Whose Red Lines, Whose Oversight? From Human-in-the-Loop to Governance Infrastructure for AI in the Global South

**Problem.** International AI governance increasingly articulates red lines (uses, capabilities, or development conditions that should not be crossed), yet consensus is running ahead of enforceability. Binding governance concentrates on deployment and use, while development-stage harms in labor, data sovereignty, market concentration, environmental impact, dangerous capabilities, and participation remain under-governed and are especially salient from Global South perspectives; human oversight often remains a principle rather than an operational mechanism. **Central research question.** How can AI red lines become participatory and enforceable governance infrastructure when authority, monitoring, and intervention are distributed across humans, AI agents, organizations, regulators, civil society, and affected communities in Global South contexts? We ask *whose* red lines legitimately define unacceptable practice; *where* meaningful oversight sits when decisions emerge across human-agent-organizational chains; and *what* evidence, escalation, capacity, and remedy turn a red line into a binding constraint. **Red-line types.** We separate four types because each has different legitimate definers and enforcement levers: *legal prohibitions* (legislatures, courts; sanctions), *technical capability thresholds* (evaluators, standards bodies; access restriction), *institutional rules* such as procurement conditions (public agencies; contract terms), and *community-defined boundaries* (affected groups; contestation and public reason-giving). When types conflict, the framework records who proposed, contested, and decided each line, so that overrides become visible and reviewable. **Framework.** Governance is modeled as six linked functions, *define* $\to$ *observe* $\to$ *contest* $\to$ *escalate* $\to$ *enforce* $\to$ *redress*. For each function we ask whether an identifiable actor holds the four conditions of meaningful control: information, competence, authority, and practical intervention capacity. This yields a $6\times4$ *chain matrix*, scored per deployment from documents and interviews (0 absent, 1 partial, 2 present). A zero in the authority or capacity column marks a *breakpoint*: a red line that exists on paper but cannot bind. AI agents may support *observe* and *escalate* but hold no *define* or *enforce* authority, and their outputs count as evidence open to contest. **Illustrative case.** A ministry procures an externally built Arabic-language eligibility-scoring system. A procurement rule defines the red line, but logs sit with a foreign cloud provider (*observe*: no information), applicants cannot appeal a score (*contest*: no standing), and no domestic body can suspend the contract (*enforce*: no authority). The matrix locates three breakpoints that a single human in the loop would not reveal. **Contribution and Global South novelty.** First, the unit of analysis moves from a system or operator, the focus of existing audit and accountability frameworks, to the governance chain around it, with a participation-to-authority continuum (affected $\to$ consulted $\to$ represented $\to$ agenda-setting $\to$ co-deciding $\to$ enforcing) that addresses "participation washing." Second, oversight becomes an institutional-design problem scored at each link rather than a matter of human presence. Third, we propose a governance-capacity threshold extending UNESCO's Readiness Assessment Methodology. To avoid penalizing under-resourced jurisdictions, a failed threshold triggers obligations rather than exclusion: deployment proceeds only with provider- and funder-financed capacity (audit access, appeal channels, local escalation routes), and readiness is assessed jointly with local institutions and civil society. This is a research proposition, not existing policy. **Research agenda.** *This is early-stage work in progress.* Cases are chosen by diverse-case selection across MENA/North Africa, varying provider origin (imported or locally developed) and regulatory maturity, starting with public-sector procurement, Arabic and dialectal data annotation chains, and cloud-dependent agentic systems. Process tracing identifies the mechanisms that produce breakpoints; claims beyond the region concern mechanisms and scope conditions, not generalizations about the Global South as a whole. Participatory prototyping, with stakeholders recruited purposively across the four red-line types, will test a "red-line observatory" hosted by an independent academic and civil-society consortium. Its authority is evidentiary: it documents incidents and routes them to bodies with jurisdiction, and is evaluated by whether routed cases receive a formal response. The aim is to treat Global South actors not only as populations protected by AI governance but as agents able to define, observe, contest, enforce, and revise its boundaries.