The Non-Proliferation Trap: How “AI safety” is Used to Lock Out the (third) World
If you opened Twitter or LinkedIn over the past 48 hours, you couldn’t miss it. Your feed was almost certainly flooded with sensational posts, viral threads, and resignation letters from frontier AI lab researchers claiming that we are racing toward self-improving superintelligence that will end the world. We are watching people on the inside publicly announce they are stepping away because the risk of AI killing all humans within a few years is “too high,” while alignment leads openly chime in to confirm that yes, they believe there’s a real chance this technology wipes us out. It is apocalyptic, it is high-drama, and it is all over my “for you” page and probably yours too.
AI agency and autonomous capabilities do carry real risks, but when you look past the volume of these apocalyptic declarations, you have to ask a straightforward question: who actually benefits from this panic? For me, the sudden narrative that frontier AI is an existential threat so dangerous that it could end humanity isn’t just organic fear. It is setting the stage for a very specific political outcome, using a strategy we have seen play out before.
Time for a little history lesson. Remeber when the US developed the atomic bomb through the Manhattan Project? The world’s most powerful nations at the time immediately recognized that they possessed a technology capable of absolute dominance. To prevent anyone else from matching that power, they constructed a narrative around existential dread, arguing that allowing nuclear capability to spread would mean global annihilation.
That framing led directly to the Non-Proliferation Treaty. On paper, non-proliferation was presented as a noble effort to save humanity from self-destruction. In practice, it locked in a hypocritical cartel. A select group of nations retained their nuclear weapons, while the rest of the world was legally forbidden from ever developing their own. If a sovereign nation outside that circle attempted to build their own, they were met with sanctions, isolation, or military intervention under the disguise of global safety.
I believe that exact non-proliferation logic is now being weaponized to build a regulatory moat around frontier AI. The existential threat and sensational tweets and tv interventions are used to justify the need of strict government licensing, compute caps, and, more importantly, centralized oversight over who is allowed to train or run high-level models. The goal is to ensure that AI capabilities remain strictly enclosed within a small group of approved players: the dominant proprietary labs in the United States (openAI, anthropic, Google), sovereign domestic efforts in China (deepseek), and regional attempts in Europe (Mistral).
If proprietary AI models become the primary interface through which humanity accesses knowledge, culture, and digital infrastructure, concentrating those models within three or four corporate entities is a dangerous failure mode. Yann LeCun has argued this point consistently: “The biggest risk of AI is the concentration of power in a few dominant providers of proprietary AI assistants. The only solution to AI sovereignty is open source foundation models.”. When a model’s weights are closed, safety becomes an opaque promise made by an evaluator you cannot inspect. When model weights are open, communities can verify, recalibrate, fine-tune, and run infrastructure locally without fear of sudden access revocation or foreign policy choke-points.
For non-Western nations, closed-source “aligned” models represent a new form of digital paternalism. An evaluator trained on Silicon Valley or Beijing values will inherently treat local cultural contexts, languages, and political realities as out-of-distribution noise, or worse, as alignment failures. Restricting model weights doesn’t make the world safer; it simply guarantees that a handful of foreign corporations and governments maintain a monopoly over global digital intelligence.
This brings us to the most frustrating part of this trajectory for me as an African researcher: Africa’s total absence from the architecture being built. It was George Santayana who said:
“Those who cannot remember the past are condemned to repeat it.”
Africa was excluded as a producer during the industrial revolution, locked out during the nuclear era, and today remains almost entirely absent from the physical and algorithmic infrastructure of frontier AI. We do not own sovereign compute clusters, we do not control frontier model weights, and our local data is routinely scraped to train foreign systems that are then sold back to us as subscription products.
This ongoing exclusion is not just the result of foreign monopolies; it is also, and unfortunately, a direct consequence of internal failure. Brilliant African researchers, mathematicians, and engineers are actively contributing to major breakthroughs inside top global labs in the West, but they cannot do that work at home because corrupt, short-sighted governments consistently fail to invest in foundational research, scientific infrastructure, or compute capacity, treating technology as something to passively consume rather than build. Can you believe there is no frontier AI lab in the entire continent? Just the same way no African country owns an atomic bomb.
If African nations continue to sit back and watch while foreign powers dictate the rules of AI access, the consequences will be foundational across three critical fronts. First, we face total sovereignty failure: relying entirely on foreign APIs means a country’s educational, economic, administrative, and medical AI systems can be disabled by a policy change or export restriction. Second, we face epistemic erasure: models calibrated on Western data will continue to hallucinate or underperform on African languages, local law, and historical contexts, enforcing a subtle form of cultural standardization. Third, we face economic subjugation: value creation migrates entirely to compute owners, leaving Africa as a provider of cheap data annotation and raw consumer traffic.
Looking across Africa and developing nations today, it is easy to see history repeating itself in real time. The viral panic and doom-mongering on our social media feeds are a deliberate setup to manufacture consent for restrictions, keeping cutting-edge intelligence behind closed doors. For countries that lack the political will or domestic infrastructure to build these frontier systems from scratch, waiting for foreign monopolies to hand down permission is a trap. The only viable path forward, the only way to prevent a permanent digital caste system, is to advocate relentlessly for open-source AI. Making model weights accessible to everyone is the single most urgent battle to ensure true safety for all and that no country or region is left behind once again.
The Non-Proliferation Trap: How “AI safety” is Used to Lock Out the (third) World
If you opened Twitter or LinkedIn over the past 48 hours, you couldn’t miss it. Your feed was almost certainly flooded with viral threads and resignation letters from frontier AI lab researchers claiming we’re racing toward self-improving superintelligence that will end the world. People on the inside are publicly announcing they’re stepping away because the risk of AI killing all humans within a few years is “too high,” and alignment leads are chiming in to confirm they believe there’s a real chance this technology wipes us out. It’s apocalyptic. It’s all over my “for you” page, and probably yours too.
AI agency and autonomous capabilities do carry real risks. But past the volume of these apocalyptic declarations, there’s a straightforward question worth asking: who actually benefits from this panic? To me, the sudden narrative that frontier AI is an existential threat dangerous enough to end humanity looks less like organic fear and more like the setup for a specific political outcome, one we’ve seen play out before.
A little history lesson. When the US developed the atomic bomb through the Manhattan Project, the world’s most powerful nations immediately recognized they possessed a technology capable of absolute dominance. So they built a narrative around existential dread, arguing that letting nuclear capability spread would mean global annihilation, to make sure nobody else could match that power.
That framing led directly to the Non-Proliferation Treaty, sold as a noble effort to save humanity from self-destruction. What it actually locked in was a hypocritical cartel: a select group of nations kept their nuclear weapons while everyone else was legally forbidden from ever building their own. Try to build one outside that circle, and you’d face sanctions or military intervention, dressed up as global safety.
I think that exact non-proliferation logic is now being weaponized to build a regulatory moat around frontier AI. The existential threat, the sensational tweets, the TV appearances: all of it gets used to justify strict government licensing, compute caps, and centralized oversight over who’s allowed to train or run high-level models. The goal is keeping AI capability locked inside a small group of approved players: the proprietary labs in the US (OpenAI, Anthropic, Google), the sovereign effort in China (DeepSeek), the regional attempt in Europe (Mistral).
If proprietary models become the primary interface through which humanity accesses knowledge and culture, concentrating that much power in three or four corporations is a dangerous failure mode. Yann LeCun has made this point repeatedly: that the real risk isn’t rogue AI but concentrated control by a handful of proprietary providers, and that open-source foundation models are the only real path to AI sovereignty. Closed weights turn safety into an opaque promise from an evaluator nobody outside the company can inspect. Open weights let communities verify, recalibrate, and run infrastructure locally, without worrying that access gets revoked by someone else’s foreign policy.
For non-Western nations, closed-source “aligned” models are a new kind of digital paternalism. An evaluator trained on Silicon Valley or Beijing values treats local culture, language, and political reality as noise, or worse, as an alignment failure to be corrected. None of this restriction makes the world safer. What it guarantees is that a handful of foreign corporations and governments keep the monopoly on global digital intelligence.
This is the part of the trajectory that frustrates me most as an African researcher: Africa’s total absence from the architecture being built. Santayana’s line about those who forget the past being condemned to repeat it feels almost too on the nose here. We were excluded as producers during the industrial revolution, locked out during the nuclear era, and today we’re almost entirely absent from the physical and algorithmic infrastructure of frontier AI: no sovereign compute clusters, no control over frontier model weights, just local data scraped to train foreign systems that get sold back to us as subscription products.
And it’s not only foreign monopolies doing this to us. Brilliant African researchers, mathematicians, and engineers are contributing to major breakthroughs inside top labs in the West right now, but they can’t do that work at home, because corrupt, short-sighted governments keep failing to invest in foundational research, scientific infrastructure, or compute capacity. Technology gets treated as something to consume, not build. There is no frontier AI lab anywhere on the continent, the same way no African country owns an atomic bomb.
If African nations keep sitting back while foreign powers write the rules of AI access, the fallout hits on three fronts. Sovereignty fails first: relying entirely on foreign APIs means a country’s educational, economic, administrative, and medical AI systems can be disabled by someone else’s policy change or export restriction. Then comes epistemic erasure, models calibrated on Western data keep hallucinating or underperforming on African languages, local law, and historical context, enforcing a quiet form of cultural standardization. And economic subjugation follows: value creation migrates entirely to whoever owns the compute, leaving Africa as a source of cheap data annotation and raw consumer traffic.
Looking across Africa and the rest of the developing world today, it’s easy to see history repeating itself in real time. The viral panic and doom-mongering flooding our feeds looks like a deliberate setup, manufacturing consent for restrictions that keep cutting-edge intelligence locked behind closed doors. For countries without the political will or domestic infrastructure to build frontier systems from scratch, waiting for foreign monopolies to hand down permission is a trap. Advocating relentlessly for open-source AI is the only way I see to avoid a permanent digital caste system. Making model weights accessible to everyone isn’t a side issue, it’s the single most urgent fight if we actually want real safety, and a world where no country gets left behind again.