Africa’s AI Rulemaking Reopens DABUS and Training-Data Copyright Questions
“Just AI” is resurfacing in African IP debates not because it makes for a neat slogan, but because the policy direction is becoming harder to ignore. After the African Union adopted its Continental AI Strategy in 2024, high-level policy dialogue in 2025 continued to urge African states to develop AI laws, regulations and national frameworks suited to their own conditions. AI governance on the continent is moving from principle to institutional design.
That shift has pushed two older disputes back into view. One is South Africa’s earlier acceptance of a patent application naming DABUS as inventor: was that a meaningful doctrinal break, or mainly a sign of what a deposit-based patent system can leave unresolved? The other is whether generative AI training on copyrighted works could fit within fair use or fair dealing exceptions in some African jurisdictions, and whether local creators should be compensated when their works feed commercial models. The harder policy question now is not whether Africa should encourage AI, but how to do so without further weakening the bargaining position of African creators, publishers and cultural industries.
Why “Just AI” is becoming an IP policy lens in Africa
The notable change is that African AI policy is no longer only about catching up with technology. The AU’s strategic direction is explicitly Africa-centric, development-focused, inclusive and responsible. Recent Research ICT Africa work on IP policy for “Just AI in Africa” sharpens that discussion by asking a more uncomfortable question: if AI training, deployment and monetisation follow existing patterns of platform concentration and global copyright asset control, who in Africa is left with the value? In many sectors, local media houses, artists, language communities and smaller creative businesses may supply the works and data while capturing very little of the upside.
In that sense, “Just AI” is not a soft normative add-on. It is a way of testing real design choices. Are copyright exceptions being drafted too broadly for commercial model training? Are patent rules being stretched by machine inventorship claims without resolving ownership and accountability? Are transparency duties strong enough to let creators know when their works have been ingested? A serious African AI policy conversation is becoming less about abstract innovation branding and more about who gets bargaining power, compensation and legal visibility in the new value chain.
DABUS in South Africa matters, but not in the way it is often presented
South Africa’s 2021 acceptance of a patent application listing DABUS as inventor remains one of the most cited AI-and-patent episodes in Africa. But it is easy to overread it. The more careful view is that the event forced the issue into public view without conclusively settling it for the continent. It did not amount to a full legislative rethink of inventorship, nor did it produce a judicial doctrine that African patent offices can simply replicate. South Africa’s patent system has long been known for its formal, non-substantive examination features, so the case exposed a legal gap as much as it signalled a policy choice.
That said, the DABUS episode still has real value for policymakers and applicants. It puts pressure on questions that cannot be postponed forever: must an inventor be a natural person, who can own rights in an autonomously generated technical solution, how should entitlement and assignment work, and what exactly is being certified when an inventor is named in a filing? For businesses, the practical lesson is clear. DABUS should not be treated as a shortcut route for AI inventorship claims in Africa. It is better understood as a warning that if definitions stay vague, downstream validity, ownership and enforcement disputes become harder, not easier.
The training-data fight turns on the limits of fair use and fair dealing
There is no single African answer to whether AI training can rely on copyright exceptions. The Research ICT Africa paper makes that point well. Fair use and fair dealing share a common lineage in common-law systems, but they do not operate identically across jurisdictions. Nigeria’s recently enacted copyright regime is more likely to leave some fair-dealing room for AI model development. Kenya’s Copyright Act expressly includes scientific research within fair dealing, while excluding computer programs. South Africa’s current Copyright Act, by contrast, still lacks both a broad flexible exception and a dedicated computational analysis provision.
That is why sweeping claims that “AI training is probably covered by fair use” are risky in an African context. For model developers, the real problem is not only doctrine but exposure. What was scraped, from where, under what licence logic, with what opt-out mechanism, and at what scale? Did the training set include news archives, books, music, images, code or local-language content? If the legal basis turns out to be wrong, the resulting liability may not come from one claimant but from thousands. In cross-border projects, dataset provenance, segmentation, licensing strategy and crawler governance are quickly becoming core IP compliance issues rather than side notes for later.
The real policy balance is between IP concentration and local creative income
The most grounded part of the current African debate is that value distribution is finally being discussed openly. South Africa’s Competition Commission has already linked AI and media sustainability, finding that AI chatbots and generative systems have used South African news content for training and development without fair compensation, while local publishers remain poorly positioned to negotiate or even manage opt-out tools. That moves the debate beyond theory. The issue is no longer only whether copyright law technically permits some training uses. It is also whether the economic architecture built around those uses leaves local creative markets weaker.
If policymakers simply import broad exceptions with no compensation, transparency or bargaining support, value may continue to flow outward from African publishers, musicians and writers to foreign model operators and dominant digital platforms. But a rigid prior-licensing rule for every act of training would also hit African researchers and startups first, because transaction costs are easier for larger incumbents to absorb. A more workable middle path would separate non-commercial research from large-scale commercial training, require meaningful disclosure and content-control tools, preserve room for local scientific and educational development, and consider collective licensing, levies or negotiated compensation where markets clearly fail to reward local creators.
For companies, the immediate work is practical rather than philosophical. Publishers should review contracts, machine-readable restrictions, crawler settings and proof of title. AI developers should keep auditable records of training sources, filtering logic, opt-out handling and local-content use. IP teams should keep the South African DABUS episode separate from the much wider copyright-exception analysis, rather than letting one headline case distort legal planning across the continent. Africa’s AI rulemaking window is now open. The real test will not be who says they support innovation, but who can build a framework in which African creative labour is not treated as a free input into someone else’s scale advantage.
