of 2018) to add Article 30-4, a flexible exception widely regarded as one of the most permissive in the world for AI training. The provision allows the exploitation of a copyrighted work, to the extent necessary, when the purpose is not to “enjoy” the work - that is, not to perceive the thoughts or feelings expressed in it as a human reader, viewer, or listener would - but rather for purposes such as technology development, data analysis, and computer information processing. Machine learning on copyrighted data falls squarely within this “non-enjoyment” category, because a model extracts statistical patterns rather than appreciating the expressive content.
The Japan Agency for Cultural Affairs later confirmed, in its general understanding on AI and copyright, that Article 30-4 applies in principle to training generative AI models, whether for commercial or non-commercial use, without requiring the rightsholder’s permission. The exception is not unlimited: a proviso withholds the benefit where the use would “unreasonably prejudice the interests of the copyright holder,” and using AI to reproduce expressive content for human enjoyment is treated as a separate, regulated act. But the baseline rule is strikingly favorable to training, and Japan has explicitly positioned this stance as a way to attract AI development.
Why business readers should care: jurisdiction matters enormously for AI training. Japan’s “non-enjoyment” exception offers far more headroom than the EU’s opt-out or the contested US fair-use defense, which is why where a model is trained - not just by whom - is now a strategic legal question.