After the AI Impact Summit in India, one message emerged with resounding clarity – hyperscaled AI is the way to go. In addition to India signing on to the US’s Pax Silica Declaration and India securing investments of 250 billion USD for investment in AI infrastructure from both governments and commercial actors, India and the US also issued a joint statement pledging to adopt “regulatory regimes that advance technological innovation and promote investment”.

We know now that almost all generative AI models are trained on data from the open web – notably, peer-produced digital and cultural commons (such as Wikipedia, free and openly licensed software, datasets, creative works) and other publicly accessible content made available through open Internet protocols. All this “stock” of data is harvested at scale via web crawling and without regard for copyright, which significantly subsidises the cost of AI development. This reuse also concentrates market power as well as agenda-setting power in the hands of a few firms.

That digital and cultural commons are crucial to AI innovation is a testament to the centrality of these commons for human flourishing.

That digital and cultural commons are crucial to AI innovation is a testament to the centrality of these commons for human flourishing. But these commons are being extracted by a small set of hyperscalers, who give little to no value back either to the communities that are the lifeblood of these commons or to society at large (in fact, they actively harm humanity, as we see in the use of generative AI for warfare). Against this context, what legal reforms are necessary, both within and outside copyright law? Drawing from previous research, I offer some reflections on emerging public-interest licensing initiatives for AI training datasets.

Copyleft licenses

A core legal tool for the digital commons is licenses that invert copyright. Copyright holders use standardized licenses to grant wide pre-facto authorizations for downstream distributions, modifications, and adaptations of their work. Free Software licenses encode the principle of copyleft – requiring adaptations and modifications of the licensed software code to be shared on similar terms as the original, to ensure a constant supply back to the “stock” of digital commons on the Internet. The Free Culture movement extended this principle to creative works. The Creative Commons license suite creates standardized license options, including widest reuse either through a public domain declaration or reuse with attribution, to more conditional reuse through similar copyleft conditions or limiting reuse only for non-commercial purposes. Across all these licenses, however, there is no discrimination based on the type of user. In fact, the open source movement went one step further to create very permissive licenses that do not contain any copyleft, thereby leaving the door open for open source artefacts to be integrated into proprietary technologies.

Free Software licenses encode the principle of copyleft – requiring adaptations and modifications of the licensed software code to be shared on similar terms as the original, to ensure a constant supply back to the “stock” of digital commons on the Internet.

But as Dmriti Kleiner noted almost two decades ago, “[t]o be ‘free’ means to be open to commercial appropriation, since freedom, in the terms of copyleft, is defined as the non-restrictive circulation of information rather than as freedom from exploitation”. The mere existence of a “stock” of digital and cultural commons, whose easy access and reproduction is facilitated by the Internet, does not ensure equitable value realization; this type of freedom and openness has in fact facilitated new concentrations of wealth and power.

Anti-extractivist licenses that expand on the logic of copyleft

Kleiner proposed the concept of  “copyfarleft”, which has informed other efforts such as the Peer Production License and the copyleft license. These licenses follow a dual-track approach discriminating between two types of licensees – they allow downstream commercialization only by actors in the social and solidarity economy, while all other actors are prohibited from commercial use and need to follow copyleft.

There is now a new wave of anti-extractivist licenses emerging in the AI context.

There is now a new wave of anti-extractivist licenses emerging in the AI context. Licenses created by African communities for their language datasets, a crucial training resource for LLMs, also contain a dual-track approach. The Nwulite Obodo license for African language datasets is completely permissive for licensees from developing countries, while licensees from the Global North can be required to either pay royalties, share infrastructure, commit to stronger copyleft, or offer any other type of value back to the licensors. The Esethu Framework is also permissive for researcher reuse and commercial reuse by African entities, while entities outside Africa need to pay a licensing fee, which is reinvested back into the creation of more data.

These are not simply means to demand compensation from Big Tech for the continued exploitation of local cultural resources.

These are not simply means to demand compensation from Big Tech for the continued exploitation of local cultural resources. Rather, they illustrate new strategic subversions of copyright to build AI ecosystems based on equity and care, where communities from the Global South can contribute their cultural and language resources on their own terms while also steering the direction of AI ecosystems in the Global South. The Masakhane community in Africa is an important example, and many community-created language dataset initiatives are now adapting and adopting these licenses.

Big AI firms want to improve the functionalities of their AI models in non-European languages (dubiously termed “low-resource languages”) in order to “localize” and thereafter impose their AI models in Global South countries as part of the “AI for good” paradigm. AI firms are rushing to acquire and digitize linguistic and cultural resources of Global South communities, while at the same time, funnelling access to culture and knowledge through the narrow prism of their AI tools and flooding the Internet with AI slop. Microsoft’s ELLORA initiative and the Google-funded Vaani initiative in India are both efforts to crowdsource local language snippets and create openly licensed, training datasets, which in turn allow Big Tech to use all these datasets at zero cost to further entrench their market dominance with multilingual AI tools. Against this context, the new anti-extractivist licenses become strategic tools for communities to retain control over their cultural and linguistic resources.

Combining strategic subversion of copyright with ethical use restrictions

Another strategy is licenses with ethical use restrictions – which are additional obligations on licensees, over and above the inversion of copyright. Examples range from Coraline Ehmke’s Hippocratic License, which prohibits downstream use that violates fundamental human rights norms, to licenses with “behavioural use restrictions” in for AI training datasets, code and models which prohibit downstream use for fully automated decision-making or predictive policing, among others.

Another strategy is licenses with ethical use restrictions – which are additional obligations on licensees, over and above the inversion of copyright.

Other licensing initiatives require licensees to pledge certain commitments. The UsageRights license for instance, requires licensees to evaluate the social and environmental impact of their actions with the licensed dataset, publish their findings openly, and commit to ecological redirection. Another licensing framework proposed by researchers for mixed datasets (i.e. datasets containing both personal and non-personal data elements) requires licensees to commit to a privacy pledge, and comply with any request for erasure of personal data made by an individual whose personal data is contained in the dataset. Licensees are also required to transmit these commitments downstream.

Takeaways for advocacy around the digital, cultural, and AI commons

The proliferation of licenses with more conditions certainly raises important questions of legal interoperability (where resources licensed under different licenses are sought to be remixed), virality (extending license conditions attached to training datasets to AI models and AI-generated outputs), and enforcement (where AI firms have and continue to ignore license terms).

More fundamentally, they also raise questions about the normative justifications of copyright law, a point raised by many at a round table organised by IT for Change on the sidelines of the AI Impact Summit.  Even if strategic, the reliance on private property rights (notably the right to exclude) as a tool to limit extractive use of the digital and cultural commons needs to be balanced against a more explicitly political approach to commoning rooted in more communal forms of cultural production and knowledge sharing and mutualization of labor and value.  At the same time, we should also be wary of romanticizing the public domain. As the indigenous data sovereignty movement reminds us, the public domain of traditional knowledge, for instance, is often treated as a “free” resource for proprietary innovations.

To this extent, we cannot (and should not) put all our promises into subversions of copyright law. While changes to copyright law are necessary, positive protections for the public domain and sanctions for extractions from the public domain need to be legislated. But on the other hand, strategic use of copyright through new licenses needs to be combined with other structural changes. It is important to address the ways in which generative AI is impacting our individual and collective capacity to engage in cultural and knowledge production.

Where AI models are trained on publicly available data, some commentators have proposed a tax or levy to be imposed on AI companies by national governments (instead of tax breaks for AI hyperscalers). This can then be used to invest in local stewardship of cultural resources, GLAM (Galleries, Libraries, Archives and Museums) and knowledge institutions, and also to fund much-needed AI and digital literacy efforts. Even if gradual, there is also a need to redirect public expenditure away from securing the services of AI firms towards local efforts for cultural and knowledge preservation.

Even if gradual, there is also a need to redirect public expenditure away from securing the services of AI firms towards local efforts for cultural and knowledge preservation.

Put simply, licenses are one tool in the toolbox to both preserve and nourish digital and cultural commons in the current AI innovation paradigm. But licenses need to be combined with more structural shifts to create a vibrant, democratic, safe and resilient information and cultural ecosystem.