Three and a half years after the emergence of generative AI as a new technology paradigm, there is broad agreement that AI companies have extracted enormous value from the digital commons without meaningful reciprocity. The harder question — one that most policy discussions still avoid — is what a workable redistribution mechanism that accounts for this value extraction would actually look like. It is becoming increasingly clear that copyright licensing, however creatively applied, cannot answer that question alone. The public information commons, whose very existence has been one of the enabling conditions of the current generation of AI systems, extends far beyond works that are actively managed by rightholders. It includes the public domain, openly licensed content, publicly funded research, government data, and the accumulated output of decades of everyday cultural expression. A compensation mechanism that operates only within the copyright economy will systematically exclude the majority of those whose contributions made these systems possible.

A compensation mechanism that operates only within the copyright economy will systematically exclude the majority of those whose contributions made these systems possible.

This is why the debate needs to shift from training-time licensing to deployment-based redistribution. Rather than trying to attach consent or compensation obligations to the moment training data is ingested — a moment at which commercial value is uncertain, transaction costs are prohibitive, and the link between specific works and the value created by downstream models and applications is technically difficult to establish — the trigger should be the commercial deployment of AI services on the market. At that point, added value has been created and can be measured, and the redistribution obligation falls on the entities best positioned to bear it. And while this will require additional modelling and experimentation, it should be possible to define the scope of the mechanism to match the scope of the underlying appropriation — or the desired contribution to society — rather than being confined to what copyright law happens to cover and channelling the resulting revenue back to a narrow class of professional creators and commercial rightholders.

This is why the debate needs to shift from training-time licensing to deployment-based redistribution.

I outlined such a framework — a levy on commercial AI services, with proceeds distributed back to information producers, custodians, and public institutions — in last year’s Beyond AI and Copyright white paper. What I want to do here is set out three conditions that any deployment-based redistribution mechanism must meet if it is to function equitably and effectively. These are not technical parameters but political tests. A mechanism that fails any of them will either entrench existing inequalities or fail to address the structural problem it is intended to solve.

Revenues must flow beyond copyright

The most important condition is also the most likely to be violated in practice. Any deployment levy that channels its proceeds exclusively through copyright-based distribution mechanisms will replicate the existing distribution of power in the information economy — and thereby exclude exactly the actors who have the weakest market position but whose contributions make up the most substantial part of the commons.

A levy whose distribution logic cannot reach these contributors is not a solution to the problem of AI extractivism — it is a redistribution of rents within the commercial information economy, leaving the commons itself uncompensated.

Existing collective management infrastructure is designed to distribute remuneration to professional rights-holders. That is a legitimate function, but an insufficient one. The training of foundation models has not drawn exclusively on the professional content economy. It has drawn on the whole open web: on public domain works, open access research, the digitized collections of libraries and archives, and content produced and shared by individuals and communities with no expectation of commercial return. A levy whose distribution logic cannot reach these contributors is not a solution to the problem of AI extractivism — it is a redistribution of rents within the commercial information economy, leaving the commons itself uncompensated. The appropriate response to extraction at this scale is not individual remuneration — most of the information that powers AI systems cannot be reliably traced to specific creators, and the transaction costs of doing so would be prohibitive — but collective redistribution: funding the institutions and infrastructure that sustain public information production for everyone, including professional creators. This requires making explicit political choices about who counts as a contributor to the information ecosystem: open access publishers, cultural heritage institutions, public service media, Wikipedia and similar openly governed knowledge projects, and — as discussed below — public AI infrastructure.

The working paper on generative AI and copyright, released by India’s Department for Promotion of Industry and Internal Trade in December 2025, illustrates both the appeal and the limits of this approach. The paper proposes a mandatory revenue-linked remuneration right for AI training — sharing the core intuition that the attachment point for compensation should be downstream rather than at the moment of training. But it remains firmly anchored in copyright: the proposed mechanism distributes to registered rights-holders through collective management organizations, with no recognition of the broader information commons.

The current copyright debate is dominated by organized rights-holders from the Global North whose organizational capacity and legal standing make them visible in policy processes that most information producers in the Global South are structurally excluded from.

The global dimension of the challenge to value redistribution makes this condition even more urgent. The current copyright debate is dominated by organized rights-holders from the Global North whose organizational capacity and legal standing make them visible in policy processes that most information producers in the Global South are structurally excluded from. A deployment levy that defaults to existing rights management structures will predictably reproduce this asymmetry. Any serious attempt to design such a mechanism must grapple with questions of global redistribution — including what share of revenues should flow toward multilateral cultural infrastructure rather than being captured entirely within the jurisdiction implementing the levy.

A levy must be part of a larger public AI strategy

A redistribution mechanism that generates revenues without simultaneously reshaping the AI infrastructure landscape solves only half the problem. The deeper issue is not just that commercial AI companies extract value from the public information commons — it is that, absent any intervention, the capacity to deploy AI at scale will remain concentrated in a small number of entities whose interests are not aligned with those of the communities and institutions that sustain public knowledge. Redistribution that funds existing information producers while leaving this concentration intact may be able to slow the deterioration of the information ecosystem, but it does not address the emergence of a new layer of powerful commercial gatekeepers over how societies access knowledge.

The historical analogy that is most instructive here is not a comforting one. The failure to build public alternatives to commercial platform infrastructure in the early internet era was not a matter of insufficient political will at a single moment. It was the cumulative result of decades of market-centric policymaking that foreclosed public options before they could develop. AI development now presents us with a rapidly closing window in which such a choice is still possible.

Once large-scale AI systems are fully embedded in how people access information, building meaningful public alternatives becomes structurally foreclosed.

A portion of levy revenues must therefore be directed toward public AI infrastructure: models, datasets, and services that are publicly funded, transparently governed, and designed to serve the public interest. As our white paper on Public AI makes clear, building viable public alternatives to commercial AI systems requires substantial investment across the full AI stack — compute, data, and models — and levy revenues alone will not be sufficient to fund this. The levy is best understood as a dedicated, structural contribution to public AI infrastructure from those who extract value from the commons, complementing rather than substituting for direct public investment.

What “public AI infrastructure” means institutionally remains genuinely open. The right model is probably somewhere between a Wikimedia-style transnational commons and public broadcasting systems — but whatever form it takes, the fundamental challenge is scale. Ideally, this would be a multilateral effort to counter the economies of scale that commercial AI providers have achieved. Realistically — given the absence of meaningful coordination in a world dominated by shifting geopolitical alignments — regional collaborations such as the European Union seem the most likely level of intervention for building out such systems.

Protecting open-source model development, research, and innovation

Any redistribution mechanism that applies indiscriminately to all uses of AI carries the risk of creating barriers to exactly the activities most worth protecting: academic research, civil society AI development, and the development of open or open-weight models that empower broad societal adoption of AI without reinforcing dependencies on proprietary commercial providers. Unrestricted academic research and the development of open-source AI models and systems are also core pillars underpinning any credible public AI strategy. This means that these activities will need to be exempted from redistribution obligations. Moving the attachment point for the remuneration obligation from the use of copyrighted works for model training to the commercial deployment of AI systems trained on publicly available information effectively creates a safe harbour for research and open source model development, providing much-needed legal certainty for researchers and public AI developers.

Unrestricted academic research and the development of open-source AI models and systems are also core pillars underpinning any credible public AI strategy.

The CREATe Centre’s proposal in response to the UK government’s consultation on copyright and AI points toward one way of achieving this. It argued for broad permissibility of AI training in the pre-commercial phase combined with clear obligations at the point of market deployment. This is precisely the logic of a deployment-based framework: the key distinction is not between types of models or actors, but between stages of the value chain. Research, experimentation, and open model development happen before commercial value is extracted — and it is the extraction of commercial value that triggers the obligation.

Looking forward

The DPIIT working paper discussed in the first section shows that the intuition behind a deployment-based approach is gaining traction beyond the jurisdictions that have dominated the AI policy debate so far. That is an encouraging sign. The question is how far to go: rather than returning to in-copyright remuneration mechanisms that have proven politically intractable everywhere, any jurisdiction that is serious about addressing value extraction while sustaining the information commons should consider going further — toward a framework that meets the three conditions outlined above.

The question is how far to go: rather than returning to in-copyright remuneration mechanisms that have proven politically intractable everywhere, any jurisdiction that is serious about addressing value extraction while sustaining the information commons should consider going further — toward a framework that meets the three conditions outlined above.

The urgency of this is hard to overstate. Focusing on consent and compensation for training — looking backwards at value extraction that has already happened — misses the point that the value is being created here and now, when AI systems are deployed at scale and the information commons is being turned into a commercial product in real time. That is where the intervention needs to happen, and the window for shaping the terms on which it happens is closing.

Two caveats of the framework proposed here are worth acknowledging. First, it is not intended as a proposal to strip creators and other information producers of agency. Rather, it shifts the focus toward the stage of the process where the redistribution of value can be most effectively organized. For real-time uses of copyrighted works — such as retrieval-augmented generation — rights-holders should still be able to set conditions based on their ability to control access, and licensing arrangements at that stage remain appropriate.

Second, this framework has been developed primarily with the power dynamics and value chains of information production in mind. It is becoming increasingly clear that there are creative sectors — such as music and audiovisual — where copyright-based approaches, whether through voluntary or collective licensing, may be better suited to ensure a reciprocal distribution of value between AI companies and creators.

In summary, the framework proposed here is not intended to displace copyright-based mechanisms where they work, but to complement them where they do not.

The series brings together expert voices and was commissioned to inform the development of the issue brief by IT for Change, ‘Governing AI for the Cultural Commons: Beyond Intellectual Property’, under the AI, Culture and Intellectual Property Subgroup of the UNESCO Global Civil Society Organizations (CSO) and Academic Network on AI Ethics and Policy.

New article every Wednesday! Watch this space for more thinkpieces and read the issue brief here.