The Red Lines Are DrawnIn the autumn of 2025, a negotiation quietly reached an impasse at Anthropic's headquarters in San Francisco. The other party was the U.S. Department of Defense. The agenda: how far could the company's AI model, Claude, be deployed for military and national security purposes? And on two specific questions, Dario Amodei refused to say yes — no matter the price offered.The first was the use of Claude for large-scale surveillance, including the interception of personal data worldwide. The second was its use in fully autonomous lethal weapons systems — systems in which no human being participates. For Amodei, these two lines were non-negotiable. Each carried a deep, specific fear.Fear of a World Where AI Alone Can KillAt the core of Amodei's fear of autonomous lethal weapons lies a fundamental rejection of removing human beings from the decision to kill. He has stated: "At the moment the trigger is pulled, there must be human eyes, judgment, and conscience. Without that, war is no longer a human act."Imagine concretely what would happen if today's AI were deployed on a battlefield. AI can analyze thousands of targets per second, quantify threat levels, and autonomously determine attack priorities. While a human soldier spends one second deliberating "should I shoot?", AI can render 100 attack judgments. But when errors of misidentification are mixed into those "judgments" — who bears the responsibility?AI acts on training data. If that data carries bias, people of certain races, clothing styles, or behavioral patterns become more likely to be misidentified as "threats." In past deployments of image-recognition AI by the U.S. military, multiple cases were reported in which farmers carrying agricultural tools were classified as armed combatants. A human soldier might sense in the atmosphere, the person's expression, or the context that "this is wrong." AI has no such intuition. If a numerical value crosses a threshold, that becomes a death sentence.What Amodei particularly fears is the very process by which autonomous AI killing decisions become "normalized." Initially they are introduced under the pretense that "humans make the final approval." But in real combat, there is virtually no time for a human to scrutinize a conclusion AI reaches in one second. Gradually, a hollowed-out process — in which humans merely agree with what AI has already decided — becomes the standard. Ultimately the situation arrives at: "AI decides; humans only push the button." That is an automated killing machine wearing the label of "human oversight."Amodei also fears the proliferation of this technology. If the United States deploys autonomous AI weapons in combat, China, Russia, and even terrorist organizations and non-state actors will race to develop the same capabilities. Unlike nuclear weapons, AI has dramatically lower development costs and requires no physical manufacturing infrastructure. An era in which a small drone can be equipped with a lethal-judgment AI using computing power roughly equivalent to a smartphone has already technically arrived. If a "AI weapons arms race" among great powers begins, Amodei argues, its endpoint will be beyond anyone's control."If we don't draw the line here and now, when we try to draw it ten years from now, no one will follow it anymore. The first line determines everything," he is reported to have told close aides.The Day the World's Personal Data Gets Absorbed by AIOn the other red line — large-scale interception and surveillance of personal information — Amodei's concerns are equally concrete and severe.The information-processing capability of modern AI far exceeds human imagination. The wiretapping and surveillance that intelligence agencies once conducted through mass human labor required narrowing the focus to "only high-value targets," due to physical limitations. AI erases those constraints. It is now technically possible to simultaneously process call records, emails, social media posts, location data, purchase histories, medical records, and facial recognition data for billions of people — and to instantly grasp each individual's behavioral patterns, ideological tendencies, social relationships, and financial status.The scenario Amodei fears most is this: when a highly advanced language AI like Claude is connected to a global communications infrastructure, it no longer becomes "a tool to monitor specific individuals" — it becomes "an infrastructure for the constant surveillance of all humanity." Who holds what political opinion. Who has made statements critical of the government. Who has been in contact with dissident figures. A system that automatically detects these things, flags them, and reports them to authorities becomes reality for the first time through AI.This is incomparable to China's "social credit system." China's system is largely limited to facial recognition in public spaces and monitoring of online speech. But a comprehensive interception system fusing AI with communications infrastructure — one that includes decryption of encrypted communications, network analysis of human relationships, and detection of contradictions with past statements — possesses surveillance capabilities of an entirely different magnitude.Amodei emphasizes this danger even within the United States. The Fourth Amendment guarantees freedom from unreasonable searches and seizures. But it is technically possible for AI to analyze the communications records of hundreds of millions of people "without a warrant" through telecommunications provider data. When the results of that analysis start to be used for administrative penalties or prosecutions against specific citizens — under the banner of "counterterrorism" or "national security" — that is no longer investigation. It is a perpetual interrogation of the entire citizenry.And the greatest problem with AI-based surveillance is that it is "invisible." A physical wiretap can be discovered. But there is no way for an ordinary citizen to know which communications an AI operating on Claude is analyzing, or how. A vague sense of "I might be under surveillance" causes people to self-censor their speech. When everyone begins filtering their words — wondering "is what I just said okay?" — that, Amodei says, is the beginning of the end of a free society.What the Department of Defense sought was precisely to have Claude take on this capability — to place Claude at the core of a global communications analysis system under the pretext of monitoring foreign intelligence activities. But Amodei knew well the historical pattern by which "today the target is foreigners" becomes "tomorrow the target is domestic dissidents." It is still fresh in memory that the NSA, once claiming to intercept "only the communications of terrorists," was revealed through Edward Snowden's disclosures to have actually collected the communications records of hundreds of millions of American citizens."The AI we built could become the most powerful surveillance tool in history. That is precisely why we must refuse to hand over the keys. That refusal is the reason we built this company." That was Amodei's answer.Secretary of Defense Pete Hegseth, who took office in January 2025, issued a memo demanding that military AI use be liberated from the constraint of corporate terms of service for "any lawful purpose," applying intense pressure. He loudly condemned the notion that "private companies should be allowed to dictate military operational judgment."Notably, regarding behind-the-scenes negotiations: the Department of Defense is reported to have made an informal proposal to Anthropic, offering to "guarantee government contracts worth $1 billion annually" in exchange for removing the red lines. But Amodei reportedly rejected this instantly, telling senior executives that "such a deal would be selling the company's soul." This anecdote illustrates that the decision was rooted in genuine conviction.The Accelerating ClockTo understand the backdrop to this confrontation, let us set the clock briefly to early 2026. Anthropic CEO Dario Amodei published a 38-page essay declaring that "superintelligent AI could arrive by 2027." Titled "The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI," the piece revealed — from a man who had once championed an optimistic vision of AI curing diseases and extending lifespans — a powerful sense of crisis about "real-world dangers" on a civilizational scale.Amodei warned that humanity is currently caught in the midst of a "rite of passage," facing the greatest national security threat in history. Now that AI has reached "adolescence," the question is whether humanity possesses the maturity to wield such power. And the answer to that question began to play out in reality as a full-scale confrontation between one company committed to ethics and safety, and a Trump administration that treats military supremacy as absolute.The Trump Administration's "Death Sentence"On February 27, 2026, the situation reached a decisive turning point. President Trump posted on his social media platform Truth Social, denouncing Anthropic as a "radical left-wing company" and ordering federal agencies to immediately cease use of the company's technology. In response, Secretary of Defense Hegseth announced that Anthropic would be designated a "national security supply chain risk" — and ordered all contractors doing business with the Department of Defense to prohibit any commercial dealings with the company.This designation was unprecedented. The "supply chain risk" label is normally applied only to companies associated with foreign adversaries, such as Huawei. There was no prior precedent for applying it to a domestic American private company arising from a contractual dispute. Even conservative legal scholars raised a chorus of voices calling it "an abuse of the law."This was effectively a "commercial death sentence" for Anthropic. Claude runs on infrastructure provided by AWS and Google Cloud — both of which are major contractors with the Department of Defense. If they complied with the order and cut off Anthropic, the company would be pushed to the brink of extinction.There is an important behind-the-scenes detail: three days before the announcement of the designation, the Department of Defense's legal counsel team reportedly submitted a memo to Hegseth refusing to sign off, citing "excessively fragile legal grounds." But Hegseth overruled them, insisting that "the political message matters more than the legal details." Objections also arose within the White House, but the matter was reportedly settled by a single word from Trump himself: "Do it."The Rise of a Rival and the Transformation of GovernanceWithin hours of Anthropic's exclusion, its former colleague and greatest rival, OpenAI, announced a contract with the Department of Defense for deploying AI models on classified networks. The timing was too convenient to be coincidental, and many in the industry interpreted it as "politically calculated."OpenAI, led by CEO Sam Altman, reached an agreement that effectively accepted the DoD's standard of "any lawful purpose." The red lines Anthropic had fought to defend — against autonomous lethal weapons — were not explicitly stated. As a result, the leadership of AI governance shifted structurally from "corporate safety guidelines" to "state-led oversight."One senior researcher reportedly said afterward: "The day after the agreement, half my team had gone silent in the internal chat. That was the answer." Altman himself is said to have confided to close aides: "If we refused, we'd end up like Anthropic. That helps no one" — suggesting the decision was made out of fear.Ultimately Altman attempted damage control by adding a provision prohibiting "intentional domestic surveillance of American citizens" — but whether safety restrictions at the level Anthropic demanded have been maintained remains unclear.The People's Revolt and the "Constitutional Crisis"The government's heavy-handed intervention and OpenAI's capitulation triggered an unexpected market reaction. Claude — the very AI that had been shut out — ironically gained explosive popular support among ordinary consumers.In early March 2026, Claude shot to number one on Apple's App Store free chart. Prominent figures publicly announced switching from OpenAI to Anthropic, as users acted in support of "corporate conscience." Meanwhile, ChatGPT saw a sharp spike in daily uninstalls.Simultaneously, liberal think tanks and civic organizations framed the issue as a "constitutional crisis." Groups such as the Center for American Progress condemned the administration's attack on Anthropic as "an unprecedented assault on a private enterprise" — an abuse of power exceeding legal authority. They argued that large-scale AI surveillance and the automation of lethal decisions are "constitutional questions that Congress must legislate," and strongly demanded the convening of public hearings.Notably, the reaction was bipartisan. Criticism arose even from within the Republican Party framing it as "government intervention in free markets." A conservative Texas senator issued a statement declaring that "government intervention with legally binding force in the terms of service of a private company runs contrary to the principles of a free-market economy" — dealing the administration no small blow.The Warfighting Secretary and the Shadow of 2027Behind this confrontation lies the ideology of a singular politician: Hegseth. Sometimes called a "Christian nationalist," he has pushed to purge diversity from the military and pursue military strength alone. The Department of Defense has begun using the old name "Department of War" alongside its official title, and Hegseth himself reportedly prefers the title "Secretary of War."One of his aides told reporters: "Every time the Secretary hears the phrase 'AI ethics,' he thinks of Silicon Valley elitism. For him, this isn't a conversation about technology — it's a culture war." This statement reveals that the conflict reflects a deep ideological fault line transcending the practical question of military AI use.For Hegseth, Anthropic's safety restrictions were nothing more than "left-wing shackles weakening the military's capabilities." AI possessing the intellect of "50 million Nobel laureates" is the ultimate weapon — and to restrict it in the name of ethics was, in his view, an intolerable act of betrayal.Time is running short before the "2027" Amodei prophesied. When the autonomous development cycle closes — in which AI designs the next generation of AI on its own — who will control that technology? Will it be ethics that respects individual rights? Or will it be the will of the state, flying the banner of "any lawful purpose"? The answer is being shaped right now, in the midst of a real political struggle.An Unfinished StruggleThe full-scale confrontation between Anthropic and the Trump administration has transcended the question of a single startup's survival — it poses the essential question of how humanity should govern a "godlike power" called AI.Anthropic has signaled its intent to contest the "supply chain risk" designation in court by seeking an injunction, and this lawsuit will serve as a constitutional test case examining the scope of presidential authority under the Defense Production Act. In Congress, a bipartisan group of lawmakers has begun working toward an "AI Governance Framework Act," and in a rich irony, this confrontation is paradoxically accelerating the legislative process on AI regulation that had long been stalled.Amodei once dreamed of a future in which AI, as a benevolent machine, would rescue humanity. But today, amid the turmoil and violence characteristic of adolescence, he stands on the front lines — working to ensure that the intelligence he helped create is not consumed by the enormous force called the state. "What we are as a species is being tested," he has said. Those words are now unfolding as a real political struggle, before the eyes of the entire world.ReferencesDario Amodei, "The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI," Anthropic Blog, January 2026.Pete Hegseth, "Memorandum on AI Utilization Standards for DoD Contractors," U.S. Department of Defense, October 2025.Center for American Progress, "Unprecedented Attack on Private Enterprise: The Anthropic Case and Constitutional Implications," March 2026.The New York Times, "Trump Administration Designates Anthropic as Supply Chain Risk," February 28, 2026.The Washington Post, "OpenAI Strikes Deal with Pentagon as Anthropic Faces Government Exclusion," February 27, 2026.Wired, "The Battle for AI's Soul: Anthropic vs. the Trump White House," March 2026.Bloomberg Technology, "Inside the Negotiations That Led to Anthropic's Government Blacklisting," March 2026.MIT Technology Review, "What the Anthropic-Pentagon Standoff Means for AI Governance," March 2026.Financial Times, "Silicon Valley's Culture War Goes to Washington," March 2026.Reuters, "Claude AI App Surges to No.1 on App Store Amid Political Controversy," March 2026.The Atlantic, "Pete Hegseth's War on Tech Ethics," February 2026.Foreign Affairs, "AI Arms Race: How the Anthropic Affair Accelerated Global Military AI Development," March 2026.