On Friday at 5:21 p.m. ET, Anthropic says it received a U.S. government directive ordering the company to suspend access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States, including foreign-national Anthropic employees. Anthropic says the practical effect was that it had to disable both models for all customers to ensure compliance. The company also says the directive did not provide specific details about the national security concern, only that the government believed it had become aware of a method for bypassing, or jailbreaking, Fable 5. Anthropic reviewed the demonstration and characterized the issue as narrow, non-universal, and capable of producing results already available from other public models.
The Era of Permissioned Intelligence
For the past three years, most of the public conversation about AI has focused on capability: how smart the models are, how fast they are improving, how much they cost, how many tokens they can process, which benchmarks they can beat, and whether the next model will meaningfully outperform the last one. That conversation made sense while AI was still understood as software. Software has features. Software has release notes. Software has bugs. Software has users. We know how to think about software, even when it surprises us.
Intelligence Decoupled From Consciousness
But frontier AI is not software in any ordinary sense. It is intelligence decoupled from consciousness. The models do not know they exist (at least we don’t think they do). They do not have a childhood, a spouse, a favorite song, an honestly earned sense of humor, or any particular desire to impress your board of directors. Yet they can reason, explain, synthesize, classify, challenge assumptions, write code, analyze law, summarize research, design experiments, tutor a child, help a developer, coach an executive, and produce useful work across many domains orders of magnitude faster than humans can.
Whether you want to call that “intelligence” is a philosophical question. Whether it performs intellectual labor has been asked and answered.
For all of human history, intelligence came bundled with human beings. If you wanted access to a physician’s diagnostic intuition, a mathematician’s abstraction, a scientist’s method, a teacher’s patience, a lawyer’s reading of a contract, or an engineer’s ability to decompose a problem, you needed access to that person or to the institutions that trained and employed them.
The internet changed access to information, not intelligence. Search made facts discoverable; it did not make judgment abundant. Raw information is not knowledge, nor is it wisdom.
AI has fundamentally changed this. Now that intelligence has become a product, someone will decide who gets the premium version, who gets the safe version, who gets the delayed version, who gets the monitored version, who gets the export-controlled version, and who gets nothing.
The New Gatekeepers
According to Anthropic, the government acted under national security authority and required the suspension of access to two frontier models. Anthropic complied, while publicly disagreeing with the basis for the action and arguing that, if the same standard were applied across the industry, it would “essentially halt all new model deployments for all frontier model providers.” Anthropic also says it supports the government’s ability to block unsafe deployments, but only through a statutory process that is transparent, fair, clear, and grounded in technical facts.
This is not a simple “government bad, companies good” story. It is not even a simple censorship story. The government has legitimate national security concerns. A sufficiently capable model can help competent bad actors move faster. Cybersecurity, biology, chemistry, autonomous agents, and model distillation are not abstract policy categories; they are real domains where better reasoning can produce both extraordinary benefits and catastrophic misuse.
To be clear, Anthropic already makes access decisions inside its own products. The company calls them safeguards. Other companies use different terms: safety policies, acceptable use, alignment, red teaming, mitigations, routing, refusals, trust tiers, enterprise controls. These policies may be well-intentioned. Some are absolutely necessary. But the structure is still the structure. A small number of people define the categories of dangerous knowledge, build classifiers to detect those categories, decide which model you are allowed to use, and determine whether you receive the best answer, a weaker answer, a deflection, or no answer at all.
The government calls it national security. The companies call it safety. Researchers call it risk mitigation. Regulators call it the public interest. The labels change; the mechanism does not. Someone decides what kind of intelligence you are allowed to access.
Orwell’s Mechanism
George Orwell’s novel 1984 was a warning about the institutional control of reality. The Thought Police did not merely punish forbidden speech; they shaped the conditions under which thought itself could form. In Orwell’s dystopian world, that required surveillance, propaganda, coerced language, memory holes, and brute force. In our world, it may require a policy document, an export-control directive, a model router, and a classifier quietly deciding that your question belongs in the wrong bucket.
I don’t want to overstate the analogy. This moment is far more interesting than the cartoon version of the argument. The people making these decisions are not villains. They may be thoughtful, careful, technically sophisticated, patriotic, and genuinely worried about misuse. They may also be shareholders, political appointees, regulators, executives, safety researchers, and institutional actors with incentives, biases, blind spots, career risk, and power.
When startups pitch me would-be world-changing tech, I always ask: “You say you’ve built the Death Star, are you planning to rent it out or go f#ck up some planets?” It’s a joke. Mostly. What I’m really asking is whether the team understands the difference between building a powerful tool and governing the consequences of that tool’s existence. The question gets harder when the Death Star is not a fictional superweapon but a frontier intelligence system built by a private company, trained with private capital, deployed through private terms of service, and then constrained by public power.
It is one thing for a company to decide whether a product feature ships on Tuesday. It is another thing for a company to decide which humans get access to the most capable reasoning systems ever created. It is also uncomfortable when governments assert that authority, especially through opaque processes, emergency letters, verbal evidence, or standards that cannot be applied evenly across the industry. The unsettling part is that both sides have plausible claims. Companies built the systems. Governments have public safety obligations. Users have interests. Researchers have expertise. Customers have dependencies. Citizens have rights. Bad actors have incentives. Everyone has a point, and none of those points answers the central question: Who gets to decide?
Access Is Power
If access to advanced AI determines who can learn faster, code faster, research faster, negotiate better, diagnose more accurately, design more efficiently, and make better decisions, then whoever controls access, controls the future.
We have seen versions of this before. Literacy created classes. Libraries created access. Universities created credentials. Printing presses created religious and political crises. Broadcast networks created distribution power. Search engines created visibility power. Social platforms created attention power. Each new knowledge system produced gatekeepers. The difference now is that the gate is not standing in front of information. It is standing in front of intelligence.
This is why the Anthropic episode should not be dismissed as an edge case in export control or a temporary dispute over jailbreaks. It is a preview of the governance layer that will sit above all frontier AI systems. That layer will decide who gets access, under what conditions, with what monitoring, with what audit trail, in what jurisdiction, for which use cases, at what capability level, and with what right of appeal. The governance layer may be corporate. It may be governmental. It will probably be both. It will almost certainly be messy, political, inconsistent, and rationalized as necessary.
Maybe it is necessary. I do not know how to responsibly provide unrestricted access to systems that will soon exceed expert-level performance in domains where misuse has real-world consequences. I also do not know how to preserve democratic access to intelligence if the most capable systems can be gated by a handful of corporations and government officials.
I don’t have a recommendation. I am writing in a moment when I cannot see the answer or even pretend to. What I can see is the shape of the question. Who gets to decide what the rest of us are allowed to know?
The Thought Police do not need to read your mind if they can decide which minds you are allowed to consult. This scares the hell out of me. It should scare the hell out of you too.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.