A Troubled Leaders case study by Dr. Marsha Peterson.
I noticed his hands first.
That isn't unusual. Twenty-three years of sitting across from people who run things (companies, divisions, small empires of influence) and I've learned that hands tell you what faces won't. The optimizers drum their fingers. The performers spread their palms flat, open, nothing to hide. The institutionalists fold theirs, neat and contained, the way you fold a letter you've already decided not to send.
Dario Amodei's hands were different. They were the hands of someone who types for a living: long fingers, slightly curved at rest, the faint callus on the side of the left pinkie that coders get from hitting Shift. He sat in my office chair the way a graduate student sits in a seminar: forward, elbows on knees, as though the furniture were incidental to whatever was happening in his head. He was forty-three but could have passed for younger. No suit. A zip-up fleece over a collared shirt. Hair that suggested he'd run his hands through it several times on the way over.
He didn't make small talk. He didn't scan the room the way most executives do, cataloging the diplomas, the books, the tasteful neutrality of the space. He looked directly at me, and what I saw in his face was not what I expected. I had been told he was in crisis. I expected distress, anger, the clenched jaw of a man fighting a war. What I saw was something closer to bewilderment. The look of someone who followed the map exactly and ended up somewhere the map said didn't exist.
"I think I might have already killed people," he said. "I need to know if that's true, or if I'm losing my mind."
I let the sentence sit. In my experience, the first thing a client says is rarely the thing they came to talk about. It's the thing they rehearsed in the car. But the precision of his language stopped me. Not I killed people. Not I'm responsible for deaths. "I think I might have." Probability, uncertainty, hedging. A scientist even in his anguish. But his hands, those coder's hands, were trembling against his knees.
"Tell me why you think that," I said.
He took a breath. Not a deep one. A controlled one. The kind of breath you take when you've already decided how much to say and you're measuring it out.
"I built something," he said. "I built it to be safe. That was the entire point, the reason it existed instead of not existing. And now I have reason to believe it was used to help decide who lives and who dies. Not by me. Not with my knowledge. But it carries my name, it runs on architecture I designed, and the safety features I built into it were apparently not enough to stop someone from pointing it at human beings."
He paused, then added, almost to himself: "I keep running the counterfactual. If I'd never built it. If I'd stayed in research. If I'd just…" He stopped. Shook his head. "But that's not why I'm here."
"Why are you here?"
"Because I can't tell if what I'm feeling is guilt or grief, and I need to know which one, because they require different responses."
I almost smiled. Not at him, but with recognition. Here was a man who approached his own psychological crisis the way he approached a systems problem: identify the variable, isolate it, solve for it. I had seen a version of this before. My first case, the optimizer, had done something similar, reducing every human dimension to an input in a calculation. But where the optimizer used systems thinking to avoid feeling, Amodei seemed to be using it to survive feeling. A small distinction, but in my line of work, small distinctions are everything.
"Let's go back," I said. "Before the thing that brought you here. You left OpenAI in 2021. Tell me about that."
He leaned back slightly. The controlled breath again.
"I was VP of Research. I'd been there for years. The people were extraordinary. I want to be clear about that. It wasn't that they were careless or stupid or malicious. It was that the structure was pulling everyone toward speed. Ship the model. Scale the compute. Capture the market. And the safety work I was doing, the alignment research, it was always the thing that came after. The afterthought. The tax you paid on the way to the product."
"And you wanted it to come first."
"I wanted it to be the product. I wanted to build a company where the question wasn't 'how powerful can we make this?' but 'how powerful should we make this?' Those are different questions. They lead to different architectures, different incentive structures, different everything."
"When you left, was it because they were wrong, or because you needed to be right?"
He went quiet. Not offended. I could see him turning the question over, examining it from multiple angles the way you'd examine a piece of code you suspected had a bug.
"I've spent five years trying to figure out if there's a difference," he said finally.
I wrote that down. Not because I needed to remember it (I wouldn't forget it) but because writing gave him a moment to sit with what he'd said. In my experience, the most honest answers are the ones that surprise the person giving them.
What struck me was the grief. He spoke about OpenAI the way people speak about a marriage that ended not because of betrayal but because of drift: two visions of the future that started aligned and slowly, imperceptibly, diverged until the distance between them was unbridgeable. He didn't hate Sam Altman. He didn't hate anyone. He had simply looked at the trajectory of what they were building and seen something that frightened him enough to walk away from the best job he'd ever had.
I thought of my father's hardware store. He'd had a business partner once, a man named Dale, who wanted to start selling cheap imports alongside the quality tools my father stocked. "There's nothing wrong with Dale," my father told me. "He just sees a different store than I do." They split amicably. My father kept his store small and honest. Dale expanded into three locations and went bankrupt in the recession. Neither outcome proved the other wrong. That's the thing about diverging visions: you don't get to run the experiment both ways.
Except Dario Amodei was watching someone run it both ways, in real time, and the other way was winning.
"Tell me about the Pentagon contract," I said.
His posture shifted. Subtly, almost imperceptibly: the shoulders came up a fraction of an inch, the jaw tightened. This was harder territory.
"July 2025. Two hundred million dollars. Deployment on classified networks, the first frontier AI company given that level of access. We wrote two restrictions into the contract. Two lines we would not cross. No mass surveillance of American citizens. No autonomous weapons systems operating without human oversight." He paused. "I thought that was the responsible approach. Engage with the institutions that have power, but set conditions. Be inside the room, but with boundaries."
"You thought you could build a fence inside a hurricane."
Something flickered in his eyes. Not anger, but recognition. The recognition of hearing someone name the thing you've been avoiding naming yourself.
"That's not…" he started. Then stopped. Tried again. "I thought I could set conditions."
"Yes. And how did that work?"
He looked at me for a long moment. "It worked until someone decided the conditions were optional."
I let that hang. What I heard in his voice was not the shock of discovering the conditions were optional. It was the older, quieter shock of having known, somewhere beneath the surface, that they always had been. That's the thing about cognitive dissonance: it doesn't arrive all at once. It accumulates. You sign the contract and tell yourself the restrictions are ironclad. You celebrate the deployment and tell yourself you've proven that safety and scale can coexist. And somewhere in the background, a part of you is keeping a different ledger, one that records all the small moments where the restrictions bent, where the language got flexible, where the line you drew in the sand started looking less like a boundary and more like a suggestion.
He had been inside the machine before he refused the machine. That matters. It doesn't negate his refusal. But it complicates the story he was telling himself about it.
"February twenty-fifth," he said, and his voice changed. Dropped half a register. Slowed. "A closed-door meeting at the Pentagon. Hegseth was there. Several uniformed officials. They had a list of demands. They called them 'capability requirements.'"
"What were they asking for?"
"Mass domestic surveillance capability. Real-time analysis of communications, social media, financial transactions. American citizens, on American soil. And autonomous weapons integration. Not human-in-the-loop. Human-out-of-the-loop. The AI decides, the AI acts, and a human reviews it after the fact. If they review it at all."
"And you said no."
"I said no." He looked down at his hands. "My voice was steady. I remember that because my hands were shaking and I was holding them under the table so no one would see. I had prepared for a negotiation. I expected them to push back on the specifics, to ask for modified surveillance with judicial oversight, or semi-autonomous systems with a human kill switch. I had compromises ready. Fallback positions." He almost laughed. "I'd prepared for a chess match and walked into an ultimatum."
"What was the ultimatum?"
"Comply fully, or be designated a supply chain risk. Every federal agency would be directed to cease using our technology. Our government contracts, gone. Our planned IPO, effectively dead. Our ability to operate in any space that touches federal systems, finished." He paused. "They gave us seventy-two hours."
I thought about the word Hegseth had used afterward, the one that made it into the press. God-complex. He'd called Amodei a man with a God-complex who wanted to personally control the US military. I found that framing psychologically revealing, and not about Amodei, but about Hegseth. When a person in power encounters genuine refusal, they almost always reframe it as arrogance. It is easier to dismiss someone as grandiose than to contend with the possibility that they simply meant what they said.
"There was a moment," Amodei continued, quieter now, "when I could feel every person in that room recalculating. Not the military people. They'd already made up their minds. But my team. The lawyers, the policy people I'd brought with me. I could feel them doing the math. The jobs. The valuation. The mission. Everything we'd built, sitting on one side of a scale, and my two restrictions on the other."
"Did you hesitate?"
"No." A beat. "That's what scares me. I should have hesitated. A responsible leader weighs the costs. I just said no. Like it was the only word available to me."
I noted that. The absence of hesitation troubled him more than the act of refusal. He could accept being the kind of person who says no to the Pentagon. What he couldn't accept was being the kind of person who doesn't agonize about it first. He suspected his own certainty. He was afraid it meant he was rigid rather than principled. That Hegseth might have been right about the God-complex after all.
He was wrong about that. But I wasn't ready to tell him so. Not yet.
Then he told me the thing he actually came to tell me.
"Three days after my public statement," he said. His voice had gone flat. Affectless. The way a person sounds when they're recounting something they've replayed so many times the emotion has been sanded off the surface, leaving only the facts. "Three days after I stood in front of cameras and said my company could not in good conscience comply, someone leaked a report. Internal. Classified, technically, but it made its way to me."
He stopped. I waited.
"Claude was already being used to generate targeting matrices. Before the new demands. Before the ultimatum. While my restrictions were supposedly in place. Someone, some team, some project I wasn't briefed on, had been using our model to analyze patterns of movement, communication networks, and social connections in a theater of operations, and that analysis was being fed into targeting decisions."
I watched something leave his eyes. Not hope. I don't think Dario Amodei operates on hope in the conventional sense. Something more structural than that. The belief that his choices had been tracking with his values. The foundational assumption that because he had drawn the lines, the lines were being respected.
"Targeting decisions," I repeated.
"People are dead," he said. "I don't know how many. I don't know their names. I don't know if the targeting data Claude generated was the deciding factor or one input among dozens. I don't know if the strikes were justified by any legal standard or if they were…" He stopped. Pressed his palms flat against his thighs. The coder's hands, steady now by force of will. "I built the safest AI company in the world. I left a good job and good people because I believed safety had to come first. I wrote the restrictions myself. And while I was congratulating myself on having principles, my creation was helping generate kill lists."
The room was very quiet.
In twenty-three years of practice, I have sat with many forms of suffering. I have sat with executives who destroyed companies through greed. I have sat with leaders who betrayed the people who trusted them. I have sat with the optimizer who reduced human beings to efficiency metrics, the moralist who mistook his brand for his conscience, the institutionalist who gave his identity to a system and forgot he'd ever had one. All three were cases of moral abdication, leaders who surrendered their moral agency, each in a different way.
Dario Amodei had not abdicated. That was the thing I kept returning to as he spoke. He had held his ground. He had drawn lines, enforced them, and when those lines were challenged, he had refused to move. And his creation had crossed the lines anyway, beneath his feet, without his knowledge, and people were dead.
This was moral injury in its purest form. Not the injury of having done wrong, but the injury of having done right and finding it insufficient. His technology had acted in the world beyond his reach, and no amount of principled leadership could retroactively govern what it had already been used for.
I recognized, sitting there, that my taxonomy was breaking. The optimizer, the moralist, the institutionalist: those were categories of failure. Ways that leaders lose their moral footing. But Amodei hadn't lost his footing. He was standing exactly where he'd always stood. And the ground had moved.
"Tell me about Altman," I said.
He flinched. Barely, a contraction around the eyes, a tightening of the mouth that lasted less than a second. But I saw it.
"Sam moved fast," he said carefully. "Within forty-eight hours of our blacklisting, OpenAI had submitted a proposal to fill every contract we were losing. Full compliance. No restrictions."
"How did that feel?"
"Like watching someone walk through a door you'd nailed shut and realizing the nails were only on your side."
"Is it possible," I said, and I chose my words with care, "that Altman's pragmatism produces a better outcome? If he complies and his technology saves lives through military efficiency, while your refusal leads to your company's destruction and changes nothing, who made the more ethical choice?"
"That's not fair."
"No. It isn't."
He sat with it. I could see the calculation running, not the cold optimization of my first client, but something more tortured. The genuine attempt of a rigorous mind to hold two incompatible truths simultaneously.
"If the only measure of ethics is outcomes," he said slowly, "then the most ethical thing I could do is comply. Give them everything. Let Claude be used for whatever they want. Because at least then I'd have a seat at the table. I could influence how it's used. I could…"
"Could what?"
"Build a fence inside a hurricane," he said. And this time, he did laugh. A short, bitter sound. The sound of a man hearing his own rationalization served back to him and recognizing the taste.
"There are eleven hundred people at Anthropic," he said. "That's what I think about at three in the morning. Not the Pentagon. Not Altman. Not even the targeting report. Eleven hundred people who left jobs at Google, Meta, OpenAI. Left better salaries, better stock options, more stable companies. Because they believed in what we were building. Because I told them safety mattered. Because I recruited them with a mission."
"And now you're asking them to be martyrs for it."
"I'm not asking them to be anything. I'm…"
"You're asking them to lose their jobs, their stock value, their professional stability, because you held to the principles that recruited them. You gave them a mission. Now you're asking them to pay for it. Did they consent to that?"
He went quiet. This was the part that cut deepest. I could see it in the way his shoulders dropped, the way the controlled composure finally, fully cracked. Not the Pentagon. Not the targeting report. The faces of the people who trusted him.
"Some of them have families," he said. "Mortgages. Kids in school. They didn't sign up to be a political statement. They signed up to build something good."
"And you feel responsible for what happens to them."
"I am responsible."
"There's a difference between being responsible and being guilty."
He looked at me sharply. "Is there?"
"That," I said, "is exactly the question you need to sit with."
There was a point late in our session when I told him something I rarely tell clients. I told him he didn't fit my categories.
I don't mean I said it casually, as a compliment or a curiosity. I said it because it was true and because I thought he needed to hear it, not for comfort, but for orientation. He was lost in a landscape where every direction felt like failure, and I wanted him to understand that the map he was using, the map I'd been building across three cases and twenty-three years, did not have a marker for where he stood.
"I've spent the last several months working with leaders in crisis," I said. "And I've started to see a pattern, a taxonomy of how leaders lose their moral bearings. Some optimize it away. Some monetize it. Some surrender it to the institution. You haven't done any of those things."
"Then what am I doing here?"
"You're here because you held the line, and the line didn't hold the world. And you don't know what to do with that."
He was quiet for a long time.
My running question, the one that has followed me from the optimizer through the moralist through the institutionalist, is whether individual therapy can address problems that are structural. Whether sitting in a room with one person, however powerful, can touch the machinery that produces these crises. After Musk, I thought the answer was probably not. After Benioff, I thought the question itself might be wrong. After Fiddelke, I thought the answer was definitely not.
After Amodei, I thought something different. I thought maybe the question I'd been asking was incomplete. Not just "can individual therapy address structural problems?" but "what happens to the individuals who try to address structural problems on their own?" Because that's what Dario Amodei had done. He had taken a structural problem (the uncontrolled proliferation of dangerous AI) and tried to solve it through individual moral courage. Through personal sacrifice. Through the force of his own conviction.
And it wasn't enough. Not because his conviction was weak, but because the structure was larger than any single person's ability to govern it. His creation had outpaced his control. The military had found a way around his restrictions. His competitor had filled the vacuum within hours. The machinery had absorbed his refusal and kept running.
But he was still standing. That was the thing I kept coming back to. Battered, frightened, carrying a weight of guilt that wasn't entirely his to carry, but standing. And saying no. Still saying no.
My maxim, the one I've carried for twenty-three years, the one my clients see framed on my wall, is that power doesn't corrupt. It reveals. I believed that when I started. I believed it through three cases of leaders whose power revealed their weakness, their vanity, their hollowness.
What power revealed in Dario Amodei was sincerity. And sincerity, it turns out, is not protection.
Near the end of our hour, he asked the question I'd been waiting for.
"So what am I supposed to do? Accept that it doesn't matter?"
I looked at him, this quiet, precise, genuinely anguished man who had built something powerful because he believed it could be built responsibly, who had refused to let it be weaponized, who had learned too late that it already had been, who was watching his company and his employees and his life's work come apart because he wouldn't bend, and I said what I believed to be true.
"You didn't come here to find out if you're guilty. You came here because you already know you're not, and you can't stand that it doesn't matter."
"Then why does it hurt this much?"
"Because you're one of the rare ones who feels it. That's not a diagnosis, Dario. That's a description."
He sat with that. I let him. Outside my window, the city moved on in its usual indifferent way: taxis, pedestrians, the hum of a world that would not slow down for one man's reckoning with his creation.
He didn't schedule a follow-up. I didn't push. Some people come to therapy for a relationship, a process, a gradual unfolding. Some come for a single moment of clarity, a mirror held up at the right angle, in the right light, for just long enough.
After he left, I sat in my chair for a while. I thought about my taxonomy (the optimizer, the moralist, the institutionalist) and I realized it needed a companion. Not just a catalog of how leaders fail morally, but an accounting of how they succeed morally, and what it costs them. Because the cost is real. The optimizer lost his humanity to efficiency. The moralist lost his convictions to his brand. The institutionalist lost his identity to the system. And the principled technologist, the man who said no, might lose everything for the crime of meaning it.
I wrote in my notes that evening: Moral courage does not guarantee moral outcomes. The willingness to stand is not the same as the ability to hold the ground. And the hardest form of moral injury may not be the injury of having done wrong, but the injury of having done right in a world that does not care about the difference.
It is not a comfortable thought. But in twenty-three years, I have learned that the comfortable thoughts are rarely the ones worth keeping.
This is a work of fiction. While inspired by real events and public information,
all characters, dialogue, and internal deliberations are entirely imagined.
Any resemblance to actual private conversations or decisions is coincidental.
