Source: Diplomacy & Law

AI training raises separate questions about domestic copyright liability and states’ compliance with international obligations, according to Edmarverson A. Santos’s legal analysis. Santos explains that the Berne Convention and TRIPS neither categorically prohibit training on copyrighted works nor expressly provide general permission for it. The analysis places reproduction rights and the limits on national copyright exceptions at the centre of the issue. Training collections can combine protected works, licensed material, public-domain content and unprotected information. The inclusion of copyrighted material alone does not establish infringement; the acts performed and the conditions of acquisition and processing also matter. Extracting an unprotected fact and copying the protected expression surrounding that fact require different legal assessments.

Santos explains that collecting, downloading, storing and preparing training material can involve reproductions requiring authorization or an applicable exception. A computational purpose does not, by itself, remove those copyright questions. The analysis cautions that numerical representations do not automatically establish the reproduction of protected works. Where a model retains recoverable expression, the assessment depends on technical evidence and applicable copyright rules. Copies used to assemble a corpus, material retained in a model and generated outputs may therefore require separate examination. Santos describes differing national approaches and emphasizes that permission under domestic law and compliance of that law with treaty obligations remain distinct issues.