The BRICS New Delhi Declaration, adopted on 12 September 2026, contains an important intervention in the debate over how artificial intelligence (AI) models are trained. Paragraph 120 calls for respect for intellectual property rights in the digital environment, including for AI training purposes, and for fair remuneration of rights-holders. It separately recognises the risk that knowledge, heritage and cultural values insufficiently represented in datasets and AI models may be misappropriated or misrepresented.
Much of the debate over AI training and copyright has centred on whether developers may use copyrighted works, whether rights-holders can prevent such use and whether they should be compensated. The BRICS position opens a broader q uestion: how knowledge that does not map neatly onto individual ownership enters AI systems and is subsequently represented – extending the debate beyond who owns or can license training material.
How the BRICS Position Has Evolved
Paragraph 120 is best read alongside the BRICS Leaders’ Statement on the Global Governance of AI adopted in Rio de Janeiro in July 2025. The Rio statement called for protection of intellectual property rights, particularly copyright against unauthorised AI use, fair-remuneration mechanisms and transparency around model inputs and outputs. The New Delhi Declaration carries those concerns forward while adding the explicit risk of misappropriation and misrepresentation of inadequately represented knowledge, heritage and cultural values.
Together, the two texts bring four concerns into the same emerging framework: copyright protection, remuneration, transparency and cultural integrity. The Bhopal Declaration adopted by BRICS Culture Ministers in August reinforces this dual concern, addressing creators’ rights and remuneration for copyrighted works alongside the safeguarding of Traditional Knowledge Systems and Traditional Cultural Expressions.
Beyond Individual Rights-Holders
The distinctiveness of this position becomes clearer when set against the copyright approaches emerging elsewhere.
In the United States, the Northern District of California held in Bartz v. Anthropic that using books to train Anthropic’s large language models was highly transformative and constituted fair use. But the court separately refused to excuse Anthropic’s acquisition and retention of pirated copies merely because some were later used for training. In Kadrey v. Meta, another Northern District of California court ruled for the AI developer, but stressed that the plaintiffs had failed to produce meaningful evidence of market harm. Neither decision therefore establishes that AI training on copyrighted works is generally lawful.
The European Union has taken a different route. Article 4 of the Directive on Copyright in the Digital Single Market permits text and data mining of lawfully accessible works where rights have not been appropriately reserved. The AI Act additionally requires providers of general-purpose AI models to maintain a copyright-compliance policy and publish a sufficiently detailed summary of training content.
These approaches differ, but their copyright architecture remains centred on works and identifiable rights-holders. The BRICS formulation adds another dimension: cultural knowledge may be collectively held, transmitted across generations or embedded in languages and traditions that do not fit neatly within conventional individual authorship and ownership. It also looks at how models represent their subjects and source materials – a concern for Global Majority countries that may serve as important sources of linguistic, cultural and other training data while having limited influence over how AI systems are developed.
From Copyright to Provenance
This is where provenance becomes relevant: maintaining a record of where material came from, the source with which it is associated, and relevant rights, restrictions or conditions governing its use.
A useful institutional analogy is the 2024 WIPO Treaty on Intellectual Property, Genetic Resources and Associated Traditional Knowledge. Once in force, it will introduce a mandatory disclosure requirement in specified patent applications involving genetic resources or associated traditional knowledge. The analogy is not exact, but its significance lies in separating disclosure of provenance from the prior resolution of ownership or remuneration. The treaty will enter into force after the required 15 ratifications or accessions. Applied to AI, a provenance framework could require developers to disclose source categories, principal datasets or repositories, language coverage, known geographic or community associations, and applicable rights or use restrictions. For linguistic and cultural material, this could create a basis for identifying particular corpora, scrutinising their sourcing and contesting documented cases of misrepresentation.
India’s Opportunity
India’s large body of linguistic, artistic, archival and traditional knowledge gives it a direct interest in this question. This concern with cultural representation is consistent with India’s emerging “third way” in AI governance, organised around inclusion, access and the priorities of developing economies. As BRICS chair, it also has an opportunity to take the issue from declaration-level commitments towards a practical framework
India is already addressing a separate part of the AI-training problem through copyright policy. In December 2025, a DPIIT-appointed committee proposed a hybrid model under which AI developers would receive a mandatory blanket licence to train on lawfully accessed copyrighted content without individual negotiations, with royalties payable upon commercialisation through a centralised collection and distribution mechanism. The proposal was released for consultation and has not become operative law.
India’s 2025 AI Governance Guidelines also recognise the need to develop standards relating to data integrity, content authentication and provenance.
At the same time, Indian courts are considering how existing copyright law applies to AI training. In ANI Media v. OpenAI, the Delhi High Court held, prima facie and for the purposes of interim relief, that OpenAI’s storage of ANI’s works for training fell within the fair-dealing provision in Section 52(1)(a) of the Copyright Act. The court stressed that the finding would not determine the final suit. The ruling is under appeal, and a Division Bench has declined to stay it at this stage.
But licensing and provenance answer different questions. A licensing framework asks whether copyrighted material may be used, on what terms and with what remuneration. Provenance asks what material entered an AI system, where it came from and what rights, restrictions or cultural associations attach to it. That distinction matters particularly for traditional and cultural knowledge that may be collectively held or sit uneasily within conventional copyright ownership.
From Signal to Framework
At the BRICS level, members could develop a common taxonomy and reporting template for training-data provenance. Developers could disclose principal datasets and repositories, source categories, language coverage, known geographic or community associations, and applicable rights reservations or restrictions.
Such disclosure would not establish ownership, determine compensation or reveal how every item of training data influences model outputs. But it would provide a traceable account of the material from which AI systems are built, giving states, rights-holders and communities a stronger basis for scrutinising sourcing and contesting documented misrepresentation.
A provenance framework would therefore complement rather than prejudge India’s eventual copyright regime. More broadly, it would allow BRICS members to act on their shared concern without requiring a common copyright system.
The significance of the BRICS intervention lies precisely here. AI training raises two distinct questions: who has rights over the material used to build AI systems, and how should the origins and cultural context of that material be recognised where conventional copyright provides no complete answer? India has an opportunity to turn that distinction into a practical governance framework.