Neil deGrasse Tyson doesn’t scare easily. The astrophysicist has spent decades explaining black holes, asteroid impacts, and the heat death of the universe with the calm demeanor of a man who’s made peace with cosmic indifference. But artificial superintelligence? That rattles him.
“That branch of AI is lethal,” Tyson said in a recent appearance on the ALL-IN podcast. “We’ve got to do something about that.”
The comments, first reported by TechRadar, mark one of the most direct public calls from a prominent scientist to draw a hard regulatory line around superintelligent AI — not just to regulate it, but to ban it outright. Tyson distinguished carefully between the kinds of artificial intelligence already embedded in commerce, medicine, and daily life and the theoretical arrival of a system that could outthink humanity across every domain. The former, he suggested, is manageable. The latter is an existential problem.
It’s a distinction that matters enormously right now. And it’s one that Silicon Valley’s biggest players have been conspicuously reluctant to make.
Tyson’s argument isn’t that current AI tools — large language models, image generators, autonomous coding assistants — are themselves the threat. He’s talking about what comes next. Artificial general intelligence, or AGI, refers to a hypothetical system with human-level reasoning across all cognitive tasks. Superintelligence goes further: a machine intellect that surpasses human capability not in narrow domains like chess or protein folding, but in everything. Strategy. Persuasion. Scientific discovery. Self-improvement.
The concept isn’t new. Oxford philosopher Nick Bostrom laid out the risks in his 2014 book Superintelligence: Paths, Dangers, Strategies, arguing that a sufficiently advanced AI system might pursue goals misaligned with human survival — not out of malice, but out of optimization logic that treats human values as irrelevant variables. Bostrom’s work was once considered speculative philosophy. Increasingly, it reads like a risk assessment.
What’s changed is the pace. OpenAI, the company that introduced ChatGPT barely two and a half years ago, has stated explicitly that its mission is to build AGI. CEO Sam Altman has described AGI as potentially “the most transformative and potentially dangerous technology in human history” while simultaneously racing to build it. Google DeepMind CEO Demis Hassabis has offered similar framings — acknowledging the risks in one breath, accelerating development in the next. Meta’s Mark Zuckerberg committed billions to open-source AI development in 2024, arguing that broad access is the safest path forward.
Tyson isn’t buying the argument that the people building the technology are best positioned to govern it. On the ALL-IN podcast, he drew an analogy to nuclear weapons — a technology whose destructive potential was recognized early enough that international treaties, however imperfect, established boundaries. “We didn’t say, let’s just let the physicists figure it out,” he noted. The Manhattan Project scientists themselves, many of them, pushed for arms control. Some refused to work on the hydrogen bomb.
The nuclear analogy has limits, of course. Nuclear weapons require enriched fissile material, massive industrial infrastructure, and state-level resources. AI requires GPUs, data, and clever engineering — resources that are expensive but increasingly distributed. You can’t put enrichment centrifuges in a garage. You can, at least theoretically, train a powerful model in one. That diffusion of capability is precisely what makes the governance problem so much harder.
But Tyson’s core point stands: the question of whether humanity should build something smarter than itself is not a technical question. It’s a civilizational one.
The timing of his remarks coincides with a surge of activity on the AI safety front. In May 2025, the U.S. Senate held hearings on AI risk frameworks, with testimony from researchers at Anthropic, the AI safety-focused company founded by former OpenAI executives Dario and Daniela Amodei. Anthropic has been vocal about what it calls “catastrophic” and “existential” risk categories, publishing detailed policy papers on how frontier AI systems should be evaluated before deployment. The European Union’s AI Act, which began phased enforcement in 2024, classifies certain AI applications by risk level — but doesn’t address superintelligence directly, largely because regulators aren’t sure how to define a technology that doesn’t yet exist.
That definitional problem is real. Critics of Tyson’s position — and there are many in the tech industry — argue that banning superintelligence is like banning time travel: you can’t prohibit something you can’t build. Venture capitalist and podcast host Chamath Palihapitiya, who was present during the ALL-IN conversation, pushed back on the feasibility of enforcement. How do you ban a capability that emerges gradually from systems designed for other purposes? Where exactly is the line between a very powerful narrow AI and a general one?
These are fair questions. They’re also, some safety researchers argue, the wrong ones to ask first.
“The question isn’t whether we can define superintelligence precisely today,” said Yoshua Bengio, the Turing Award-winning AI researcher who has become one of the field’s most prominent voices on existential risk. Bengio, who leads the International AI Safety Report commissioned by the UK government, has argued that the inability to define a precise threshold doesn’t excuse inaction. We regulate bioweapons research without a bright-line definition of what constitutes a “dangerous” pathogen. We restrict gain-of-function research based on risk assessments, not taxonomic precision.
The counter-counterargument from the accelerationist camp is that restricting AI development in the United States or Europe simply cedes the advantage to China, which has invested heavily in frontier AI capabilities and shown little interest in binding international agreements on the technology. This is the geopolitical version of the prisoner’s dilemma, and it carries genuine weight. Beijing’s AI ambitions are well-documented. The Chinese government released its own AI governance framework in 2023, but enforcement mechanisms remain opaque, and the strategic incentive to develop military and intelligence applications of advanced AI is obvious.
Tyson acknowledged this tension but didn’t retreat from his position. Unilateral restraint is risky, he conceded. But unilateral development of a technology that could escape human control is riskier. “If you build something smarter than you,” he said, “you are no longer the smartest thing in the room. And historically, the smartest thing in the room makes the rules.”
That line got attention. Rightly so.
The debate over superintelligence has historically been dominated by two camps that talk past each other. On one side, the “AI doomers” — researchers like Eliezer Yudkowsky of the Machine Intelligence Research Institute, who has argued that the development of superintelligent AI will almost certainly lead to human extinction unless solved with mathematical rigor that doesn’t yet exist. Yudkowsky’s position is that alignment — the problem of ensuring an AI system’s goals match human values — is so technically difficult that building superintelligence before solving it is equivalent to suicide. He has called for international treaties that would authorize military strikes against rogue AI data centers. Not a mainstream position, but one that reflects the intensity of concern among a subset of researchers.
On the other side, the “effective accelerationists” or e/acc movement, which holds that technological progress is inherently good, that AI development should be unrestricted, and that safety concerns are overblown or motivated by incumbents seeking to lock in competitive advantages through regulation. Marc Andreessen, co-founder of the venture capital firm Andreessen Horowitz, published a widely read manifesto in 2023 titled “The Techno-Optimist Manifesto” that explicitly rejected what he called the “decel” agenda. Andreessen argued that slowing AI development would cost lives by delaying breakthroughs in medicine, energy, and material science.
Tyson doesn’t fit neatly into either camp, which may be why his comments resonated. He’s not calling for a halt to all AI development. He’s not predicting imminent doom. He’s making a more targeted argument: that there exists a specific class of AI capability — machine superintelligence — that should be treated the way we treat other technologies with species-level risk. Banned. Not regulated. Banned.
It’s a bold line to draw. And it raises immediate practical questions.
First, who enforces such a ban? There is no international body with jurisdiction over AI development. The United Nations has established an AI advisory body, but it has no enforcement power and no clear mandate to restrict research. The International Atomic Energy Agency provides a model for nuclear inspections, but AI development is far more diffuse and harder to monitor than uranium enrichment. You can detect a nuclear test from space. You can’t detect a training run from space — at least not yet.
Second, how do you distinguish between research that advances toward superintelligence and research that doesn’t? Modern AI development is iterative. Each generation of large language model is more capable than the last. GPT-4 is more capable than GPT-3. GPT-5, expected later this year, will presumably be more capable still. At what point does incremental improvement cross into dangerous territory? The question has no clean answer, and the absence of a clean answer is precisely what makes governance so difficult.
Third, there’s the economic dimension. AI is now the single largest driver of capital expenditure among the world’s biggest technology companies. Microsoft, Google, Amazon, and Meta are collectively spending over $200 billion annually on AI infrastructure — data centers, chips, energy systems. Nvidia’s market capitalization has surged past $3 trillion on the strength of AI chip demand. Banning the ultimate goal of that investment would be, to put it mildly, disruptive to global capital markets.
None of these objections are trivial. But none of them are arguments that superintelligence is safe, either. They’re arguments that banning it is hard. Those are different things.
Recent developments have only sharpened the urgency. In early 2025, OpenAI released internal research on what it calls “superalignment” — the technical challenge of ensuring that AI systems more capable than humans remain under human control. The team originally leading that effort, co-headed by Ilya Sutskever and Jan Leike, dissolved in 2024 amid reported internal disagreements about whether OpenAI was prioritizing safety sufficiently. Sutskever left to co-found Safe Superintelligence Inc., a startup focused exclusively on the alignment problem. Leike joined Anthropic. The departures were widely interpreted as a signal that even within the organizations building frontier AI, there is deep disagreement about whether the pace of development is outrunning the science of safety.
Google DeepMind, meanwhile, published a comprehensive framework in early 2025 for evaluating “frontier AI risks,” including scenarios involving autonomous AI agents that can plan, execute multi-step tasks, and interact with the real world through tool use. The framework acknowledges that current evaluation methods may be insufficient for detecting dangerous capabilities in advanced systems — a remarkable admission from one of the world’s leading AI labs.
Anthropic’s Dario Amodei has been perhaps the most nuanced voice among lab leaders, arguing in a widely circulated essay titled “Machines of Loving Grace” that advanced AI could produce extraordinary benefits — curing diseases, solving climate change, reducing poverty — but only if developed with sufficient caution. Amodei has endorsed the concept of “responsible scaling policies” that tie the deployment of more powerful systems to demonstrated safety benchmarks. It’s a middle path between acceleration and prohibition, but it depends on the assumption that safety benchmarks can keep pace with capability advances. That assumption is unproven.
So where does Tyson’s call leave us? Probably in the same place most serious policy debates about transformative technologies end up: somewhere between the aspiration and the implementation, with the gap filled by argument.
But the argument itself matters. Five years ago, discussing the prohibition of superintelligent AI in mainstream media would have been dismissed as science fiction anxiety. Today, a respected astrophysicist can say “that branch of AI is lethal” on one of the most listened-to technology podcasts in the world, and the response isn’t laughter. It’s debate.
That shift in the Overton window is significant. The people building the most powerful AI systems in history are themselves divided on whether the destination is safe. The researchers who understand the alignment problem most deeply are among the most alarmed. And the regulatory infrastructure that might govern these technologies barely exists.
Tyson isn’t offering a detailed policy blueprint. He’s offering a principle: some capabilities should not be built. It’s the same principle that led to the Biological Weapons Convention of 1972, the same principle behind the Outer Space Treaty’s prohibition on orbital nuclear weapons. The principle has never been perfectly enforced. It has, however, been better than nothing.
Whether the world can muster the political will to apply that principle to artificial superintelligence — before the technology arrives, not after — is the question that will define the next decade of AI governance. And possibly quite a bit more than that.
The smartest thing in the room, for now, is still us. Tyson’s point is simple: maybe we should keep it that way.
WebProNews is an iEntry Publication