What Really Needs a Pause: Fear-Driven AI Narratives and Nationality-Based Exclusion
There is a serious conversation to be had about powerful AI systems. Frontier models are dual-use technologies. They can help attackers, lower barriers to misuse, and create governance challenges that democratic societies have not yet solved. Anyone who dismisses these risks is not taking the technology seriously.
But the reverse is also true. Anyone who discusses powerful AI only through the lens of danger is also not taking the technology seriously.
This is why the recent PauseAI article on the U.S. government’s restriction of Anthropic’s Fable 5 and Mythos 5 models is troubling. The core event it discusses is real. Anthropic itself stated that the U.S. government issued an export-control directive suspending access to these models for foreign nationals. But the framing of the article turns a complex governance problem into a simple morality play: powerful AI is dangerous, government restriction is therefore a safety victory, and the lesson is that we need more pausing.
That framing misses several essential points.
The first and most disturbing omission is the discriminatory structure of the action itself. According to Anthropic, the directive applied to “any foreign national, whether inside or outside the United States,” including foreign-national Anthropic employees. That is not a small implementation detail. It means that access is restricted not primarily by conduct, role, trustworthiness, auditability, institutional affiliation, or demonstrated misuse, but by nationality.
This matters profoundly. Science is international. The modern AI field was not built by one nation, and it certainly was not built only by U.S. citizens. Many of the people who contributed centrally to machine learning, deep learning, computer vision, robotics, medical AI, natural language processing, and AI safety have worked across borders, universities, laboratories, companies, and citizenship categories. A person can be a trusted employee, a leading researcher, a contributor to public knowledge, and still be excluded by such a rule because of passport status.
Andrej Karpathy is a useful example of the absurdity of this kind of framing. He is one of the best-known figures in modern AI, associated with major contributions in deep learning education, computer vision, Tesla Autopilot, OpenAI, and more. He is widely described as Slovak-Canadian. Whether or not a specific policy would affect him personally in a given institutional setting is not the point. The point is that the global AI community is full of people like this: researchers whose work shapes the field, but whose nationality does not fit a simple domestic-versus-foreign security story.
If access to scientific tools is restricted on the basis of nationality, then a safety policy can itself become unjust. A democratic society should be extremely careful before normalizing categories that treat “foreign national” as a proxy for risk. There may be cases where export controls are justified, but such measures require a high burden of evidence, transparency, proportionality, and due process. Otherwise, they risk undermining the very values they claim to protect.
The second omission is the positive side of AI. The PauseAI article largely ignores what powerful AI can achieve and what it has already achieved. That absence is not neutral. It shapes the reader’s emotional response.
AI is not only a possible amplifier of cyber offense. It can also be a tool for cyber defense. Anthropic itself has described work in which Claude Mythos was used to identify high- and critical-severity vulnerabilities in important software. If a model can help attackers discover vulnerabilities, then the same underlying capability can help defenders find and fix them first. Removing access from responsible actors may not make society safer if irresponsible actors can still use alternative systems, open models, leaked models, domestic models, or future replications.
The same dual-use logic applies in biology and medicine. AI systems have already transformed protein structure prediction. AlphaFold and related systems have accelerated structural biology and biomedical research by making molecular structures more accessible to scientists. AI-supported medical imaging is also moving from theory into practice. In breast cancer screening, studies have shown that AI can reduce radiologists’ workload while maintaining or improving detection performance. These are not speculative utopian promises. They are concrete examples of AI already producing public benefit.
This is the context missing from the article. A serious argument about pausing or restricting AI must consider not only the danger of access, but also the danger of lost access. What happens when hospitals, universities, small companies, public-interest researchers, cybersecurity defenders, and non-U.S. scientists are blocked from the best tools? What happens when defensive work is slowed while attackers adapt? What happens when a safety policy strengthens centralization, geopolitical asymmetry, or national discrimination?
The answer is not obvious. That is precisely why the debate must be careful.
The most worrying part of the PauseAI article is not that it takes a strong position. Advocacy is legitimate. Civil society organizations have an important role in technological governance. But advocacy becomes dangerous when it presents a selective narrative as if it were a complete analysis.
The article emphasizes fear, urgency, and catastrophic possibility. It underemphasizes uncertainty, competing risks, positive use cases, proportionality, and the rights of those excluded by the policy. It treats the restriction almost as a confirmation of the PauseAI worldview, rather than as an occasion to examine a difficult tradeoff. This is a common failure mode in advocacy: first decide the narrative, then select the facts that make the narrative emotionally compelling.
That method should worry us. It is not enough for an organization to have good intentions. A movement that claims to defend humanity from dangerous systems should also defend the standards of public reason. It should show its readers the full problem, not only the part that activates fear. It should acknowledge benefits as well as harms. It should separate evidence from speculation. It should avoid turning complex governance questions into simple stories of heroes, villains, and inevitable catastrophe.Otherwise, advocacy risks becoming part of the misinformation environment it claims to oppose. Not misinformation in the crude sense of inventing facts, but in the subtler and more influential sense of arranging true facts into a one-sided emotional architecture. This is how public debate becomes less enlightened. The reader is not invited to reason through a difficult problem; the reader is guided toward a preselected conclusion.
That is what needs a pause.
We do not need a pause on democratic debate about AI. We need more of it. We do not need a pause on AI safety. We need serious safety work. We do not need a pause on governance. We need better governance urgently.
But we do need a pause on fear-driven narratives that exclude inconvenient facts. We need a pause on treating foreign nationals as a default risk category. We need a pause on advocacy that speaks the language of public interest while appealing primarily to anxiety. We need a pause on arguments that demand enlightenment values from technologists while failing to practice those values in their own rhetoric.
A better path is possible. Powerful AI systems should be evaluated rigorously. Dangerous capabilities should be tested. High-risk deployments should be audited. Access to the most capable systems may need tiered controls, monitoring, logging, institutional accountability, and export rules in narrowly justified cases. But such governance should be based on conduct, capability, safeguards, and accountability wherever possible, not blanket suspicion by nationality.
It should also preserve the positive uses of AI. We should want the best possible tools in the hands of responsible medical researchers, scientific institutions, educators, accessibility experts, cybersecurity defenders, and public-interest organizations. We should want international collaboration, not technological nationalism disguised as safety. We should want evidence-based governance, not panic-driven restriction.
The real choice is not between reckless acceleration and a total pause. That is a false binary. The real challenge is to govern powerful AI without abandoning fairness, openness, scientific collaboration, and intellectual honesty.
If PauseAI wants to contribute to that debate, it should do so by presenting the full picture. It should discuss risks, but also benefits. It should discuss catastrophic possibilities, but also opportunity costs. It should criticize unsafe deployment, but also criticize discriminatory access policies. It should demand responsibility from AI companies, but also practice responsibility in its own public communication.
A movement for safety should not win by narrowing the public imagination. It should win by expanding public understanding.
That is the standard we should expect from anyone asking society to pause.
PauseAI article being discussed:
References
Anthropic statement on the U.S. government directive restricting access to Fable 5 and Mythos 5: https://www.anthropic.com/news/fable-mythos-access
Anthropic announcement of Claude Fable 5 and Claude Mythos 5: https://www.anthropic.com/news/claude-fable-5-mythos-5
Anthropic page on Claude Mythos and Project Glasswing: https://www.anthropic.com/claude/mythos
PauseAI homepage and stated position: https://pauseai.info/
Wikipedia page for Andrej Karpathy, for basic biographical context: https://en.wikipedia.org/wiki/Andrej_Karpathy
AlphaFold 3 paper in Nature: https://www.nature.com/articles/s41586-024-07487-w
Nature article on the expansion and impact of AlphaFold resources: https://www.nature.com/articles/d41586-026-00787-3
MASAI randomized trial on AI-supported mammography screening in The Lancet Oncology: https://www.thelancet.com/journals/lanonc/article/PIIS1470-2045(23)00298-X
Nature Medicine study on AI-supported mammography screening in Germany: https://www.nature.com/articles/s41591-024-03408-6
FDA list of AI-enabled medical devices: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices
If you liked this blog post, I recommend having a look at our free deep learning resources or my YouTube Channel.
Text and images of this article are licensed under Creative Commons License 4.0 Attribution. Feel free to reuse and share any part of this work. AI was used to support the creation of this article.