The dominant AI policy conversation has a structural problem: it is optimised for the wrong threat. While real, measurable harms — mass layoffs, soaring data-centre energy consumption, algorithmic bias, automated disinformation, AI-assisted targeting in warfare — demand immediate regulatory action, the global discourse has been captured by speculative existential risk narratives. The result is a policy environment that treats hypothetical superintelligence as more urgent than documented discrimination. This is not an accident. The doomsday narrative serves a specific commercial function. By framing advanced AI as a potential weapon of mass destruction, incumbent tech giants create the rhetorical preconditions for regulatory regimes that only they can satisfy. The argument is seductive: if AI is dangerous enough to end civilisation, surely only a "trusted few" — meaning well-resourced, Western-headquartered corporations — should be allowed to develop it. Under the banner of "safety," monopolists write rules that kneecap open-source competitors, the very models that offer the Global South its only realistic path to sovereign AI capability. The colonial parallel is not rhetorical flourish — it maps structurally. Frontier AI models cost upwards of $100 million in compute alone to train from scratch. The entire Global South is effectively locked out of foundational model development, forced into dependency on exported systems built on assumptions that do not reflect local realities. Kenya's AI-driven Social Health Authority system is a case study: an imported model, ignorant of the country's vast informal economy, systematically overestimated incomes and overcharged the poorest citizens for healthcare premiums. The failure was not a bug — it was the predictable outcome of dependency on externally designed systems. Open-source AI is the only viable bridge. It allows local innovators to adapt frontier capabilities to local data, local needs, and local development priorities — energy, agriculture, healthcare, climate mitigation — rather than the chatbot-centric roadmap set by Silicon Valley. Crushing open-source under safety regulation does not make the world safer. It eliminates the primary mechanism developing nations have to build independent tech ecosystems. The dependency trap extends beyond economics. AI monopolies create cognitive lock-in: users outsource critical thinking to proprietary systems, generating data that feeds back into more capable models that further erode human agency. This is not a Global South problem alone — wealthy nations face the same extraction cycle. The difference is that the Global South has no fallback position. The article's author argues that the path forward is neither AI bans nor unregulated development, but decentralised, transparent, globally distributed AI ecosystems. Safety research and rigorous testing for near-term harms must be funded aggressively — but not weaponised to consolidate power. Open-source innovation must be defended as a public good. Human accountability must remain absolute, resisting the anthropomorphisation of systems that possess no humanity. The core tension is now visible: every regulatory framework proposed in the name of safety must be interrogated for who it protects and who it excludes. The existential risk discourse is not wrong to exist — but its dominance in policymaking, at the expense of documented harms and equitable access, is itself a form of extraction.
