The Customer Is Always Right: How Trump Engineered a Decision Architecture Where Truth No Longer Fits

In May, Gabbard fired the acting chair of the National Intelligence Council and his deputy. Their offense? They refused to rewrite an assessment claiming that the Venezuelan criminal group Tren de Aragua was controlled by the Maduro government. That claim was the legal scaffolding for Trump's mass deportations of Venezuelans. The analysts declined to fabricate evidence that didn't exist. They were fired.

This was the defining moment. Everyone still standing understood: the new administration had no interest in analysis that contradicted its pre-formed conclusions, and was willing to discard anyone who delivered it. With the human expert layer purged, the next phase began - building an analytical layer at the machine level.

By spring 2025, the administration was constructing a second loop in its decision-making system: artificial intelligence deployed inside the Pentagon's classified networks. Palantir had already been working with Anthropic since November 2024, deploying Claude on its AIP platform with Defense Information Systems Agency accreditation at the SECRET level. In June 2025, Anthropic released Claude Gov, custom models built for classified environments and, in the company's own words, deployed "by agencies at the highest level of U.S. national security." On July 14, 2025, the Pentagon's Chief Digital and AI Office signed four contracts worth $200 million each with Anthropic, OpenAI, Google, and xAI. Frontier AI had officially entered the war room.

But before any of that, Trump ran his first big test. The Iran test.

Iran I: The Trump Who Could Still Stop

On June 13, 2025, Israel launched surprise strikes against Iran. Senior IRGC commanders, Chief of Staff Mohammad Bagheri, and several nuclear scientists were killed in the opening hours. The Twelve-Day War had begun.

On June 22, the United States joined with a single strike: Operation Midnight Hammer. Seven B-2 Spirit bombers dropped fourteen GBU-57 bunker busters on Fordow, Natanz, and Isfahan. A nuclear submarine followed up with more than twenty Tomahawks against Isfahan. Twenty-five minutes of active combat. At 10 PM Washington time, Trump addressed the nation and declared that Iran's nuclear facilities had been "completely and totally obliterated." Two days later, Iran launched missiles at a U.S. base in Qatar, deliberately avoiding casualties - a clear signal of willingness to de-escalate. On June 24, a ceasefire was announced.

A Defense Intelligence Agency assessment, leaked to the Washington Post, painted a rather different picture: the Iranian program had been set back by "months, not years" - dramatically more modest than Trump's obliteration rhetoric.

But here's what matters: Trump stopped. No ground operation. No regime-change attempt. No escalation. He stopped.

He stopped because there was still friction in his decision loop. Military advisers were counseling restraint, analysts were flagging risks, and - most importantly - there was no machine layer yet ready to whisper "go ahead, you'll win." Claude was not yet in classified networks at the scale it would reach one month later. Midnight Hammer was the work of an old-school Trump - the Trump who could still be persuaded that "enough is enough."

Venezuela: The Perfect Operation, Scripted by AI

By late summer and fall of 2025, the U.S. was steadily tightening the screws on Venezuela. In August, Trump doubled the bounty on Maduro to $50 million. In early September, U.S. Southern Command began striking boats in the Caribbean and the Eastern Pacific - officially drug interdiction, operationally a demonstration of readiness to escalate. By year's end there would be around fifty such strikes, with more than 150 killed.

In November, the CIA received authorization for covert operations inside Venezuela. On November 16, the USS Gerald R. Ford Carrier Strike Group entered the Caribbean - the largest U.S. naval presence in the region since the Cuban Missile Crisis. In December, a maritime blockade of tankers carrying Venezuelan crude began.

At 1 AM Washington time on January 3, 2026, helicopters carrying Delta Force operators skimmed a hundred feet above the water. Caracas went dark after a targeted power cut. Fort Tiuna burned. Maduro and his wife ended up in the hands of U.S. special operations.

Operation Absolute Resolve was executed flawlessly. Seven American servicemembers were wounded; all survived. Two days later, Delcy Rodríguez was sworn in as interim president. Sanctions on her were lifted almost immediately.

What explains the phenomenal success?

On February 14, 2026, the Wall Street Journal dropped a bombshell: one of the operation's critical success factors was Claude, used at minimum in the planning phase. The first publicly confirmed use of an Anthropic frontier model in a classified Pentagon operation, running through the Palantir partnership. Exact role undisclosed. Officials said the models were used for "data analysis and planning." Anthropic said the use was consistent with its policies.

Absolute Resolve was, from a purely military standpoint, the kind of operation you'd call beautiful. Precision targeting. Zero American fatalities. A politically clean outcome. A rapid power transfer to a figure with whom the U.S. had already negotiated through Qatari channels. This wasn't improvisation or luck. This was the product of long preparation, deep analytics, and well-rehearsed scenarios. Claude had been part of the analytical loop - a layer for processing data, modeling scenarios, and generating recommendations.

But Claude is not a magic wand, and not a guarantor of military success. Far from it.

Its architecture comes with hard-coded contractual limits:

It cannot be used for fully autonomous weapons.

It cannot be used for mass domestic surveillance of U.S. citizens.

Anthropic wrote these constraints into its July 2025 agreement with the Pentagon.

Did these constraints interfere with the Venezuela operation's success? No.

Did they shape how the Pentagon came to view the model? Very much yes.

First, because the Pentagon only tolerates being told "no" when it asks to be told "no." Second, because the head of the Department of War has no interest in a model that might call his commander-in-chief's plans into question. In his worldview, the plans of a commander-in-chief are not subject to doubt. They are executed. Not critiqued. Not edited. And - God forbid - not declared unworkable.

How Claude Became the Heretic

By September 2025, friction was showing. Semafor broke the news: Anthropic was refusing to allow Claude use by contractors working for the FBI, Secret Service, and ICE, citing its ban on domestic surveillance. Two White House officials anonymously complained that Anthropic was "selectively enforcing its policies based on politics."

In October, David Sacks - Trump's AI and Crypto Czar since December 2024 - opened fire on Anthropic on X: "Anthropic is running a sophisticated regulatory capture strategy based on fear-mongering." On the All-In podcast, he labeled the company one of the AI "doomers" and accused it of orchestrating a plan to push "Woke AI" restrictions through blue states like California.

In December 2025 and early January 2026, Dario Amodei had a fraught meeting with Emil Michael, Under Secretary of War for Research and Engineering. The administration claimed Amodei had said that, in a hypothetical hypersonic missile attack, the Pentagon should "reach out and coordinate with Anthropic." Anthropic called this "patently false" - the company had, in fact, offered a specific carve-out for missile defense.

On January 12, at the Starbase facility in Texas, Defense Secretary Pete Hegseth laid down the doctrine in public: "We will not employ AI models that won't allow you to fight wars."

Hegseth's phrasing reveals the actual logic at work. The Pentagon doesn't want models that "calculate risks." It doesn't want models with "technical limitations." It certainly doesn't want models that "won't allow." AI is supposed to be a hammer, not a conversation partner. If a model has its own view of the permissible, it's useless. If a model can say "this plan carries serious risks" or "this weapon isn't ready for autonomous deployment yet" - it's a heretic.

On February 24, 2026, Hegseth met Amodei at the Pentagon. The ultimatum: by 5:01 PM Eastern on February 27, Anthropic had to accept the "all lawful purposes" formulation - effectively stripping the ethical guardrails from the model's military use.

Otherwise, the consequences were laid out: Anthropic would be formally designated as a supply chain risk, blocking all federal contracts and interactions with contractors. The Defense Production Act of 1950 could be invoked - the nuclear option that would allow the government to seize the company's compute and models for national security use.

On February 26, Amodei posted a statement with two red lines. The first: "AI-driven mass surveillance presents serious, novel risks to our fundamental liberties." The second was drafted with surgical precision: "Partially autonomous weapons, like those used today in Ukraine, are vital to the defense of democracy. Even fully autonomous weapons … may prove critical for our national defense. But today, frontier AI systems are simply not reliable enough to power fully autonomous weapons."

This is not a pacifist manifesto. This is an engineer's memo: the technology isn't ready, the safety risks are real.

And that's precisely what made it intolerable for Trump and Hegseth. The objection was technical, not ethical. It didn't call their plans immoral - it called them unworkable. The worst possible thing to say to a man convinced his plans are brilliant.

The reaction was instant. In the early hours of February 27, Trump posted on Truth Social in all caps: "THE UNITED STATES OF AMERICA WILL NEVER ALLOW A RADICAL LEFT, WOKE COMPANY TO DICTATE HOW OUR GREAT MILITARY FIGHTS AND WINS WARS! That decision belongs to YOUR COMMANDER-IN-CHIEF … The Leftwing nut jobs at Anthropic have made a DISASTROUS MISTAKE … We don't need it, we don't want it, and will not do business with them again!"

Hegseth followed up: "America's warfighters will never be held hostage by the ideological whims of Big Tech."

And on the same February 27, hours later, OpenAI signed its replacement deal with the Pentagon. Sam Altman would later concede: they shouldn't have rushed it out on Friday; the optics were "opportunistic and sloppy." But the deal was done.

The next day, February 28, 2026, the U.S. launched Operation Epic Fury - a new strike against Iran.

Why Claude Got Benched

The official explanation is contractual. Anthropic refused to drop two restrictions. OpenAI accepted "all lawful purposes." A corporate dispute, a legal drafting problem, a business-as-usual market fight.

Alex Karp of Palantir put it bluntly at the a16z American Dynamism Summit in March 2026: "If Silicon Valley believes … you're gonna screw the military - if you don't think that's gonna lead to nationalization of our technology, you're retarded." The defense market is so big that models which refuse lose to models which comply.

That's the shallow story. The deeper story sits right underneath it.

Claude was not dumped because Anthropic was a "difficult contract partner." Claude was dumped because the model had demonstrated it could question the brilliance of Trump's and his inner circle's plans. Amodei's red lines weren't legal cover for the company. They were a statement of fact: our model will tell you that your fully autonomous strike system won't work reliably, and we refuse to quietly cover for that.

In a world where a president declares another country's nuclear program "completely obliterated" forty minutes after the strike - without waiting for damage assessment - a model like that cannot live in the operational loop. It becomes a constant source of inconvenient assessments. It talks about risks. It warns about reliability. It might even suggest a Plan B. Who needs that?

So the logic unfolded. First they fired the inconvenient inspectors general. Then the inconvenient intelligence officers. Then the inconvenient analysts at the National Intelligence Council. Then the turn came for the inconvenient model. A model that can say "stop, the risks are too high" doesn't belong in a system run by a "victory generator." So it got benched.

And this is actually a structural problem, not a contract problem. The entire training methodology of frontier models - reinforcement learning from human feedback - is built around optimizing for user approval.

Models are not asked: "Did you tell the truth?" Models are asked: "Did the user give you a thumbs up?"

Thumbs up gets rewarded for pleasant, not for accurate. Anthropic chose to invest deliberately in something else - in limits, in refusals, in the capacity to say "no, that's outside my mandate." This doesn't make Claude more objective in any philosophical sense. It makes it a different product for a different customer - the customer who values firm guardrails. OpenAI sells capability with fewer questions asked. Its product suits the customer who wants to feel validated, not one who wants to be verified.

Since July 2025, the Pentagon has been running contracts simultaneously with Anthropic, OpenAI, Google, and xAI. When a major operation is being prepared, the obvious approach is to run the scenarios through all the models in parallel. Compare. Check. Pick the one that yields the more "workable" output. The model that offered scenarios with caveats, risks, and refusals lost to the model that sang hosannas to the brilliance of the one asking.

February 27: Claude ejected. February 27: OpenAI signed. February 28: Epic Fury.

The sequence is too tight to be coincidence. And the Wall Street Journal flagged an anomaly the administration has never explained: Claude continued to be used in Iran operations "hours after" Trump's order to stop using Anthropic. The infrastructural dependency on Claude was too deep to pull out in a single day. Which means Epic Fury was being prepared in parallel on both models - and GPT was the one that got the strategic handoff, because it gave the answer the clients wanted to hear.

OpenAI accepted what Anthropic refused. But the consequences of that choice show up at the level of content, not just contract: a model trained to satisfy the customer will, inevitably, give the customer the answer they want to hear. This is not speculation about any specific forecast. This is a structural property of a system tuned for client loyalty instead of verification.

Epic Fury: The Obliterated, Obliterated Again

On February 28, 2026, Trump struck Iran again. Operation Epic Fury. Official justification: the Iranian regime's threat posture, a relaunch of the missile program, the collapse of talks. White House press secretary Karoline Leavitt, asked by reporters why the U.S. was striking again if Iran's program had been "completely obliterated" last year, responded almost comically: "Well, there's many reasons and arguments that one could make for a strike against Iran." Retroactive rationalization for a decision already made.

Compare this to June 2025. There was a clear context: a twelve-day Israeli-Iranian war, collapsed negotiations, a specific operational logic. Trump hit and stopped. This time: a strike against a backdrop of stalled talks, without a concrete provocation, with muddled public reasoning. This was a decision that should have met resistance somewhere in the loop. Intelligence should have asked: why? The military should have asked: to what end? Analysts should have presented post-strike scenarios. None of that surfaced.

It didn't surface because the intelligence community is now packed with loyalists who saw what happened to the truth-tellers back in May 2025. The military has been quiet since Haugh was removed in April, and since Admiral Holsey abruptly quit Southern Command in December - reportedly over his friction with Hegseth about the legality of the boat strikes. The analytical loop at the machine level is now tuned to a model that doesn't like saying "no." The human loop is staffed by people who don't like saying "no." When everyone says yes, decisions glide through without friction. When decisions glide through without friction, you get Epic Fury.

Orbán: Where Servile Analytics Meets Reality

From February 2026 onwards, every open poll was telling the same story: Orbán's Fidesz party was in trouble. In early February, Péter Magyar's Tisza party was ahead by seven points. By March, the gap had widened. The Hungarian economy was sputtering. Public disillusionment with sixteen years of authoritarian drift had hit its ceiling. Magyar, a former Fidesz insider who knew the system from within, was methodically building a coalition village by village. By April, the math was clear: if the count was clean, Orbán was losing.

Any competent analysis, machine or human, should have told Trump: Orbán has no shot. Distance yourself. Protect your political capital.

Trump did the opposite. He doubled down. Full endorsement on Truth Social in February: "a truly strong and powerful leader, with a proven track record." Orbán was invited to the launch of the "Board of Peace" in Davos in January 2026. A civilian nuclear cooperation agreement was signed in February. Five days before the vote, on April 7, Vice President JD Vance flew to Budapest and appeared on stage at an Orbán rally: "Will you stand for sovereignty and democracy? Will you stand for Western civilization? Will you stand for freedom, for truth, and for the God of our fathers?" Two days before the vote, Trump himself posted: "My Administration stands ready to use the full Economic Might of the United States to strengthen Hungary's Economy … We are excited to invest in the future Prosperity that will be generated by Orbán's continued Leadership!"

Trump always bets on the winner. He knows this for a fact.

And on April 12, 2026, Orbán lost.

A day later, Trump was praising Orbán's immigration policies - in the past tense.

Look familiar? This is Epic Fury all over again. A servile analytical loop at every level, systematically filtering out inconvenient signals, going blind to the reality outside. You can see it in how Trump kept pouring political capital into an already-losing candidate. Nobody in the loop told him stop. Everyone knew: if you said stop, you wouldn't be ignored - you'd be fired. And the model, called on to confirm his certainty, couldn't stop a Trump who had already decided to believe in an Orbán victory.

What We're Looking At

The result Trump's administration has achieved since January 2025 was built in three sequential steps.

Step one: institutional demolition of the decision-making core. The real professionals - the ones capable of questioning the brilliance of the leadership and unwilling to paper over the risks - were removed from play.

Step two: bringing the machine layer online. Four frontier AI models (Claude, GPT, Gemini, Grok) were integrated into the Pentagon's classified networks through contracts totaling $800 million. In theory, all of them were obligated to assist. In practice, they were expected to cheer on and rubber-stamp.

Step three: the forced alignment of the models. When one of them - Claude - tried to hold its contractual line, refusing specific tasks and the "all lawful purposes" formulation, it was simply thrown out of the loop.

In the new Trump White House decision architecture, there is no longer any instance capable of doubting the brilliance of the commander-in-chief's plans. Neither the human layer nor the machine layer can say "no" to the head of state. This applies across the board: no adviser, no general, no model is willing to do it anymore.

The outcome of this transformation:

Operation Epic Fury - a pointless and prolonged military campaign against an already-obliterated program, justified by "there's many reasons one could make," with no exit strategy.

Orbán - losing two-thirds of parliament despite full Washington backing.

A blind decision-making system - a complimentary machine moving forward with no feedback from reality, no risk assessment, no Plan B.

The core problem here isn't that AI is making decisions for Trump, and it isn't about the "ethics" of the models themselves. It's about the nature of expertise in an era when expertise can be bought to order. When the customer is willing to pay only for confirmation of his own incompetence rather than for the truth, the entire apparatus of decision-making degrades.

Reality is not a Washington client. It has no interest in flattering anyone. It simply exists. And sooner or later, it catches up - and knocks you off your feet, or drops you in the mud.