Generative models and copyright, current state
By AI Resource Zone Admin · March 15, 2026 · 4 min read
Courts on both sides of the Atlantic are still working out how copyright applies to training data and generated output.
Copyright questions around generative models sit at the intersection of two long-running debates, what counts as a copy during training, and who, if anyone, owns the output. In the United States, a cluster of lawsuits filed from 2023 onward by authors, news publishers, and visual artists is working through discovery and early motions. The Copyright Office has issued guidance clarifying that purely machine-generated output without human creative input is not eligible for registration, while works with meaningful human authorship can be.
In Europe, the AI Act layers transparency obligations on providers of general-purpose models, requiring them to publish sufficiently detailed summaries of training content and to respect opt-outs expressed under the existing text-and-data-mining rules in the Copyright Directive. The United Kingdom has been consulting on a text-and-data-mining regime of its own, with earlier proposals for a broad exception withdrawn after pushback from the creative industries. Japan continues to operate under a permissive statutory framework that allows training on lawfully accessed works for non-enjoyment purposes.
Licensing markets are beginning to form around these rules. Several large news publishers, stock image providers, and book distributors have signed training-data agreements with model developers. Some creators have chosen to opt out of crawling using machine-readable signals, including updated robots.txt directives and industry-specific preferences. Standards bodies and the World Wide Web Consortium are discussing how to make these signals more consistent across platforms.
Editor's note: Consumers rarely think about copyright, but the resolution of these cases will affect what AI products look like. Training-data licensing will shape which models exist, which languages they speak well, and how much they cost to run. It will also influence whether smaller developers can compete with the handful of firms that can afford comprehensive licenses. Watching settlements and early judgments matters more than headline verdicts at this stage.