South Africa's national AI policy is being rewritten from scratch, and the scholar helping rewrite it has just told the government what she thinks is missing: rules for open-source AI licensing.
On July 2, 2026, Research ICT Africa (RIA), the Cape Town think tank led by Alison Gillwald, published "Rewriting the Rules for an Era of Just AI," warning that open-source licensing is at risk of becoming a vehicle for "digital colonialism" unless African governments write safeguards directly into law. The report's target is what it calls "openwashing" — big technology companies attaching an open-source label to a model while withholding the training data, compute, or fine-tuning pipelines a local developer would need to actually reproduce or audit it, all while drawing on African-language data and labor for free.
The timing is not incidental. Gillwald sits on the seven-member independent panel that South Africa's Department of Communications and Digital Technologies convened on May 13, 2026, to salvage the country's national AI policy — chaired by Wits University's Benjamin Rosman and including University of Pretoria computer scientist Vukosi Marivate, attorney Heather Irvine, and CSIR's Jabu Mtsweni, per TechCentral's reporting. The panel exists because the original draft policy — gazetted April 10, 2026 as Government Notice 3880 in Gazette 54477 — was formally withdrawn via Gazette 54478 on June 12, 2026 after editors at three academic journals confirmed that citations attributed to them in the policy's 67-reference bibliography did not exist, as ITWeb reported. Communications Minister Solly Malatsi's own explanation was blunt: the department's drafters had fed prompts into a generative AI tool and published the output without checking a single source — the same technology the policy was meant to govern.
The Steelman: Openwashing Is a Real Phenomenon
RIA's underlying concern is not manufactured. The Open Source Initiative — the body that maintains the canonical definition of open-source software — spent years building consensus around what it now calls the Open Source AI Definition (OSAID), precisely because model releases branded "open" routinely fail a basic test: could a skilled developer use the disclosed materials to build a substantially equivalent system? OSAID requires four freedoms — use, study, modify, and share — and demands enough detail about training data provenance, labeling, and sourcing to make that reproduction possible. Many models marketed as open-source today disclose weights but not data, which is exactly the gap RIA is pointing at. When African researchers, journalists, or public agencies build on a model whose training corpus and governance terms are opaque, they are absorbing downstream liability — bias, IP exposure, unreviewable outputs — without the audit rights the "open" label implies. That is a legitimate governance problem, not a rhetorical one, and a national AI policy that says nothing about it would be incomplete.
Where the Prescription Overreaches
But the leap from "openwashing exists" to "African governance frameworks must address it directly" through binding licensing rules risks solving a labeling problem with a market-access problem. Africa currently holds under 1% of global data-center capacity, and more than 36 African countries have data protection laws either in force or advancing, according to WeeTracker's July 2026 survey of the continent's digital-sovereignty push — meaning compliance fragmentation, not compute scarcity, is already the binding constraint on AI deployment in most African markets. A South African standard for what counts as sufficiently "open" AI, if it diverges from OSAID or from the EU AI Act's open-source carve-outs, adds exactly the kind of bespoke national licensing test that makes global labs reluctant to release anything at all into a jurisdiction — the opposite of RIA's own stated goal of ensuring Africans can inspect, adapt, and build on frontier models rather than merely consume them.
The better instrument sitting in plain sight is disclosure, not licensing gatekeeping. A policy could require that any AI system procured by a South African public body, or marketed in South Africa as "open source," publish a data-provenance statement aligned with the existing OSAID criteria — piggybacking on a standard already being fought over globally rather than inventing a parallel African one. That gives RIA's transparency goal real teeth without creating a compliance regime foreign labs might simply route around by skipping the market.
Why the Process Matters as Much as the Substance
The department's credibility problem compounds the stakes. A panel assembled to fix a policy discredited by unverified AI output has an obvious incentive to overcorrect toward maximal caution on every AI-adjacent question, including open-source licensing — RIA's timing, arriving weeks before the panel's expected consolidated report, will make it a reference point either way. The panel is targeting Cabinet resubmission by November 2026 and public consultation in January 2027. Whether South Africa becomes a model for open, evidence-based AI governance in Africa — or an example of licensing rules that quietly priced the continent out of the open-model ecosystem it is trying to join — will turn on whether the redraft distinguishes disclosure from restriction.