Since August 2025, providers of general-purpose AI models sold into the European Union have been required to publish summaries of the data used to train them, under Article 53(1)(d) of the EU AI Act. A review of those disclosures turns up African-language corpora, African digital publishers and African code repositories inside training sets for models now licensed across the continent, but no recorded instance of an African author, musician or developer being notified, asked, or paid for that use.

The gap is not accidental, this article argues, but built into the Act's design. The EU AI Act is a product-safety and market-regulation instrument, not a copyright-enforcement mechanism, and its disclosure template requires reporting by dataset and source category rather than by individual work. A line reading "web-crawled data including African-language content from Common Crawl" satisfies the letter of the law without giving any specific rights holder the evidence needed to bring a claim.

The underlying infringement, the author contends, is already established under African law. Nigeria's Copyright Act 2022 and Ghana's Copyright Act 2005 both recognise a reproduction right that large-scale scraping implicates, and the Berne Convention's three-step test permits exceptions only where they do not unreasonably prejudice an author's legitimate interests. This standard mass commercial scraping fails in jurisdictions with no text-and-data-mining exception.

Nigeria has no such exception; Kenya's is narrowly limited to research and news reporting; South Africa's long-stalled Copyright Amendment Bill, which would introduce fair-use exceptions, was found constitutional by the Constitutional Court in mid-2026 but remains unsigned. What African rights holders lack is not the underlying right but the practical means to enforce it against foreign developers, jurisdiction, discovery of training data, and enforceable judgments abroad.

The EU's own opt-out mechanism compounds the problem. Under the CDSM Directive, rights holders can reserve their works against text-and-data mining through machine-readable means, but the burden falls on the rights holder and the odds of an individual African author or small publisher having implemented such a reservation before being scraped, "approach zero." Whether that silence should count as consent is itself contested under Berne Article 5(2), which bars conditioning protection on formalities.

This article is sceptical that solutions from the US, UK or EU translate to African circumstances. US courts have reached mixed results applying fair-use doctrine to AI training. Bartz v. Anthropic found training on lawfully acquired books fair use while rejecting the same defence for pirated copies, but such rulings shield only US developers and offer African claimants no recourse. The UK abandoned two attempts at a text-and-data-mining exception amid industry opposition, most recently in March 2026.

This article proposes three responses: a joint ARIPO–OAPI protocol, modelled on the Marrakesh Treaty's cross-border framework, creating a non-waivable minimum remuneration right for AI training use of African works, administered through national collecting societies; coordinated monitoring by African collecting societies of Article 53(1)(d) disclosures, potentially through CISAC affiliations; and any future domestic text-and-data-mining exception carrying a mandatory remuneration obligation and an opt-in register.

We conclude that the AI Act's disclosures are the first evidentiary opening of their kind, however coarse. African institutions must now build an accountability framework on that opening and this will determine whether the AI transition redistributes value to African creators or whether it will simply produce a more detailed record of their exclusion from it.