Editing Openai/6924739c-2300-800e-9743-65c152c213f1 (section)

==== ### ====

What the Directive does
* Introduces two TDM exceptions: - Art. 3 – mandatory exception for research organisations/cultural heritage institutions doing TDM for scientific research. - Art. 4 – broader exception allowing anyone to perform TDM unless rightsholders opt out using machine-readable reservations.
* Scholarship and policy analyses confirm these provisions were intended to cover automated analysis and AI training on large corpora.

Relevance
* This is the leading legislative model for handling AI-style training by exception rather than by fair use.
* It would have been a natural comparative touchstone for: - Arguing that when legislatures want to bless AI training on protected content, they do so expressly via TDM exceptions; - Or conversely, that absent such an exception, broad unlicensed training remains presumptively infringing.

===== What happened =====
* Photographer sued LAION over use of images scraped from the web and incorporated into an AI training dataset.
* The Hamburg court held that reproductions for training could fall under Germany’s TDM exception for scientific research (Section 60d UrhG, implementing DSM Art. 3), and that AI training counts as TDM.

Relevance
* One of the first court decisions squarely addressing AI training and TDM in Europe.
* Shows a judicial approach that: - Accepts AI training as TDM; - Uses statute-based exceptions rather than judge-made fair use; - Still leaves room for rightholder opt-outs and limits (especially when outputs reproduce works).

A U.S. court could look to this as a contrast model: in the EU, statutory exceptions + opt-outs are doing work that in the U.S. is being asked of fair use.

===== Key features =====
* Japan’s 2018 amendment created Article 30-4, which permits the use of copyrighted works for “information analysis”, including AI development and training, without needing authorisation, subject to certain conditions (for example, not using the works to enjoy them as such, and guarding against certain market harms).

Relevance
* Another major economy that has explicitly legalised AI training on copyright works via a broad TDM-style exception.
* Very useful as comparative material to show that where a jurisdiction wants to ensure AI training is lawful, it does so affirmatively—again reinforcing the idea that in the U.S., courts are being asked to fill a policy gap via fair use.

===== What it says =====
* Section 29A allows making copies of works for text and data analysis for non-commercial research, provided the user has lawful access to the work.
* There has been active UK discussion about broadening this into an EU-style TDM exception that would more clearly cover AI training.

Relevance
* For Westlaw-style subscription access, this is particularly on point: - It suggests a model where lawful subscribers can data-mine content in limited research contexts without permission, but commercial exploitation (like building a competing product) remains contested or excluded.
* It’s a neat comparative frame for where ROSS-type uses might sit if this were litigated under UK law.