Source text: legal position, compensation models and primary sources

Proposed roadmap

The timeline is based on voluntary and technical solutions being tested before broad compulsory legislation. At the same time, it avoids allowing voluntarism to become a permanent excuse for postponing regulation: if the pilots’ coverage, transparency or distribution is inadequate, an extended collective licence or targeted statutory licence should be capable of being activated following an open evaluation.

Primary sources

• The EU DSM Directive 2019/790, particularly Articles 3, 4 and 12 – the basic rules for research TDM, general TDM subject to reservations, and collective licences with extended effect.
• The EU AI Act 2024/1689, particularly Article 53 – copyright policy and training summaries for providers of general-purpose AI models.
• The European Commission’s GPAI Code and implementation information – practical compliance with transparency and copyright requirements.
• The Swedish Copyright Act, Act 1960:729 – exclusive rights, TDM, contractual licences, liability and compensation.
• Government Bill 2021/22:278, Copyright in the Digital Single Market – the Swedish preparatory legislative materials for implementing the DSM Directive.
• United States Copyright Act, §§ 102, 107, 504 and 506 – originality, the limits of protection, fair use and liability.
• U.S. Copyright Office, Copyright and Artificial Intelligence, parts two and three – human authorship, AI-generated outputs, training, licences and assessments of potential reforms.
• Japan’s Copyright Act and the Agency for Cultural Affairs’ General Understanding on AI and Copyright – Article 30-4 and non-consumptive use.
• China’s Copyright Law 2020 and the interim regulations for generative AI services – rights, exceptions, sanctions and requirements for lawful training sources.
• India’s Copyright Act 1957 and § 52 – computer-generated works, fair dealing, copyright societies and liability.
• South African Constitutional Court, CCT306/24 (2026) – constitutional review of proposed fair-use and digital exceptions.
• The Berne Convention and the TRIPS Agreement – international minimum rights and the limitations of the three-step test.

• DSM Directive 2019/790 (EUR-Lex)
• AI Act 2024/1689 (EUR-Lex)
• Act (1960:729) on Copyright (Riksdag)
• U.S. Copyright Office: Copyright and Artificial Intelligence
• EPRS: Generative AI and Copyright (European Parliament)

Further reading and policy material

• The European Parliament study Generative AI and Copyright – Training, Creation, Regulation (2025): the gap between generative AI and the EU’s current TDM rules, and proposals on transparency and licensing.
• The European Parliament resolution on copyright and generative AI, adopted in March 2026: political demands for respect for opt-out, stronger negotiating power and collective licensing.
• European Commission reports on collective rights management and extended licensing: institutional design, representativeness and experience of ECL models.
• U.S. Copyright Office, part three on generative AI training: the most detailed official comparison of voluntary licensing, compulsory licensing, opt-out and fair use.
• WIPO’s work on AI and copyright infrastructure: attribution, compensation, registers and cross-border rights management.
• W3C TDM Reservation Protocol: a technical protocol for machine-readable TDM reservations and links to licensing policies.
• C2PA Content Credentials: cryptographically bound provenance and documentation of the origin and alterations of digital content.
• IPTC guidelines for AI opt-out, version 2.0: practical methods for publishers and image-rights holders to express reservations concerning AI and data mining.
• Academic comparison of control and compensation in AI training: trade-offs between exclusive rights, collective licensing, rights to compensation and market models.

The information in the underlying material has been assessed as at 10 August 2026. The legal position is changing rapidly; new rulings and EU decisions may alter the assessments, and the text is updated with dates and version history.