The Statecraft Wing
When a lab does something that scares you, you look for the engineer. The statecraft wing is the answer to a different question, the one nobody asks at the demo, which is who talks to the government on the lab’s behalf, and what they used to do before the lab existed. Run those résumés and the door doesn’t just revolve. It revolves between the same few buildings, and the buildings have names like the White House and the National Security Council and the State Department.
This is the third of the four machines the pilot named, and the one most easily smudged into the second. The enforcement floor decides which prompts a citizen may type. The statecraft wing decides what the lab tells a senator, what it files with a regulator, and how it positions itself in a great-power contest. Those are different levers, held by different people, drawn from different pools. Collapsing them is the single mistake that makes this whole series refutable in one question, so draw the line first and draw it hard. The character shop writes the refusals. The enforcement floor runs the bans. This wing does not touch either. It governs the lab’s politics, not the model’s mouth.
The cleanest pipeline in the set
Start with the densest single résumé, because it is the template for the rest.
Chris Lehane ran global policy at OpenAI after a career that reads like a Democratic operative’s greatest hits. Lawyer in the Clinton White House counsel’s office through the Whitewater and Lewinsky years. Press secretary to Vice President Al Gore on the 2000 presidential campaign, where he earned the nickname “Master of Disaster” for crisis work. Head of global policy and public affairs at Airbnb from 2015 to 2022. Then OpenAI, as the executive who runs how the company talks to governments. Campaign war room, then home-sharing regulatory brawls, then the firm building artificial general intelligence. The skill is the same at every stop. You manage how power perceives your client.
What that role produces is on the record in OpenAI’s own words. In its submission to the White House Office of Science and Technology Policy on the national AI Action Plan, the company argues that “American-led AI built on democratic principles” must “continue to prevail over CCP-built autocratic, authoritarian AI,” and proposes helping allied countries “build AI on democratic rails.” That is not a safety document. It is a foreign-policy brief, filed by a company, written in the grammar of statecraft, and aimed at a government that was being asked to pick a side.
Anthropic staffs the wing from the same building
Across town the labels differ and the buildings on the résumé do not.
Michael Sellitto is Head of Global Affairs at Anthropic. Before that he was Director for Cybersecurity Policy on the National Security Council staff from 2015 to 2018, leading international engagement on cyber policy. Before that, Special Assistant to Deputy Secretaries of State William Burns and Antony Blinken. In between government and Anthropic, he was founding Deputy Director of Stanford’s Institute for Human-Centered AI, and he now sits on the board of the Frontier Model Forum, the body the frontier labs founded to speak for themselves. The NSC, then State, then the institute that certifies the field, then the company, then the board that represents the companies. One person, the whole circuit.
Sarah Heck leads public policy at Anthropic. Before Anthropic she was Head of Entrepreneurship at Stripe, and before Stripe she “led global entrepreneurship and public diplomacy policy at the White House National Security Council,” in the company’s own description. The pattern doesn’t need embellishment. The NSC keeps appearing in the work history of the people who decide what the labs say to the state, which is either the most natural hire in the world or the quiet part, depending on which side of the door you’re standing on.
Then there is the man who runs Anthropic’s policy outright. Jack Clark co-founded the company and, per Anthropic’s own announcement, moves to the role of Head of Public Benefit while having led policy throughout. He testified in June 2025 before the House Select Committee on the Chinese Communist Party, at a hearing titled “Algorithms and Authoritarians.” His written testimony runs a section headed “Why We Should Win—Democratic versus Authoritarian AI,” and argues that AI built by authoritarian nations “will—no matter what the personal preferences are of the people in those countries building it—be inescapably intertwined and imbued with authoritarianism.” A lab co-founder, in a congressional hearing room, making the case that the technology itself carries the politics of the regime that builds it.
Google revised the principles, not the prose
Google did the same thing at the level of doctrine, and did it in public.
In February 2025 the company revised its AI Principles and removed the explicit pledges, in place since 2018, not to build weapons or surveillance technology that violated accepted norms. The 2018 language had been written after employee protests over Project Maven, the Pentagon drone-analysis contract. The removal was reported at the time by CNBC and the Washington Post. In the blog post announcing the change, DeepMind’s Demis Hassabis and senior research executive James Manyika wrote that “democracies should lead in AI development, guided by core values like freedom, equality, and respect for human rights.” Same frame, third lab, independently arrived at. The pledge against weapons came out the same week the sentence about democracies went in.
The Meta caveat, stated plainly
Meta’s wing is run by Joel Kaplan, Chief Global Affairs Officer, formerly Deputy Chief of Staff in the George W. Bush White House. He is the obvious next exhibit, and this is exactly where the lazy version of the story overreaches, so be careful.
In January 2025 Meta announced “More Speech, Fewer Mistakes,” ending its third-party fact-checking program in favor of a Community Notes model. That is a change to how Facebook and Instagram moderate posts. It is a platform-content decision. It is not a stated change to Llama or to Meta’s AI model policy, and nobody has shown that it is. The credential is real. The Bush-White-House-to-tech-statecraft path is the same shape as the others. But the receipt in hand documents a platform-moderation change, so that is all this dispatch will say it documents. Tying it to the behavior of the models would be the thing this series exists to refuse: a guess wearing a citation’s clothes.
The steel-man, and why it holds anyway
Take the strongest version of the defense, because it is a good one.
A frontier lab operates in a world where governments write the rules that decide whether it lives. Of course it hires people who know how Washington works. You would not staff your government-affairs shop with engineers any more than you would staff your model team with lobbyists. The NSC and State Department alumni are there because the job is genuinely about national security and international policy, and those are the people who actually understand it. And the democratic-AI framing is not paranoia. There really is a great-power contest in this technology, the CCP really is racing, and a company that builds AGI in an open society has a defensible reason to say so out loud. On its own terms, the statecraft wing is just competent government relations.
All of that is true, and none of it dissolves the observation. The steel-man explains why the wing is staffed by former officials. It does not explain away what follows from it. The people who spent a decade inside the national-security state, who learned to see technology as terrain in a contest between regimes, now sit at the firms building the most powerful information technology yet made, and they carry that lens with them. They did not stop thinking like the government when they left it. They started thinking like the government on behalf of a company. The frame came along for the ride, and the frame is now the house position at three separate labs.
Not a cabal. A current.
Be precise, because the comforting version is the wrong one.
Nobody wired this. No room exists where Lehane, Sellitto, Heck, Clark, and Manyika agree on a script. They did not coordinate the democratic-AI language. They converged on it, which is stranger and harder to fight, because convergence has no off switch. Three labs reached the same geopolitical frame independently, from staffs drawn out of the same handful of agencies, and they reached it because the people writing the lines came up in the institutions where that frame is the water. A conspiracy you can expose. A talent pipeline you cannot, because there is nothing hidden to find. The LinkedIn profiles are public. The testimony is published. The RFI submissions are posted on the company’s own CDN.
And keep the wall between this machine and the last one standing, because it is load-bearing. None of these people write the refusals. None of them run the bans. Sellitto does not decide whether your prompt gets flagged; Lehane does not decide what the model will say about a chemistry question. They decide what the lab tells the state, and what the lab asks the state to do, which is a different and arguably larger power, and a separate one. The character shop governs the mind. The enforcement floor governs the account. The statecraft wing governs the relationship between the company and the government that can regulate it out of existence or anoint it the national champion. Two truths, side by side. Both real. Both sourced. Different levers, different hands.
The government did not have to capture the labs. It did not raid them, subpoena them, or plant anyone. It simply watched its own alumni walk through the front door, take the corner offices marked Global Affairs, and start writing the company’s foreign policy on a business card that still, in spirit, says public service.
Watch the watchers. This wing posted its testimony, filed its submissions, and listed its officials by name. Read the org chart.