Even as AI is framed as a universal driver of growth, AI governance debates remain largely focused on downstream risks, leaving the conditions under which it is produced largely unexamined. In an effort to shift this focus, IT For Change along with other member organisations of the Global Digital Justice Forum and collaborators including the Planetary AI project convened the roundtable “AI Governance from the South: Redlines to Baselines”, alongside the India AI Impact Summit in New Delhi.
A core premise of the discussion was that AI must be understood as an infrastructure. This framing redirected attention away from the spectacle of AI applications towards the production networks that sustain them. Participants pointed to how large-scale models depend on layered forms of extraction, including data scraped from public domains, labour organised through opaque subcontracting chains, and resource-intensive physical systems. What is often presented as seamless innovation is in fact sustained by familiar patterns of extraction and arbitrage, whether in the outsourcing of data work across the Global South, or the siting of data centres in regions where environmental oversight is weak.
Yet prevailing governance approaches focus on identifying harmful use-cases (deepfakes, bias in hiring systems, misinformation, among others), and treat these underlying conditions as externalities. Participants pointed to, for instance, how growing scrutiny of biased algorithmic outputs is not matched by attention to the labour conditions under which training data is produced. The result is a global assemblage that persistently displaces ecological and social costs onto already vulnerable geographies in the South and siphons value to a small set of enterprises based predominantly in the North.
It is within this context the framework of baselines becomes analytically and politically powerful. Emerging from Global South experiences of extraction, this framing foregrounds the conditions under which AI systems are produced. Baselines articulate the minimum conditions that must hold across the AI value chain—for instance, lifecycle-level disclosure of labour inputs and impacts. This shifts the locus of governance upstream, from managing symptoms to structuring production.
A second thread, and one that resonates strongly with Planetary AI, was the insistence that AI development must be situated within a finite world. Current trajectories assume infinite scalability, demanding ever more data, compute, and infrastructure. However, AI expansion is already entangled in competing claims over resources. The siting of data centres, and the intensified extraction of minerals required for hardware, is placing growing pressures on land, water, and energy systems, often with limited participation of affected communities in decision-making. In response, participants called for AI infrastructure development to be subject to mandatory prior assessments of resource use and territorial impact. They further stressed that affected communities and local governments must not only be consulted in decisions surrounding AI infrastructure development, but must be guaranteed enforceable rights over environmental information and the power to halt projects that exceed local ecological thresholds.
A third concern centred on how control over critical inputs to AI (compute, high-value datasets, cloud infrastructure) is concentrated. This allows dominant enterprises to set the terms on which others can build, through pricing structures, licensing structures, and technical dependencies that are difficult to exit. Even publicly funded AI initiatives are often routed tied up with privately controlled infrastructures, locking states and public institutions into long-term dependencies. Participants argued that foundational digital infrastructures, including DPIs, must be treated as public goods and governed accordingly, with strong safeguards to sustain public value creation. They further underscored the need for interventions to decentralise control over AI inputs, including obligations over dominant firms to ensure equitable access for public institutions, researchers and small firms.
These reflections offer a counterpoint as well as a direction for AI governance processes, including the upcoming Global AI Dialogue in Geneva. Multilateral governance forums have continued to prioritize voluntary commitments framed as “responsible innovation”, even as the need for binding obligations has become starkly clearer. In contrast, the framework of baselines demands confronting the logics of extraction, concentration, and externalisation that underpin contemporary AI development. It centres the need for enforceable constraints and redistribution of control across the AI value chain. The credibility of AI governance processes increasingly depends on whether they can move beyond managing risks to breaking entrenched concentrations of power, thereby enabling alternative pathways for AI development.
To learn more about the issues and discussion points explored during the roundtable, please see the issues brief published here.