What the Anthropic-Pentagon standoff reveals — and what was already known
By Chris Kohler, Human Peace Project
On February 27, 2026, a deadline passed.
The Pentagon gave Anthropic — maker of the AI model Claude — until 5:01 PM Eastern to surrender its safety guardrails or face being labeled a national security threat. A supply chain risk. The same designation reserved for Chinese telecom giant Huawei. Applied to an American company. For refusing to agree that its technology could be used, without restriction, for mass surveillance of American citizens and for weapons that kill without a human making the final decision.
Anthropic held its line. The U.S. government stopped using Anthropic’s products entirely. The Pentagon moved to designate the company a national security risk.
For over a decade, I have watched the conditions for this moment assemble themselves — not with any certainty about when or how it would arrive, but with the knowledge that something would eventually force the real question into public view. Not the science fiction version. The actual one, embedded in policy, in procurement, in the legal frameworks that govern what machines are permitted to do to human beings without a human hand on the trigger. I write here to those who don’t want to risk waiting for the science fiction version.
The Human Control of Weapons Act was drafted in 2014 precisely because that question was already urgent — because the trajectory was already visible to anyone paying attention to where technology policy and weapons development were heading. Congress couldn’t move then. It moves slower now, in a world where AI capability doubles every eighteen months and major policy framework takes a decade to pass if it passes at all.
That gap — between the pace of the technology and the pace of democratic response — is where we are living. That is what this moment is. My question then and now:
Who controls the machines we are building to make decisions about life and death — and on what human basis do they make those decisions?
A Room Several Blocks from
the White House
In the fall of 2014, I walked into an office building several blocks from the White House — the Center for a New American Security, which has since relocated and now overlooks the White House. I was heading toward the security desk when a large man intercepted me in the lobby. He knew my name. I had no idea how. I had not gone through any formal vetting process. There was no reason anyone in that busy lobby should have been able to pick me out of the crowd.
He took me to a private elevator. We went up. I was walked directly into a conference room, bypassing the usual institutional front end — the receptionist, the waiting room, the small signals that tell you what kind of place you are entering. I was at the table almost before I had time to think.
During the months I was invited to sit around that table, it was occupied by people wearing military medallions, Department of Defense and Department of State attorneys, Ivy League ethicists, weapons systems experts, computer scientists. In the center of the table was a black triangle-shaped device. I was told the President could be listening at any time.
I had been introduced as a representative of the Christian community. That was true as far as it went — I am a Christian, and the moral framework I brought into that room was not abstract. It was specific, old, and demanding: an insistence, running through certain strands of Christian witness for centuries, that a human being cannot abdicate direct moral responsibility for violence to any system, institution, or machine. I was clear when I spoke from that place without misrepresenting anyone else’s voice. There are tens of millions of Christians in this country for whom the questions I was raising are not abstract policy debates but moral obligations of devotion.
I had no security clearance. No advanced degree in weapons systems or international law. In that first meeting, the work here eventually named the Human Peace Project had no name yet — though it had been the work of my life long before I walked into that building. I had arrived at the issue of autonomous weapons through a lifetime of peace and environmental activism, through grief, through watching institutions fail human beings at the most intimate possible scale, and through the simple observation that no one else seemed to be doing what needed to be done.
After the meeting ended, as I was exiting the room, a man in a blue suit who had not been sitting at the table but was sitting in a chair against the wall stopped me and gave me his card. Department of Defense attorney. In private, one on one, he told me he was glad I had shown up. He explained that the military was not monolithic — that there was a genuine split, going back to World War One, between those who understood that weapons operating without human control had repeatedly backfired on the commanders responsible for them, and a newer generation shaped by urban combat in Iraq and Afghanistan who saw autonomous weapons as a solution to watching their fellow soldiers and civilians die in close-quarters fighting. He mentioned, without elaborating, that 3000.09 had been necessary — that there was some resistance at the top of the military to going down the autonomous weapons path.
He didn’t say which side he was on. He didn’t have to.
What a General Told Me
There was another gathering — entirely separate from CNAS. I do not recall seeing anyone from CNAS there. At this meeting, the military’s four branches each made presentations to the defense manufacturing community about their plans for autonomous weapons. About a hundred people filled that room. Top defense industry sales executives worked it. The people building the systems and the people buying them, in the same space.
Part of me deliberated about whether to engage with that world at all. I decided it was an opportunity to gather information. I never suffered the illusion I was going to change anything in there.
I joined the line of people waiting to speak with the senior military figure present — a retired general with decades at the highest levels of the Pentagon, the man to whom everyone in that room looked as the authority. When I reached him I asked what I hoped sounded like an innocent question, and without revealing my agenda. The military had been using autonomous weapons in various forms for decades, I said. What exactly was the big fuss about now?
He answered without hesitation. I was just another one of the hundred or so people in that room. He spoke plainly.
He said autonomous weapons were going to be used offensively.
A general’s few simple words answered what I had come to Washington to figure out. What was the language of DoD Directive 3000.09 dancing around? Now I might finally know. The man who said it was not advocating. He was describing. Which made it worse.
What any resistance at the highest levels of the military was actually about, I could only speculate. But a senior military figure had just told me plainly that the real fuss was about offensive weapons.
As the room began to empty I stood alone, the defense industry sales executives talking shop within earshot. A man I had noticed standing quietly by himself against the wall throughout the proceedings walked quickly past me. He didn’t stop. He didn’t look at me. But as he passed, close enough where only I could hear him say quietly, almost under his breath: “Thank you.”
Then he was gone.
I don’t know his name. I have no idea what his affiliation was to any of the institutions represented that day. My instincts tell me he was military. I don’t know exactly what he was thanking me for — perhaps something I had asked of one of the presenters. Maybe he had seen me doing my work elsewhere in DC. Maybe he overheard my interactions with some of the defense marketers in that room. But he was clearly unwilling to be seen with me. Whatever it cost him to say it, he said it anyway.
I have carried that “thank you” for more than a decade. Somewhere in that room was a person who understood what I was trying to do. He couldn’t say so openly. But he said so.
Another moment I have thought about many times since: In that room I approached the director of a significant AI project — a woman whose position put her at the center of exactly the technological development I was there to understand. I handed her one of my Human Peace Project cards and explained why I was there.
She started to shake.
She knew something. Whatever she knew, holding that card in her hands brought it out.
The Language
In November of 2013, while lobbying in DC on another topic, I brought up my personal concern of autonomous weapons in what I thought would be a receptive congressional office. That office, Hank Johnson’s, was receptive. A Chief of Staff of this US Congressional representative told me to come back with a legislative proposal on the issue of autonomous weapons.
I went home and started reading. Somewhere in that reading I found DoD Directive 3000.09 — the Obama administration’s 2012 policy on autonomous weapons. My first reaction was simple: The directive’s language is constructed to permit what it claims to prohibit.
What I first found were the words: semi-autonomous and autonomous.
The Obama directive to the military drew a line around “autonomous” weapons — systems that could select and engage targets without human intervention. But it left open a door for “semi-autonomous” systems. That distinction may sound reasonable. It isn’t. And “autonomous”? No better. Both are rhetorical constructions, not technical specifications. Together they create a spectrum with no fixed moral point — a continuum along which targeting decisions can be progressively handed to machines while a nominal human somewhere in the chain provides what lawyers could call “meaningful control,” another phrase flexible enough to mean almost anything. “Autonomous” gives the appearance of a hard limit. “Semi-autonomous” provides the escape route. Together they are a permission structure dressed as oversight policy.
Who wrote that language, and whether the person who signed it fully understood what they were permitting, I cannot say here with certainty. What I can say here is what that language does. It opens a door and calls the door a wall.
Mark Gubrud — a physicist and one of the most technically fluent voices on autonomous weapons in the world — wrote the first draft of what would become the Human Control of Weapons Act. He also brought to the proposed legislation the framing that became its title and its spine: human control. Rather than trying to construct an airtight definition of what an autonomous weapon is, what a machine can and can not do — a definitional fight the Pentagon would always win — the Act would assert what any weapons system must have: What a human being must do. That reframing was Mark’s greatest contribution to the proposal, drawn from his deep engagement with the AI research community, and it was exactly right.
I took that first draft from Atlanta to Chicago by Megabus, specifically to sit in front of Marc Chagall’s America Windows at the Art Institute to work through not only what I was willing to put my name to, but to separate myself from any and all of my usual distractions while being guided by my conscience and that art. I took Greyhound from Chicago to DC. I then took the proposal to Capitol Hill and to other DC contacts.
Closing both the “fully autonomous” and “semi-autonomous” escape hatches was vital. So was an addition that came after the first draft: an exception for genuinely defensive systems. The argument for human control of weapons cannot ignore the legitimate need to protect the living from incoming harm. That exception is narrow, carefully bounded, and surrounded by accountability requirements. It cannot become another escape hatch.
I was ambivalent about the name itself. “Human Control of Weapons Act.” It very aptly describes the Act while sounding so awful. It is perfect, but sadly perfect.
Mark Gubrud’s technical contribution was indispensable. Without his scientific precision the document would not have been serious. What I brought was different: decades of reading what institutions say versus what they mean, and my own moral reference points that determined what I was willing to bring to Congressional offices, cabinet-level departments, DC NGOs, and the world.
Today, in his public statement refusing the Pentagon’s demands, Anthropic CEO Dario Amodei drew his line around “fully autonomous weapons.” The Pentagon said it had no interest in “autonomous weapons that operate without human involvement.”
The escape hatch is still open. The language has barely changed in twelve years.
“Autonomous.” “Semi-autonomous.” “Human involvement.” “Meaningful.” These words hardly mean what they sound like they mean. They are the architecture of permission — the verbal infrastructure that allows the technology to advance while appearing to be constrained. I recognized this permissive construction in 2014. I am watching it operate again today, in real time, in a dispute that the entire world is now watching.
Where Anthropic’s Line Falls Short
Anthropic CEO Dario Amodei’s stand deserves respect. But a careful reader will notice that his objection is primarily technical: Today’s AI is not “reliable enough” for fully autonomous weapons. That is a meaningful limit. It is not a permanent one. A system that becomes reliable enough, in the logic of his argument, might eventually clear the bar.
Let me be clear: The requirement that a human being take positive action to initiate the engagement of each specific individual target is not a reliability standard. It does not move when the technology improves. It is a moral requirement. A human being cannot hand off to any machine — however sophisticated, however capable — the weight of deciding that a specific person, in a specific moment, will die. Not because machines always decide worse than humans. Because machines cannot be held accountable for what they decide. The accountability disappears into the architecture. That is not a flaw in autonomous weapons. It is their appeal. They abdicate humans from their part in a moral act. We are hoping to govern our way out of a moral responsibility with a machine module.
Again, “fully autonomous” and “semi-autonomous” are not meaningfully different categories. They are different points on a continuum whose destination is the same. The language that separates them is not a technical specification. It is a permission structure.
There is one further distinction worth quickly naming here. The moral calculus is not the same for defensive and offensive autonomous systems. A weapon designed to protect the living from incoming harm occupies different moral ground than one designed to seek and destroy. That distinction is important. The current debate has not yet caught up to it.
What Happened in Caracas
The public drama — the ultimatum, the deadline, the Truth Social post, the OpenAI deal announced hours later — obscured a simpler and more disturbing question: What was Claude actually being used for?
My best read, informed by what we know about how these systems work together, is something far more mundane than the science fiction of lethal autonomous robots. Palantir’s platform ingests intelligence — signals, communications intercepts, movement data, surveillance feeds. A large language model — the technology underlying systems like Claude — sitting on top of that infrastructure does what large language models do: It synthesizes ambiguous, incomplete information into coherent, confident-sounding output.
In the operation to seize Nicolás Maduro, that output was probably something that looked like an intelligence product — a prioritized picture of who was where, who posed a threat, who needed to be neutralized to get a team in and out. A targeting list.
We have seen this before. Israel’s AI systems — Lavender, The Gospel, Where’s Daddy — processed mass surveillance data to generate lists of people to kill in Gaza. The mechanism in Caracas was in all likelihood the same mechanism, applied to a different theater. People who were not Maduro died. Could be an Anthropic employee saw enough of the output to be troubled by it. That employee raised a concern. And that concern, surfacing at the wrong moment for the wrong people, triggered everything that followed.
Regardless of what happened in Caracas, here is what matters technically, and why the OpenAI deal that replaced Anthropic within hours deserves scrutiny rather than relief: OpenAI’s claimed guardrail is cloud-only deployment. The argument is that by keeping the model in the cloud rather than on edge hardware, you prevent it from being integrated directly into weapons systems. That is a real architectural constraint. It is also beside the point. The targeting — the Palantir data fusion, the intelligence synthesis, the list of people — happens through the cloud. The weapon doesn’t need the AI on board. It needs the AI upstream. Cloud deployment doesn’t close that gap. It is the gap.
OpenAI’s agreement with the Pentagon references existing law as its protection against autonomous weapons and mass surveillance. What laws was OpenAI referring to? “Lawful use” is the phrase that does the work. We have seen this construction before — in DoD Directive 3000.09 (as described above in What I Saw in the Language). The “fully autonomous” and “semi-autonomous” distinction was a permission structure dressed as a 2012 policy. “Lawful use” referred to by OpenAI is the same policy ambiguity updated for 2026. What the law does not yet contemplate, it cannot prohibit. And the law has never contemplated AI operating as it is operating now.
Congress was asked to close the legal gaps. The Human Control of Weapons Act was presented to Congressional offices in 2014. The gatekeepers held the gate. The technology did not wait. It never does. What was a policy gap twelve years ago is now operational reality — in Gaza, in Caracas, and in whatever comes next.
Anthropic’s employee who raised a concern did what the Human Control of Weapons Act was designed to make structurally possible: A human being, in the chain, said wait. The Pentagon’s response was to remove Anthropic from the chain entirely and replace it with a partner willing to accept “lawful” as sufficient. OpenAI’s Sam Altman admitted the deal was rushed. He did not explain what, exactly, he had agreed to.
The language has barely changed. The stakes have.
How I Got Here
I need to tell you something about how I arrived here, because it matters to what I am asking of you.
I have been doing this since 1976, when as a thirteen year old I used the results of a science project to lobby local government about sewage overflow into streams. In college, friends and I crafted a giant Nuclear Freeze postcard signed by Bernard Lowndes and hundreds of others to be presented to the Reagan White House. We asked Julia Child to sign it. She declined. Others declined, too.
I have been asking ever since — of whoever needed to be asked, in whatever room I could get into. In the decades that followed: Environmental battles at the local and state level. Child advocacy work that taught me to write for legal audiences. Anti-war organizing. A state Soil and Water Commissioner once quipped she thought I had more stop work orders than any other environmentalist in the state. I pushed for and successfully used a bipartisan legislative hearing to stop public corruption, exposed activists manipulating petition language after they had signatures and other underhanded dealings, and carried the cash that bailed a future United States Senator out of jail. I have sat across tables from people whose names you would recognize and said no when no was the right answer, and kept those conversations private because that is what integrity requires.
I am not recounting this to establish credentials. I am recounting it because the question I am always asked, implicitly or directly, is: how did you end up in those rooms? The answer is that I have been finding my way into rooms where decisions were being made since I was thirteen years old. Not because I had the right connections or the right degrees, but because I care, I do my homework, and I show up.
The path to Washington on autonomous weapons began years earlier, in a congressional office, when I asked whether Congress was doing anything on the issue of autonomous weapons. The answer was no. I was told to prepare something and come back.
There is also this: I did not come to autonomous weapons policy through a think tank fellowship or a defense contractor contract or a university appointment. I came through grief, sadness, worry, and a sense of responsibility to my world.
My mother died badly, inside an institution that protected itself rather than her. I will leave it at that. Before her death I had been her primary caregiver for years. She was on 24-hour oxygen, could not walk without a walker, and was often in a wheelchair or bedridden. I knew she would likely die well before me. That is the way it is supposed to be. But the way she died left a silence in me that had no bottom. I lost nearly a third of my weight in the months that followed.
The depression was deep and long. I had been through depression before. Years before, bedridden and addicted to prescribed painkillers, I documented what appeared to be illegal development in woods behind my home — climbing a railroad tie retaining wall in pain every day to photograph the violations. That battle lasted two years and ended in the developer’s bankruptcy. The climbing was excruciating and also physical therapy. I came off the pharmaceuticals somewhere in those two years.
A retired editor of a national magazine, a man I met at a conference on drone warfare where Cornel West spoke, heard this story over lunch. He told me it was a story about recovering from depression. He offered to write a play about it. I never followed up. I was too busy, particularly with caring for my mother.
After my mother died, I could not pull myself out of depression. Fortunately, I finally remembered what that writer said about the play he wanted to pen. So I went looking for a brand new cause. Getting into action for a cause larger than myself was the antidote for my depression.
I found the weaponization of artificial intelligence. And that writer may have saved my life.
I thought: What could be worse than imparting our violent, fear-driven, controlling nature into machines that can think and act faster than we can?
Have you already felt the start of it? The office that lost your file. The medical appointment where someone checked boxes while you spoke. The denial letter with no name on it. The phone line where the human never comes. The people on the other end of those experiences were not indifferent by character. They had become indifferent by function. Each one a component, processing a case, following a protocol, insulated from the human consequence of the next decision by the human who made the one before it. The machine made them that way. Or rather — they had built the machine that made them that way. They keep stepping inside it.
That is the machine we are putting into the machine.
Not artificial intelligence. Us. Our who-lives-and-who-dies processes, our institutional indifference, our diffusion of accountability across enough human components that no single one feels responsible. But through weaponized AI it will be accelerated, automated, scaled beyond any individual’s ability to see or stop.
Yet it is not AI that does any of this. Humans, as the machine parts we can be, will do it to each other through AI — faster and with far less accountability than any human institution ever has.
What Nobody Was Doing
I grew up in the shadow of the CDC. My next-door neighbor was a CDC virologist. On the other side of my house lived a CDC parasitologist. Eight doors down lived a man — also an assistant scoutmaster to many of us — who would go on to significant roles in global vaccination programs at the WHO. Across the street from him lived a doctor I understand to have been connected to the discovery of herpes, and known for holding what may be one of the oldest existing samples of HIV — preserving it because he claimed the genetic science was not yet sufficiently advanced to analyze something so rare without destroying it. His wife was like another mother to me.
My own family, when they weren’t abusing, were neglecting. These neighbors, along with a handful of school teachers, raised me. A few of them taught me to gather my own primary evidence. To trust careful observation over received consensus. To hold a question longer than is comfortable. To resist the institutional conclusion when what you have actually seen tells you something different. I also witnessed those CDC people argue across their backyard fences and debate at cocktail parties, while speaking more as one in public, like machine parts.
Many of those childhood formations shaped a lot of my life that followed — including how I read 3000.09.
Alone in my house researching the issue, I read what Human Rights Watch was saying — an echo of the problem, but not enough. I checked FCNL, my favorite peace organization. They seemed to be focused entirely on climate. Nobody was doing what needed to be done.
So I asked myself, as I had learned to do: What would it take? I had lobbied on Capitol Hill before. I understood strategic selling — the discipline of identifying who the real decision-makers are, mapping the actual paths to those people, building the relationships that open the right doors, and doing it with precision rather than broadcasting into the void. I had spent decades as the person between buyers and talent in the marketing world — the one who made things happen. I handled marketers’ manipulations of their customers. I read every contract and proposal for what it was actually committing everyone to. I had written many judicial reports on child sexual abuse and could write in legal language.
And I had a direct line: The staffer in Hank Johnson’s office who had told me to come back with something. It was clear this issue was the one I was called to do at that time in my life.
I presented the Human Control of Weapons Act in Washington — to Johnson’s office, to Jim McGovern’s office, and to others, including across the aisle as well.
McGovern’s office was reportedly the Democratic Party’s gatekeeper on weapons systems policy.
After those visits, things on Capitol Hill dried up.
What This Age Asks of Us
I want to be honest with you about something I do not often say aloud.
My time in Washington, and what I learned when I returned to the world with the information I had gathered, left me genuinely fearful that there may not be anything any of us can do. The machine parts of concentrated power — the billionaires, the top dogs in the defense industry, the political leaders they corrupt, the interdependent institutions that protect themselves before they protect people — do not yield easily to ordinary citizens with ordinary means. I watched it operate up close. I came home with few illusions.
If you are waiting for Congress to pass a Human Control of Weapons Act, or for the Pentagon to voluntarily constrain the financial drives of the defense industry selling to them plus lobbying your elected leaders with campaign donations and or gifts, or for any institution to save us from the consequences of machine autonomy owned and controlled by the elite few — I understand that hope. I cannot confidently share it as a viable solution to the world which seems to be coming.
Call your representatives. Trust that nations and international bodies will eventually construct frameworks. These things are worth doing. Do them.
But what I am actually asking — what I think this moment genuinely requires, as the power of AI doubles every 18 months, faster than policy can be enacted, as the powerful relentlessly pursue more money and power — is something older and less glamorous than political action. It is to watch one another’s backs. To refuse the divisions that political leaders depend on for their relevance and their power. To not fall for the idea that the person on the other side of whatever line your leaders have drawn for you this week is your real enemy.
The question of who decides whether machines can kill without human accountability is not a partisan question. It is not a question the left has answered or the right has answered or any institution has answered. It is a question that belongs to all of us, and the people with the most at stake in the answer are the ones with the least voice in producing it.
War has always benefited the wealthy and cost the poor most dearly. That has not changed. The technology of control through violence, or the threat thereof, will not change this.
Watch each other. Tell each other the truth. Don’t let them divide you.
That is my appeal.
The Human Control of Weapons Act
The Human Control of Weapons Act was a starting place — a floor, carefully constructed, with room for legitimate defensive uses and no room for the language games that allow the floor to be removed. It was presented to Congress in 2014. It was not taken up. The gatekeepers held the gate.
It is here now, on this site, for anyone to read. Not because I believe an Act of Congress will save us. But because the arguments embedded in its language are sound, and because the people who need to make the decisions that actually matter — the engineers, the soldiers, the citizens, the human beings who will be asked to trust these systems with their lives and the lives of others — deserve to know that someone already did the work of thinking this through, and that the thinking holds.
Read it, share it if it moves you. And whatever you do next, do it in the knowledge that this question — who controls the machines we are building to make decisions about life and death — has always belonged to you.
The Human Control of Weapons Act is available in full at HumanPeaceProject.net