Organizing Against the Algorithm: Collective Response as a Governance Lever for Gradual Disempowerment in South Africa's AI Infrastructure Buildout

I like this paper a lot. The central move — taking the gradual disempowerment framework and asking what bottom-up resistance to it looks like, then pulling in labour economics and social movement theory to answer that — is genuinely original. Most of the disempowerment conversation is top-down (alignment, governance from above), and you're right that the question of whether affected populations can organize against it is underweighted. Bringing those two literatures into contact is a real contribution.

The strongest part is the comparative case framework in Section 3. Scoring the four cases on trigger, actor, demand structure, and outcome, and then pulling out the pattern — collective response forms when there's a visible harm and an identifiable target, and AI labour substitution is missing both turns a story into a falsifiable claim. And then Section 5 maps each recommendation directly onto a missing condition (disclosure creates the target, early-warning creates advance notice, the civil-society seat creates standing). That tight link between diagnosis and prescription is what makes it more than commentary. The Cassava case is well-chosen and I appreciate that you took the sovereignty framing seriously instead of treating the project as a villain the energy/labour/capital-structure breakdown is sharp.The honest limit is that this is conceptual. Four cases generate a hypothesis, they don't test one, and you say so plainly, which I respect. The thing that would worry me if I were building on this: the early-warning mechanism in 5.2 is a surveillance system pointed at deployment, and you flag — correctly — that it could be turned around to monitor and suppress organizing instead. I'd want that risk treated as a core design constraint, not a closing caveat, because "worker-controlled data governance" is doing enormous load-bearing work in one sentence. Two things would lift it: even a handful of interviews with unions near the Johannesburg or Cape Town sites would move this from armchair analysis toward evidence, and a short legal check on whether South Africa's labour framework can even accommodate diffuse, non-single-employer disruption would tell you whether 5.3 is actionable or aspirational. The dual-use note about this energy being redirectable at AI that has real developmental benefit was an honest thing to include.

The paper brings something fresh to the table by looking at collective action as a real form of AI governance, instead of just focusing on regulations or technical alignment like most people do. I thought the comparison of those four South African cases was smart—it helped pull out clear conditions for when collective action actually works, like visible harm, obvious targets, and straightforward demands. Then using that lens to explain why the Cassava AI Factory hasn’t sparked much organized pushback yet felt like a solid bridge between gradual disempowerment and labor economics, which you don’t see discussed much in AI governance conversations.

The biggest weakness is that the empirical side feels a bit thin. The Cassava case raises good questions around energy costs, job losses, and foreign companies capturing most of the value, but a lot of it stays at the level of open questions rather than clear evidence. For instance, it notes that the actual distribution of costs and benefits hasn’t been made public, so it’s hard to tell if the suggested fixes are addressing real harms or just potential ones. The four-case comparison also shows interesting patterns, but four examples aren’t really enough to prove those are the main drivers of successful organizing.

The policy suggestions make sense given the framework—especially things like requiring companies to disclose how benefits and costs are distributed, and setting up early-warning systems for sectors. Still, the paper would feel a lot stronger if it could show these changes would actually shift outcomes on the ground, not just increase transparency. Some interviews with unions, workers, or policymakers, or comparisons with similar big infrastructure projects, would have helped back that up.

Overall, it’s a thoughtful conceptual piece with a clear theory of change and an honest acknowledgment of its limits. The next step is to test the framework with real primary evidence and show that collective organizing can actually work as a practical governance tool in AI projects, rather than just a promising idea.

This is a solid and specific case study that identified three core questions for why South Africa’s collective civil society response has failed to materialize in the case of the development of a new AI data center in Johannesburg. It names the key actors influencing the development of the data center. It lists 3 clear policy recommendations to fill this gap. The limitations section was thoughtful and accurate. The inclusion of the role of labor unions and collective bargaining was a novel and powerful contribution.

It would benefit from more concrete vision for the Sectoral Early-Warning Mechanism, as it was unclear from the text what the purpose and proposed structure of this mechanism could be and how it would measure or identify displacement risk. I also suggest the authors take a more expansive view of the civil society observer recommendation and consider broader public-private partnerships and working groups with civil society inputs, instead of project-based contributions which may run into complications. It would strengthen the paper to include additional state-linked entities beyond CSIR that would be instrumental in adopting these policy recommendations for immediate impact and call to action for specific agencies. It would also strengthen the analysis to include which civil society organizations could be instrumental and relevant for this specific case study.

The paper shows structural policy changes that could improve civil society’s responses to AI developments, and it hypothesizes that the same strategy can be applied to AI developments, but it does not provide a clear path forward to if this collective action can be applied to AI successfully. It simply assumes that if it was successful in historic cases, this can be replicated in the case of AI data centers. The authors must explicitly bridge the gap between historic labor mobilization and AI infrastructure.