In this fourth update on IP developments concerning generative AI, we highlight three significant developments: the UK government’s new consultation on AI and copyright, California’s groundbreaking transparency requirements for AI developers, and the evolving industry approach to AI training data. We invite you to consult with us to revisit your use of data sets as well as refine your terms of use or other legal terms to comply with these various legal regimes.

UK

The UK government has launched a consultation on AI and copyright (closing February 25, 2025) (the “Consultation”) proposing three significant changes to the current framework. First, it suggests expanding the Text and Data Mining (TDM) exception to include commercial purposes, applying to all works to which users have lawful access, including freely available internet content. Under this “opt-out” approach, rights holders would need to actively declare they do not want their work to be used in AI training through machine-readable formats, which may be standardized.

Second, as an integral part of the TDM exception, the Consultation proposes mandatory transparency requirements for AI developers such as disclosure of training data sources and web crawlers. The proposal suggests following a balanced approach, acknowledging both the practical challenges of detailed disclosure for AI developers (particularly small businesses) and the need to protect trade secrets while ensuring meaningful transparency.

Third, it proposes to eliminate Section 9(3) of the Copyright, Designs and Patents Act 1988, which currently provides copyright protection for computer-generated works without human authors[1]. The UK government notes that most leading AI nations, including the US and most EU Member States, do not provide copyright protection to works created without a human author, and there is little evidence that such protection increases AI output or investment in AI technology. Meanwhile, AI-assisted works that exhibit human creativity would continue to be protected under existing copyright law.

While this Consultation aims to align with EU regulations while addressing implementation challenges, it has already drawn significant criticism from various stakeholders. The proposed framework, particularly its opt-out mechanism, raises important questions about the balance between facilitating AI innovation and protecting creators’ rights. The  outcome of the Consultation may serve as a significant reference point for jurisdictions worldwide grappling with these complex issues.

California

While the UK is contemplating new transparency requirements, California has already enacted legislation in this area. The AB-2013 Generative AI legislation, signed into law in October 2024 (the “Law”), introduces extensive transparency requirements for AI developers. Starting January 1, 2026, any developer making generative AI technology available to California residents must publish detailed documentation about their training datasets, including sources and ownership, number of data points, whether they contain copyrighted material or personal information, the period of data collection, and whether synthetic data generation was used.

Notably, the Law defines “developer” broadly to include both original developers and those making “substantial modifications” to generative AI systems. Such modifications encompass new versions, retraining, and fine-tuning that “materially change functionality or performance”, significantly expanding the law’s reach.

Given the Law’s expansive definition of ‘developer’ and its extraterritorial reach, companies developing or modifying generative AI systems should begin preparing their documentation processes well ahead of the January 2026 deadline, particularly if their services might be accessed by California residents.

Industry Approach to AI Training Data

As noted in our previous client updates, Open AI and other generative AI companies have been facing lawsuits from various news outlets and artists (see e.g., Andersen v. Stability AI Ltd.).

In parallel to such legal challenges, and amid increasing regulatory scrutiny as shown in this update, some of the leading AI companies are proactively securing commercial agreements for training data access. OpenAI, for instance, has recently announced several significant partnerships with major media organizations including Future, Axel Springer, News Corp, and the Financial Times.

Earlier this month, Anthropic reached a settlement in the copyright infringement lawsuit filed by major music publishers over Claude’s alleged use of protected song lyrics in its training data. The settlement requires Anthropic to implement copyright protection guardrails and establish procedures for addressing potential infringement.

Looking ahead to 2025, we anticipate an acceleration in licensing agreements between AI companies and content owners, as they seek to mitigate legal risks and ensure compliance with emerging transparency requirements. However, with many existing AI models potentially containing unlicensed training data, organizations should review their generative AI providers’ compliance with data transparency requirements and verify the legitimacy of training data sources, as both providers and enterprise customers could face legal exposure from unauthorized content use.

The Herzog team invites you to consult with us to revisit your use of data sets as well as refine your terms of use or other legal terms in order to comply with the various legal regimes.

[1] That is unless the consultation reveals evidence of positive effects from computer-generated-works protection.