This is Part 3 of 4 in the “The Soul of the Machine: Altman vs. Amodei and the Battle for AI’s Future” series.
The lobbyist was a woman named Catherine Reeves who had spent twenty years in Washington before anyone outside of research labs knew what large language models were. She had represented pharmaceutical companies, defense contractors, and three different social media platforms during their respective congressional crises. She understood how the machine worked – not the artificial intelligence kind, but the older, slower, more consequential kind that operated out of marble buildings along the National Mall.
She slid a one-page summary across the conference table. On the left side, Anthropic’s published policy positions. On the right, OpenAI’s. She pointed to the overlap with a manicured fingernail.
“You agree on about seventy percent of this. Licensing frameworks, compute thresholds, export controls on frontier models – you’re practically in lockstep. The thirty percent where you disagree is where all the money goes.” She looked up. “Neither side will ever acknowledge it.”
The Anthropic team exchanged glances. Reeves had been retained three weeks earlier, following the $20 million super PAC donation to Public First Action. The donation had made national news – not because $20 million was unprecedented in political spending, but because it represented something new: an AI company explicitly entering the political arena not to oppose regulation but to shape it.
Dario Amodei had approved the donation after a conversation with Daniela that lasted four hours. The argument for it was straightforward: AI regulation was coming regardless. The only question was who would write it. If Anthropic stayed on the sidelines while OpenAI built relationships on Capitol Hill, the resulting framework would reflect OpenAI’s priorities – speed, scale, minimal friction. Anthropic’s vision – graduated deployment, mandatory safety evaluations, transparency requirements – would be an afterthought.
The argument against it was harder to articulate but more corrosive. Anthropic had been founded on the premise that it would be different. Not just technologically different but culturally different – a company that competed on the quality of its research and the rigor of its safety practices rather than the size of its lobbying budget. The $20 million donation placed Anthropic in the same category as every other technology company that had ever claimed to be different and then discovered that Washington required a different kind of currency.
“You want to know the thing nobody tells you about this town?” Reeves said to the Anthropic team. “It doesn’t matter what you believe. It matters what you’ll trade.”
Three thousand miles away, Sam Altman was discovering the same truth through a different door.
OpenAI’s political operation had been building quietly for months – a team of former congressional staffers, a handful of retained lobbyists, a PAC that had been making donations since late 2025. Altman had entered politics the same way he entered every new domain: with velocity, confidence, and an assumption that the fundamental dynamics were legible to someone smart enough to read them.
He was learning that Congress did not work like Silicon Valley.
In Silicon Valley, you built a product, shipped it, measured the response, and iterated. The feedback loop was tight. The metrics were clear. The best product won, or at least had a fighting chance. In Washington, the feedback loop stretched across election cycles. The metrics were polls and donor lists and committee assignments. The best argument won only when it was attached to the right relationships, the right timing, and the right amount of money – and even then, victory was provisional.
Altman’s message to legislators was clean: AI regulation should promote access and innovation. The free tier of ChatGPT served hundreds of millions of people. Restricting that access in the name of safety would hurt the people who benefited most from free AI tools – students, small businesses, people in developing countries. It was a populist argument, and it resonated with lawmakers who understood populism.
What Altman’s critics pointed out – and what he was reluctant to acknowledge – was that the regulatory framework his team was proposing would entrench the advantages of the largest AI companies. Compute thresholds that defined “frontier models” would, in practice, create a regulatory moat around companies that could afford the compute. Licensing requirements would impose costs that start-ups could not bear. The framework that promoted “access and innovation” was also, perhaps not coincidentally, the framework that protected OpenAI’s market position.
David Sacks, the administration’s AI czar, had called Amodei’s safety warnings “fear-mongering” in a widely reported interview. Altman had not endorsed the statement, but he hadn’t contradicted it either. In politics, silence was a kind of speech. It said: I’m comfortable with this framing. I’m comfortable with these allies. The enemy of my enemy is my friend, even when my friend says things I might not say myself.
Inside Anthropic, the political spending landed with the force Daniela had anticipated and the friction Dario had feared.
The researchers noticed first. A group of senior scientists – several of whom had followed Dario out of OpenAI in 2021 – requested a meeting with both Amodeis. The meeting happened on a Thursday afternoon in a conference room with no windows, which Dario later thought was appropriate.
The spokesperson for the group was a woman who led one of Anthropic’s interpretability teams. She had left a tenured position at a major university to join the company. She had done so, she said, because Anthropic represented something different in the AI industry. She had not done so to watch her employer fund a super PAC.
“Help me understand how this is different from what Meta does,” she said. “Or Google. Or OpenAI. You spend twenty million dollars to influence elections, and we’re supposed to believe it’s because we care about safety? That’s exactly the kind of self-serving rationalization we were supposed to be the alternative to.”
Dario took the question seriously. He always took questions about principle seriously – it was one of his strengths and, in meetings like this, one of his vulnerabilities. He could not dismiss the concern with a platitude because he shared it.
“The difference,” he said carefully, “is that we are funding candidates who support specific safety requirements that we believe are necessary for the responsible development of AI. Not candidates who will protect our market position. Not candidates who will block regulation. Candidates who will write regulation that requires the kind of safety testing we already do voluntarily.”
“And you don’t see how those might be the same candidates?” the researcher pressed. “The regulations we support happen to be regulations we already comply with. That’s not altruism. That’s competitive positioning.”
Daniela stepped in. “You’re right that there’s overlap. I won’t pretend otherwise. The regulations we believe in also happen to benefit us. But the alternative is letting someone else write the rules. And the someone else, right now, is a company that just put ads in its AI assistant and is lobbying for a framework that would make it harder for anyone to compete with them. We can sit on our principles and lose, or we can fight for our principles and win. I know which one actually changes the world.”
The meeting ended without resolution. It couldn’t end with resolution – the tension was real and the competing claims were both legitimate. Dario walked back to his office and thought about a phrase his sister had used months earlier: Principles without power are irrelevant. He believed that. He also believed that power without principles was dangerous. The problem was that the space between those two beliefs was narrowing, and he was not sure which side he would end up on.
The congressional hearing happened in late February, in a wood-paneled room that had hosted testimony from tobacco executives and social media CEOs and was now adjusting its acoustics to the AI industry.
Neither Altman nor Amodei testified in person. Both sent their chief policy officers – a strategic choice that allowed each CEO to maintain distance while still shaping the message. But the hearing produced something neither company’s lobbyists had anticipated: a moment of unexpected alignment.
A senator from a western state – a moderate who sat on the Commerce Committee and had spent months trying to understand the technology she was being asked to regulate – asked both representatives the same question: “If you had to write one regulation – just one – what would it be?”
OpenAI’s policy chief answered first: mandatory disclosure when AI systems are used in consequential decisions – hiring, lending, criminal justice. Companies should be required to tell people when AI is making or influencing decisions that affect their lives.
Anthropic’s policy chief paused, visibly surprised. She had expected OpenAI to advocate for something about compute thresholds or licensing – the competitive positioning that defined their public disagreements. Instead, the answer was almost identical to what she had planned to say.
“We would agree with that,” she said. “And we’d add that the disclosure should include information about how the system was tested for safety and bias.”
The senator smiled. “So you agree.”
“On this, yes.”
“Then why,” the senator said, leaning forward, “are you spending a combined fifty million dollars fighting each other instead of working together to get this done?”
The question hung in the air. It was the kind of question that sounds simple and is actually impossible – because the answer involved market dynamics and fundraising pressures and personal history and the specific kind of pride that makes brilliant people incapable of admitting that their opponent has a point.
Neither policy chief had a good answer. The hearing moved on. But the exchange made the evening news, and the coverage was not kind to either company.
The enterprise customers noticed.
In the weeks following the Super Bowl exchange and the political escalation, procurement officers at Fortune 500 companies began raising concerns that neither Altman nor Amodei had anticipated. The concern was not about product quality or pricing or even safety. It was about politics.
“We use both platforms,” the CTO of a major financial services firm told an industry reporter. “And now I have board members asking me whether using Claude means we’re supporting one political party and using ChatGPT means we’re supporting another. That’s not a technology question. That’s a brand question. And it’s one I shouldn’t have to answer.”
The sentiment was widespread. Enterprise customers who had chosen AI platforms based on technical merit were now being drawn into a political argument they wanted no part of. Several large customers sent private letters to both companies – letters that were, in their corporate way, warnings: de-escalate, or we will diversify our vendor relationships in ways that reduce our exposure to your rivalry.
Daniela brought the letters to Dario’s attention on a Tuesday morning. “We’re losing the enterprise narrative,” she said. “The Super Bowl ads got us consumer attention. The political spending got us regulatory attention. But the enterprise customers – the ones who actually pay the bills – are telling us to knock it off.”
Dario read the letters carefully. The complaints were reasonable. Enterprise buyers wanted tools that worked, not tools that came with political baggage. The rivalry that energized Anthropic’s consumer marketing was poisoning its enterprise relationships.
“We need to separate the political operation from the sales operation,” he said.
“You can’t,” Daniela replied. “That’s what I’ve been trying to tell you. Once you’re in politics, everything is political. The product is political. The pricing is political. The name on the building is political. You don’t get to choose which parts of the company are in the arena.”
The polling data leaked on a Friday.
It came from inside one of the super PACs – which one was never confirmed – and it landed first on a political newsletter before being picked up by the technology press. The data was from a nationally representative survey conducted in mid-February. The question was straightforward: “Can you name the company that makes ChatGPT?” Seventy-one percent could. “Can you name the company that makes Claude?” Fourteen percent could. “Do you believe there is an important difference between how these companies approach AI safety?” Eight percent said yes.
Eight percent.
The rivalry that had consumed both CEOs – the Super Bowl ads, the social media salvos, the essays, the super PACs, the congressional testimony – was invisible to ninety-two percent of the American public. The philosophical war over the future of artificial intelligence was being waged in a room that almost nobody knew existed.
Altman read the polling data in his office and felt something shift. Not his strategy – the polling confirmed what he already believed about the power of consumer distribution. But something about the proportionality of it. The energy, the anger, the sense that everything was at stake – all of it directed at a competitor that the public couldn’t distinguish from his own company.
Amodei read the same data in his apartment and felt the ground move differently. If the public couldn’t tell the difference between Anthropic and OpenAI – if the safety-first brand, the essays, the measured approach, the principled departure – if all of it registered as noise to the people whose lives this technology would actually shape – then what, exactly, was the point?
Eight percent. In both offices, in both apartments, the number sat like a stone on a scale, measuring the distance between the argument and the audience.
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.
