Beyond Open Source: AI Licensing and the Push for Equity

In January 2025, DeepSeek released R1, a reasoning model designed for tasks such as mathematics, coding and logical reasoning. DeepSeek released the model's code and weights under the MIT Licence, one of the most permissive licences available. The terms allow commercial use, modification, and derivative works, including training other models on R1 (DeepSeek-AI 2025).

The release was widely seen as a turning point because it challenged the long-held assumption that only wealthy countries and large companies could access advanced AI (Quansah 2025). For Africa, the significance was both practical and symbolic. Since R1 was openly licensed, African developers could download it, adapt it and fine-tune it for local languages and markets, without paying licence fees or seeking permission.

The release also intensified the debate on open-source AI and was followed by a wave of open-weight releases. Hugging Face reports that open releases from Chinese companies became noticeably more active between February and July 2025. For example, Baidu went from no releases on its platform in 2024 to more than 100 in 2025, and ByteDance and Tencent increased their releases eight to nine times (Hugging Face 2026). In the United States, OpenAI released gpt-oss in August 2025, its first open-weight language model since GPT-2 in 2019, under the permissive Apache 2.0 licence (The Register 2025). In July 2025, the US government's AI Action Plan also called for a supportive environment for open-source and open-weight models (The White House 2025: 4).

Although R1 and similar open-weight releases would not resolve Africa's constraints in developing and deploying AI on its own, the continent still needs computing power, data infrastructure, research capacity, and supportive policy (Quansah 2025); the call for openness in African AI has continued beyond the DeepSeek moment. In Africa, open source is presented as an equaliser, as free access to code, models, and data would let African developers avoid building from scratch and adapt global tools to local languages and needs. Governments have articulated similar positions. In September 2025, Nigeria's Minister of Communications, Innovation and Digital Economy unveiled N-ATLaS, an open-source language model for Yoruba, Hausa, Igbo and Nigerian-accented English, which coverage presented as part of Nigeria's pursuit of AI sovereignty (TechCabal 2025).

In practice, however, licences determine what counts as open. A licence specifies who may use a model or dataset, for what purposes, on what conditions, and may also specify the governing law and jurisdictions for resolving disputes. Standard open licences grant every user the same rights. When these licences are applied to AI, this defining feature raises questions for Africa that the debate on openness often overlooks. When an African developer and a multinational technology company receive identical terms for the same model or dataset, but differ greatly in computing power, capital and access to markets, who benefits most from openness (Tarkowski 2026)? Researchers have found that openness alone does not reduce AI power concentration, and that the language of openness is often used in ways that increase it (Widder et al. 2024). African data and language communities have raised similar concerns, particularly that Global South-created open resources risk being absorbed into closed AI systems, a dynamic described as "commonswashing" (Chandrasekhar and Ávila 2026). Consultations with these communities show that they do not object to sharing their data in principle. Their concern is that current open-sharing arrangements work against their interests (Tarkowski 2026).

This article therefore considers how AI licensing functions and what it seeks to achieve. It then examines how African stakeholders have used licensing to address equity concerns, and assesses how recent global and policy developments affect these initiatives.

AI licensing and open source

A licence is a legal permission that tells others what they may do with something that a licensor controls, and on what conditions. In software history, licences became important as computers became commercial products. Early software was often shared freely among researchers and manufacturers, but later it was treated as private property, and licences controlled who could use, copy, or modify it. In response, the free software movement released the GNU General Public License in 1989, which used copyright law to keep software free to share, and in 1998 the Open Source Initiative published a definition of open source to set common standards for such licences (ETHW n.d.).

Two broad families of open licences emerged from this period:

• Copyleft licences, such as the GNU General Public License, which require anyone who distributes modified versions to release them under the same terms.

• Permissive licences, such as the MIT, BSD and Apache licences, allow modified versions to be distributed under different terms, usually requiring little more than attribution. The Apache License 2.0, introduced in 2004, also grants users a licence to any patents covering the software (University of Pittsburgh Library System n.d.).

Most AI licences today build on these models. Some AI developers release models under existing permissive licences, as DeepSeek did with the MIT Licence. Others have written new AI licences, including behavioural clauses first proposed in 2018, that restrict certain applications of a model (McDuff et al. 2024).

The growing use of software licences for AI models has exposed a practical difficulty. In AI, the subject of a licence is more complicated, because a model is a bundle of separate parts. The Model Openness Framework, developed with the Linux Foundation, identifies 17 components across a model's lifecycle (White et al. 2024). For policy purposes, they can be grouped into six layers:

 

 

The weights are among the most valuable layers, and most model licences are written around them. Yet legal scholars argue that copyright largely does not protect weights and outputs, which calls into question whether there is any protectable subject matter to license. Part of their reasoning is that weights are produced by an automated training process with minimal human authorship, and that even if this hurdle were cleared, copyright would likely treat weights as functional artefacts that copyright does not protect (Henderson and Lemley 2025). Without copyright protection, a licence on weights may operate only as a contract, binding those who agreed to it.

The rules on outputs are still developing. In the United States, a federal appeals court held in 2025 that a work generated entirely by AI, with no human author, cannot be registered for copyright, and the Supreme Court declined to review that decision in March 2026 (Cleary Gottlieb 2025; SCOTUSblog 2026). The US Copyright Office has stated that AI-assisted works are protectable only when a human contributes sufficient expressive elements, and that prompts alone are insufficient (US Copyright Office 2025). In August 2025, Kenya's Copyright Tribunal held that Kenyan law recognises authorship only for humans. It found that AI-generated aspects of a work are protected only where the author can show sufficient human intervention or effort giving the work an original character, as section 22(3)(a) of the Copyright Act requires. The Tribunal did not decide who the author was in that case because there was insufficient evidence of how the work was created (Copyright Tribunal of Kenya 2025). Both jurisdictions therefore require a meaningful human contribution before AI-assisted outputs can be protected.

From open source to open-source AI

Open source was first defined for software. The Open Source Initiative definition sets out the conditions a software licence must meet, including free redistribution, access to source code, and no discrimination against persons, groups or fields of use (Open Source Initiative n.d.).

As machine learning spread, it became clear that this definition did not map neatly onto AI. A model is more than code, as its behaviour depends on its trained weights and on the data used to train it. The Open Source Initiative began its work on AI after GitHub Copilot showed that machines could write functioning code and highlighted data's new role as an input to AI systems (Open Source Initiative 2024). Companies also raised concerns that they were describing their models as open source even when their licences restricted use or their training data was withheld (The New Stack 2024).

In response, the Open Source Initiative began researching the question in 2022 and started a public co-design process in 2023 to write an open-source definition for AI (Open Source Initiative 2024). More than 50 volunteers from nearly 30 countries, including Africa, took part, and the Open Source Initiative presented the draft at the Deep Learning Indaba in Dakar in September 2024 (Open Source Initiative 2024).

The resulting Open Source AI Definition 1.0, released on 28 October 2024, sets out four freedoms derived from free software:

• to use an AI system for any purpose without asking permission;

• to study how it works;

• to modify it; and

• to share it.

To exercise these freedoms, users must have access to three elements:

• detailed information about the training data;

• the complete code used to train and run the system; and

• the model's parameters, including its weights (Open Source Initiative 2024).

The treatment of training data was the most contested point. The definition does not require releasing the data itself, only enough information for a skilled person to build a substantially equivalent system. Some in the open-source community argued that any definition without a training-data requirement compromises the freedoms the original definition protects (The New Stack 2024). Others, including Mozilla, supported the compromise, arguing that a standard no developer could meet would do more harm than good (Open Source Initiative 2024). Meta said it disagreed with the definition, under which its Llama models do not qualify (The Verge 2024).

The debate produced a distinction that is now widely used. A model whose weights can be downloaded is often called "open weight", while "open-source AI" is reserved for systems that also provide the code and data information needed to study and rebuild them. Other frameworks take a graded approach. The Model Openness Framework ranks models into three classes based on how many components are released under permissive open licences (White et al. 2024).

Many models described as open still fall short of these standards. A survey of more than 45 generative AI systems found that many are at most open-weight, with developers withholding information about training data. This practice has been described as “open-washing" (Liesenfeld and Dingemanse 2024). DeepSeek’s R1 illustrates the distinction. Its weights are released under a fully permissive licence, but DeepSeek did not release the datasets or training code used to build it, as Hugging Face researchers who set out to reproduce the model noted at launch (Hugging Face 2025). A peer-reviewed version of the R1 paper, published in September 2025, added training-data samples and a description of how the data was generated, but it did not release the full training data, and the training framework remains internal to DeepSeek (DeepSeek-AI 2025). R1 is therefore more accurately described as an open-weight model.

The objectives of AI licensing

AI licences serve several purposes, and a single licence often combines more than one. Comparing the licences discussed in this article shows seven common aims.

• Granting permission and legal certainty: Permissive licences such as the MIT Licence clearly inform users that they may use, change and share a model. DeepSeek's choice of the MIT Licence for R1 is an example.

• Controlling commercial use and scale: Some licences allow free use up to a threshold. Meta's Llama 3 licence requires companies with more than 700 million monthly active users to request a separate licence (Meta 2024). Nigeria's N-ATLaS model sets its threshold at 1,000 active end-users (NCAIR 2025).

• Restricting harmful uses: Responsible AI Licences add behavioural use clauses that prohibit certain applications (Hugging Face 2022). Their authors argue that licences can enforce responsible use where legislation is slow or difficult to pass (Contractor et al. 2022).

• Protecting a competitive position: OpenAI's terms of use prohibit users from using its outputs to develop models that compete with OpenAI (OpenAI 2026). Meta's Llama 3 licence similarly prohibits using Llama outputs to improve other large language models (Meta 2024).

• Securing credit and traceability: Licences often require attribution or naming. Llama 3 derivatives must display "Built with Meta Llama 3", and renamed versions of N-ATLaS must carry the suffix "Powered by Awarri".

• Allocating risk: Almost all model licences provide the model "as is", without warranties, so that users bear the risk of errors and harms.

• Sharing benefits: A newer group of licences, many developed in Africa, uses licence terms to return value to the communities whose data trains AI. The next section discusses these.

These aims can pull in different directions. Permission and openness favour wide access, while protection, restriction and benefit-sharing add conditions. How a licence balances them therefore reflects the interests of its authors.

 

Africa and AI licensing: Where do we stand?

1. Why African actors are rethinking licences

As stated earlier, standard open licences treat every user the same, despite differing abilities to use them. These differences stem from disparities in computing power, capital, and other data needed to build products. This reflects a wider pattern in AI: even the most open AI systems depend on resources held by a few large companies, and openness alone does not reduce that concentration of power (Widder et al. 2024).

Licensing practice in this area is also inconsistent, as seen in an audit of more than 1,800 text datasets which found that licences were missing in over 70% of cases and wrong in over 50% on popular hosting sites, and that low-resource language data tended to carry restrictive licences (Longpre et al. 2024). Communities that create African language data frequently have limited influence over how it is subsequently reused.

2. Current licensing approaches in Africa

Over the past two years, African researchers, companies, and governments have developed licences with different terms for different users, depending on location, ownership, or scale of use. Two of these, the Nwulite Obodo Open Data Licence (NOODL) and the Esethu licence, are designed for African language datasets and require users outside Africa to share benefits or pay fees. The third, the terms of use for Nigeria’s N-ATLaS model, require a commercial licence once a user’s deployment exceeds a set number of end-users. The table below compares the three, and the paragraphs that follow describe each in turn.

 

The NOODL, developed in 2024, builds on standard open licences. It uses a single licence between three parties: the owners of African datasets; users in or from Africa and developing nations; and users outside Africa and developing nations (Okorie and Omino 2025: 102). Under this licence:

1. Both groups of users may use the data, but must share any modifications, adaptations or derivatives with users in Africa and developing nations under the same licence. 

2. Users in Africa and developing countries may use the data commercially only outside Africa and developing nations, which prevents them from selling products built on the data back to the communities that created it. 

3. Users outside Africa and developing countries carry the same obligations and must also provide a benefit specified by the community sharing the dataset, such as monetary compensation, access to infrastructure, access to human and physical resources, or equal collaboration.

Its first deployment was the DhoNam Dholuo speech dataset in Kenya (Tarkowski 2026).

The Esethu licence, developed by Lelapa AI and inspired by NOODL, allows free research use and commercial use by African entities. Non-African commercial users pay a fee, and the proceeds must be reinvested in creating more African language data. An entity counts as African if it is registered and headquartered in Africa, or is mostly African-owned (Rajab et al. 2025).

N-ATLaS, Nigeria's government-backed model, is a fine-tuned version of Meta's Llama 3. Its terms allow free use, adaptation, and redistribution for up to 1,000 active end-users; prohibit uses such as surveillance and military deployment; and are governed by Nigerian law (NCAIR 2025).

3. Coverage and limitations of these licences

NOODL and Esethu govern datasets, while N-ATLaS is the main African example of a model licence. This leaves a significant gap, because it is uncertain whether a data licence's terms continue to apply once the data has been used to train a model. Taken together, the three licences face seven challenges in achieving their aims of sharing benefits with data communities and widening access on fair terms:

a. Uncertain legal basis: A data licence takes effect when a person copies or adapts the licensed data. It remains unsettled whether a trained model’s weights constitute a copy or adaptation of the data used to train them (Henderson and Lemley 2025). Analysing Creative Commons share-alike licences under EU law, Szkalej and Senftleben conclude that, in most cases, a trained model breaks the link with the licence because it is unlikely to contain protected parts of the training works. The main exception is where a model has memorised protected material (Szkalej and Senftleben 2024: paras 14-16, 42). Copyright exceptions add a further layer. Creative Commons licences do not apply where a copyright exception covers the use, and in places like the EU, text- and data-mining exceptions can cover the training process itself (Szkalej and Senftleben 2024: paras 21-24). These conclusions rest on EU law; therefore, African copyright laws have not been tested on the same questions. The same uncertainty affects N-ATLaS. If its weights are not protected by copyright, its terms of use operate only as a contract.

b. Limited reach of contract: A licence normally draws its force from copyright where the licensor holds exclusive rights over the work, thus anyone who copies or adapts it without accepting the licence infringes those rights. Where the material is not protected by copyright, as may be the case for model weights and for a model trained on licensed data, there are no exclusive rights to enforce. The licence can then operate only as a contract, binding only those who accepted its terms (Henderson and Lemley 2025). The developer who downloaded a dataset may be bound by its terms. Companies that later obtain, adapt or deploy the resulting model have no agreement with the data community and are therefore unlikely to be bound. Licence obligations could reach these later users only through a chain of contracts, in which each recipient accepts the licence terms and requires the next recipient to do the same. Even such a chain would have gaps, since anyone who obtained the model outside it would not be bound if the model and its outputs are not protected by copyright (Szkalej and Senftleben 2024: paras 53–58, 72). The N-ATLaS terms face the same limitation: someone who obtains the model from anyone other than NCAIR may not be bound by them.

c. Difficulty of attribution: Models are typically trained on many datasets at once, and no established method isolates the contribution of any single dataset to a model’s value. Benefit-sharing obligations that depend on such a contribution are therefore difficult to calculate and apply. Tracing is a further difficulty. Where the licensed material makes up only a small part of the training data, it may be impossible to link the model's outputs to that material, making a breach of the licence hard to prove (Szkalej and Senftleben 2024: para 59).

d. Practical barriers to enforcement: Benefit-sharing requires continuous monitoring of how data is used (Tarkowski 2026). At the model level, a data community would also need to show that its data was used in a model that may be hosted in another country, and then pursue a remedy across borders. NOODL’s designers acknowledge that once datasets are publicly released under an open licence, licensors may be unable to trace, let alone act against, those who breach its terms (Okorie and Omino 2025: 104). The N-ATLaS user threshold raises a similar problem, as monitoring how many end-users a deployment serves would be difficult once the model has been downloaded. These requirements are demanding even for well-resourced actors. AI developers have not yet gone to court for damages or orders to enforce the terms attached to their own models (Henderson and Lemley 2025).

e. Limited use and gaps in coverage: NOODL has so far been applied to a single dataset, and its benefit-sharing framework is not yet fully developed (Tarkowski 2026). The author also notes that NOODL’s obligations for developed countries do not extend to China, where several major AI companies are based, and adds that the gap arises because the licence treats all countries the World Bank does not classify as high-income as developing countries (Tarkowski 2026). In July 2026, the World Bank still classified China as an upper-middle-income economy (International Sociological Association 2026). However, China's income per person is close to the high-income threshold, and Chinese commentators expect it to be reclassified within one to two years (Asia News Network 2026). Once that happens, Chinese users would fall within the benefit-sharing tier.

f. Inherited and overlapping terms: N-ATLaS is a fine-tuned version of Llama 3; therefore, Meta's licence applies alongside the N-ATLaS terms. The two sets of terms impose different naming requirements. Llama 3 derivatives must begin their names with “Llama 3”, while the N-ATLaS terms require the suffix “Powered by Awarri”. They also choose different governing laws, California for Llama 3 and Nigeria for N-ATLaS (Meta 2024; NCAIR 2025). Neither set of terms refers to the other, and neither says which prevails if they conflict. The Llama 3 licence also prohibits using its outputs to improve any other large language model apart from Llama 3 and its derivatives (Meta 2024). This restriction would appear to apply to N-ATLaS as well. If so, Nigerian developers could not use N-ATLaS outputs, such as synthetic text in Yoruba, Hausa or Igbo, to train a separate model built on a different base. This sits uneasily with the AI sovereignty aims attached to the model at its launch.

g. Conditions that limit openness: Free use of N-ATLaS ends at 1,000 active end-users, after which a separate commercial licence is needed (NCAIR 2025). A Nigerian start-up that builds a successful product on the model would therefore face a permission barrier as its product begins to scale. The user cap and the prohibited uses also restrict who may use the model and for what purposes. Although N-ATLaS was presented as an open-source model, it would therefore not qualify under the Open Source Definition or the Open Source AI Definition (Open Source Initiative n.d.). For the same reason, it may fall outside the EU AI Act’s open-source exemption, which requires a licence that allows free use, modification and distribution (European Union 2024).

4. The equity debate

As seen so far, the main measures of openness and the African licences pull in different directions. The Open Source Definition forbids discrimination between users and restrictions on fields of use, and the Model Openness Framework rewards permissive standard licences. NOODL, Esethu, and N-ATLaS all add usage conditions; under these measures, they count as less open.

Their designers argue that this differentiation is deliberate and necessary, because applying identical terms to users with unequal resources would itself produce unequal outcomes. NOODL’s developers draw on the South African Constitutional Court's decision in Blind SA, which held that the law cannot treat people as similarly situated when they are not. Applied to datasets, they argue that giving identical terms to users with very different access to computing power and infrastructure produces inequity, and that equitable openness requires different conditions for better-resourced users (Okorie and Omino 2025: 100-101). Similar arguments come from Indigenous data governance. The CARE Principles (collective benefit, authority to control, responsibility and ethics) were developed in response to the tension between protecting Indigenous rights in data and supporting open data and machine learning. They present community control and collective benefit as complements to the FAIR principles of findable, accessible, interoperable and reusable data (Carroll et al. 2020). Te Hiku Media's Kaitiakitanga licence applies this approach to Māori language data, keeping data sovereignty with local communities and prohibiting uses such as surveillance and discrimination (ITU 2022).

The differentiation approach also faces criticism and unresolved difficulties. NOODL has encountered resistance from proponents of openness, who regard equity licences as a potential enclosure of the commons. Its designers also concede that adding more licences contributes to a proliferation of options that may confuse dataset creators and users (Tarkowski 2026; Okorie and Omino 2025: 103).

There is also no agreement on how to group users. Esethu separates African from non-African users because its developers argue that power inequalities also exist among countries in the Global South (Tarkowski 2026), while for some datasets, grouping users by sector may work better than grouping them by location (Okorie and Omino 2025: 103). For policymakers, the central question is whether openness should still be judged mainly by the absence of conditions, or whether conditions intended to reduce inequality should also be accepted as legitimate.

Global developments and the OpenMDW 1.1 licence

While African actors have experimented with conditional licences, the global trend has shifted toward simple, permissive licences written specifically for AI. The most significant recent example is the OpenMDW licence (Open Model, Data and Weights).

The Linux Foundation released OpenMDW 1.1 on 28 May 2026, developed with Amazon, Meta, IBM, Microsoft, and others (TechTarget 2026). NVIDIA adopted it in June 2026 for several of its model families, and in mid-August 2026 the Linux Foundation submitted it to the Open Source Initiative for approval as an open-source licence, and it is still waiting for approval (BetaNews 2026; Techstrong 2026).

The licence text is short. Its main features are as follows (Linux Foundation 2026):

• One licence for the whole model: It covers the model's architecture and weights, together with any data, documentation and software released with it.

• Broad permission: It allows anyone to use the materials without restriction, under all copyright, patent, database and trade secret rights they contain. Naming all these rights addresses uncertainty about whether weights are protected by copyright.

• Few conditions: Anyone who redistributes the materials must include a copy of the licence and keep the original notices.

• No control over outputs: The licence places no restrictions or obligations on outputs generated with the model.

• A litigation clause: If a user sues, claiming the materials infringe a patent or copyright, they lose their rights under the licence unless the suit responds to one brought against them first.

• Responsibility for third-party rights: Users are responsible for clearing any rights of other people that may be embodied in the materials, such as copyright in training data.

OpenMDW is a useful step towards clarity, and it resolves several of the problems described earlier in this article. From an African perspective, however, it also confirms the mainstream direction. It treats all users identically and contains no mechanism for benefit-sharing. It also places the duty to clear rights in training data on users, leaving data communities with no practical route to recognition or return. The licences being developed in Africa and the licences being standardised globally are therefore moving in different directions.

Policy responses and enforcement challenges

1. The EU AI Act and open source

The EU AI Act (European Union 2024) is the most developed example of a law that links regulatory obligations to licensing. Under Article 53(1), every provider of a general-purpose AI model must prepare technical documentation, provide information to downstream developers, adopt a copyright policy and publish a summary of the content used for training. Article 53(2) waives the first two obligations for models released under a free and open-source licence that allows access, use, modification and distribution, provided the weights and information on the model's architecture and use are made publicly available. The waiver does not apply to models with systemic risk. Under Article 51(2), a model is presumed to have the high-impact capabilities that trigger this classification when the computing power used to train it exceeds 10^25 floating-point operations, so the largest open models are regulated regardless of their licence.

This approach has two consequences relevant to this article. First, the Act gives legal weight to the term "open source" without defining it precisely, which researchers warn incentivises open-washing (Liesenfeld and Dingemanse 2024). Second, the waiver rewards licences without conditions on use. A licence with a user cap, such as N-ATLaS, or with benefit-sharing tiers, such as NOODL, may not qualify.

2. Enforcement challenges, even for large companies

Licences are only as strong as their enforceability. Beyond suspending accounts, as in the disputes between Anthropic and DeepSeek, no AI model developer has tried to enforce its terms of use through monetary claims or court orders. Legal scholars argue that these terms are likely unenforceable, partly because weights and outputs are largely not protected by copyright (Henderson and Lemley 2025).

Another DeepSeek episode illustrates the point. OpenAI said it had evidence that DeepSeek had used its outputs to train a competing model, contrary to its terms of service. OpenAI and Microsoft reportedly blocked accounts suspected of being linked to DeepSeek (Bloomberg 2025). DeepSeek, for its part, released R1 under a licence that expressly permits using the model to train other models (DeepSeek-AI 2025).

If one of the best-resourced AI companies relied mainly on blocking accounts to protect its terms, African data communities face an even greater challenge. A community holding a NOODL or Esethu licence would need to show that a model hosted abroad used its data, then take action across borders. This suggests that licences alone are unlikely to secure benefit-sharing, and that they may need support from public law and from institutions able to act for communities. NOODL’s authors propose that governments establish and enforce a standard for AI developers when using others' datasets (Okorie and Omino 2025: 104), and Tarkowski notes that NOODL requires a trusted institution to manage access and ensure its terms are honoured.

Conclusion

The DeepSeek moment showed that permissive licensing can widen access to powerful AI, including for developers in Africa. It also showed the limits of licensing, since even a leading AI company relied mainly on blocking accounts when it believed its terms had been broken.

For Africa, the central issue is who benefits from openness. Open licences give everyone the same rights, but African developers and data communities are not equally placed to use them. NOODL, Esethu and N-ATLaS are early attempts to address this through licence terms. They remain experiments: they are largely limited to datasets, no court decisions on them have been reported, and they sit uneasily with the dominant definitions of open source and with the EU AI Act's open-source provisions. Global standardisation, through OpenMDW and similar licences, is moving towards fewer conditions.

Several steps could help close this gap. Some build on proposals already in the literature; others are offered here for consideration:

• Design licences that follow the data: Dataset licences could make permission to train conditional on the developer releasing the resulting model under the same terms and binding later recipients, the contractual chain proposed by Szkalej and Senftleben (2024). In the EU, rightholders can strengthen their bargaining position by reserving their rights against commercial text and data mining, preferably in machine-readable form. African dataset publishers whose data developers use under EU law could explore the same mechanism.

• Use transparency obligations as evidence: Frameworks such as the EU AI Act require all general-purpose model providers, including those releasing open models, to adopt a copyright policy and publish a summary of training content (European Union 2024: Art. 53(1)(c)–(d)). African data stewards could use these summaries to detect use of their datasets, and African regulators could consider equivalent disclosure requirements for models offered in their markets.

• Use public procurement: Governments are major buyers of AI systems. Procurement terms could require vendors to document the provenance of their training data and confirm compliance with licences attached to African datasets, giving licence terms practical force without new legislation.

• Pool enforcement capacity: Individual communities cannot monitor global model releases. Collective institutions, modelled on the collective management organisations that administer copyright on behalf of rightholders, could hold and enforce licences across many datasets.

• Reform funding conditions: Funders should let grantees choose from a range of open licences rather than requiring a single licence (Okorie and Omino 2025: 103–104). They could also go further. Funders could set aside budgets to administer and monitor licences, require grantees to agree licensing plans with data communities, and require the developers they fund to respect the terms of any African datasets used in training.

• Coordinate regionally: Regional bodies such as the African Regional Intellectual Property Organisation or the African Union could develop model clauses and common definitions for equity licences. This would reduce the proliferation problem and give licence terms a consistent meaning across African jurisdictions.

• Recognise equity conditions in openness standards: Standard-setters such as the Open Source Initiative and the authors of the Model Openness Framework could consider a distinct, clearly labelled category for licences that attach benefit-sharing conditions, so that these licences are not simply classified as closed.

• Build and license African models: African developers and governments that train or adapt their own models gain control over the terms attached to them, as Nigeria has done with N-ATLaS, although inherited licences from base models still apply.

Licensing alone will not deliver equity in AI. However, it is one of the few tools African actors can use directly, and its future direction is still being decided.