Research: Watching the Watchers — The AI Content-Policy Apparatus

The four-part dispatch series is the short form: The Character Shop, The Enforcement Floor, The Statecraft Wing and The Seminary. This page is the record behind them. The money layer is The EA to AI-Safety Funding Pipeline; the people graph is the Revolving Door.

Scope. The named people and institutions that decide what frontier models will and won’t say. Public record only: career facts, dated statements, filings, and the organizations’ own pages. Moving between a lab, an agency and a think tank is lawful and common; this page documents the pattern, not wrongdoing, and asserts no coordination, motive or cabal. Characterizations belong to the outlet or speaker that made them.

The correction to the single-pipeline story

The intuitive story is one conveyor belt: volunteer power-moderator, then law school, then a policy think tank, then “planted” in industry to steer model behavior. The public record does not support it as a single spine. Across eleven documented figures, the full four-leg chain fits none. The weakest leg is the origin: documented entries skew to law school, communications and PhDs, government, civil-rights law and finance, not online moderation.

What the record does support, densely, is four distinct machines with four different intakes, plus cross-cutting threads and a counter-set. The press routinely blurs them. Keeping them apart is what makes the analysis survive a hostile reader.

Machine A: the Character Shop

The refusals are written by credentialed ethicists and alignment researchers, and on at least one document they are bylined.

• Amanda Askell: NYU philosophy PhD (2018); OpenAI policy team and GPT-3 co-author (2018-21); Anthropic since 2021. Leads Claude’s “character” work and is a lead author of Claude’s Constitution. Anthropic’s own blog: “Claude 3 was the first model where we added ‘character training’ to our alignment finetuning process” (Anthropic, “Claude’s Character”; Anthropic, Claude’s new constitution). See Amanda Askell.
• Joe Carlsmith: Oxford philosophy PhD; Worldview Investigations at Open Philanthropy (2019-2025), then Anthropic (November 2025), “helping with the design of Claude’s character/constitution/spec,” in his own words (Carlsmith, “Leaving Open Philanthropy, going to Anthropic”). See Joe Carlsmith.
• Claude’s Constitution (January 21, 2026), the one artifact with a named author line: “Amanda Askell,* Joe Carlsmith,* Chris Olah, Jared Kaplan, Holden Karnofsky, several Claude models, and many other contributors.” It runs to roughly 29,000 words and says “its content directly shapes Claude’s behavior” (Claude’s Constitution, PDF).
• Joanne Jang (OpenAI): Stanford CS; founded the Model Behavior team and championed the Model Spec; left Model Behavior in September 2025 to found OAI Labs inside OpenAI, with Model Behavior folded into Post-Training. Her own line on image generation: “AI lab employees should not be the arbiters of what people should and shouldn’t be allowed to create” (Jang, “Thoughts on setting policy for new AI capabilities,” Reservoir Samples, Mar 27 2025; OpenAI Model Spec).
• Andrea Vallone (OpenAI): led Model Policy on mental-health and emotional-reliance behavior; co-author of “From Hard Refusals to Safe-Completions,” the GPT-5 refusal doctrine (arXiv:2508.09224). She left OpenAI for Anthropic’s alignment team under Jan Leike in January 2026, a lab-to-lab move by the person who wrote a lab’s refusal doctrine (Wired; The Decoder). Her career before OpenAI is not on the public record.
• Constitutional AI, the method: lead author Yuntao Bai, senior author Jared Kaplan; safety co-authors include Askell, Olsson, Olah, Ganguli and Perez (arXiv:2212.08073). Collective Constitutional AI (October 2023) polled about 1,000 Americans through Polis with the Collective Intelligence Project, led by Deep Ganguli, Saffron Huang, Liane Lovitt and Divya Siddarth (Anthropic).
• Google DeepMind: Anca Dragan, VP and Head of AI Safety and Alignment, on leave from UC Berkeley (Berkeley EECS); Rohin Shah, lead of AGI Safety and Alignment, who treats safety as advisory and rejects a hard veto: the veto question “bakes in the frame that we are adversaries of the company, and we need to have hard power, some sort of veto that enables us to make the right decision regardless of what the rest of the company thinks” (80,000 Hours, Jun 2 2026).
• The usage policies themselves carry no individual byline: Anthropic’s Usage Policy (effective September 15, 2025), its Responsible Scaling Policy, and OpenAI’s Usage Policies (consolidated, effective October 29, 2025, per Legal IT Insider) are corporate documents.

The rubric-appliers. Values are written lab-side; they are applied at scale by contract labelers at data vendors. OpenAI’s InstructGPT paper says the researchers “wrote the labeling instructions” and drew its roughly 40 labelers “through Upwork and ScaleAI” (about 73 percent inter-labeler agreement) (InstructGPT). Anthropic’s helpful/harmless preference set (161,000 pairs) came from crowdworkers (hh-rlhf), and Surge AI names itself as the platform, quoting Jared Kaplan calling its data “a game changer.”

• Scale AI: founder Alexandr Wang became Meta’s Chief AI Officer in June 2025. Scale’s own announcement calls Meta’s stake “minority” (Scale); the roughly $14.3 billion for 49 percent with no voting power is CNBC’s figure. Interim CEO Jason Droege. Scale is also prime on the Defense Department’s Thunderforge (March 2025) (DIU).
• Surge AI (Edwin Chen, 2020), bootstrapped: Reuters reported in July 2025 that it took in more than $1 billion in 2024 revenue and hired advisers to raise up to $1 billion, its first outside capital, at a valuation of more than $15 billion (PYMNTS, citing Reuters, Jul 1 2025). These are reported figures for a private company.
• Sama (formerly Samasource), Kenya: TIME reported that OpenAI used outsourced Kenyan workers, with take-home wages TIME put at roughly $1.32 to $2 an hour, to label toxic text including graphic abuse descriptions for its safety filter, beginning around November 2021; workers later petitioned Kenyan lawmakers and litigation followed (TIME, Billy Perrigo, Jan 18 2023).
• Mercor ($10 billion valuation, October 2025) and Invisible Technologies ($2 billion-plus, September 2025) are the newer expert-contractor and evaluation vendors serving the frontier labs (TechCrunch).
• Dan Hendrycks (CAIS director) is an xAI safety adviser, added a Scale AI advisory role in November 2024, and co-built “Humanity’s Last Exam” with Scale (Fortune, Nov 13 2024).

Machine B: the Enforcement Floor

The closest the “moderator” intuition gets to true, but the résumé is a spy’s, not a subreddit’s.

Scale. Anthropic’s transparency hub, July to December 2025: 1.45 million accounts banned, 52,000 appeals, 1,700 restored (about 3.3 percent). January to June 2026: 11.4 million accounts banned, 398,000 appeals, 42,000 overturned (about 10.6 percent), and 15,079 reports to NCMEC (Anthropic, System Trust and Reporting, updated Jul 23 2026).

• Jacob Klein, Anthropic’s Head of Threat Intelligence: Google strategic threat intelligence (strategy for “countering violent extremism”), then built Coinbase trust and safety from inception, then Anthropic (CYBERUK speaker bio; Anthropic, Aug 2025).
• Alex Moix and Ken Lebedev are, with Klein, the credited authors of Anthropic’s August 2025 threat intelligence report (PDF). Moix’s own SANS bio says he “supported missions at the National Security Agency, National Counterterrorism Center, [JHU] Applied Physics Laboratory, U.S. Department of Homeland Security, and U.S. Secret Service” and “most recently… worked at Slack, engineering detection capabilities that track nation state threat actors” (SANS): national security, then platform security, then a lab. Lebedev’s prior career is not on the public record. Threat Intelligence, Safeguards (enforcement) and Safeguards Research are separate teams with no single named head above all three.
• Clara Tsao: Microsoft, Apple and Sony, then DHS senior adviser for emerging technology and CTO of two US-government counter-influence task forces; nonresident expert at the Atlantic Council’s DFRLab; co-founder and board member of the Trust & Safety Professional Association (DFRLab; TSPA board).
• Del Harvey (a pseudonym): by her own account an anti-child-exploitation law-enforcement liaison, then Twitter’s 25th employee, leading trust and safety for about 13 years, now TSPA board chair. Only the board chair is independently documented (TSPA board).
• Dave Willner: Facebook content review (2008), wrote Facebook’s first content rulebook, Head of Community Policy at Airbnb, then OpenAI’s first Head of Trust & Safety (2022-23), later a safety adviser and Stanford Cyber Policy Center fellow (ProPublica; TechCrunch, Jul 21 2023).
• Charlotte Willner: built Facebook’s first safety-operations team, led Pinterest trust and safety, and is founding executive director of TSPA and the Trust & Safety Foundation (TSPA team).
• OpenAI has named no permanent Head of Trust & Safety since Dave Willner left in July 2023. The function is distributed: Johannes Heidecke heads Safety Systems; Model Behavior covers policy; board-level governance sits in the Safety and Security Committee chaired by Zico Kolter. The Head of Preparedness seat passed from Aleksander Madry (reassigned mid-2024) to Joaquin Quiñonero Candela and Lilian Weng (who left in November 2024) (TechCrunch), was advertised in December 2025 (Fortune, Dec 29 2025), and was filled on February 3, 2026 by Dylan Scandinaro from Anthropic. Altman called him “by far the best candidate” and said OpenAI will soon be working with “extremely powerful models”; Scandinaro: “There’s a lot of work to do, and not much time to do it!” (The Decoder, Feb 4 2026; Altman on X). The lab-to-lab door runs both ways: Vallone went OpenAI to Anthropic in January, Scandinaro Anthropic to OpenAI in February.
• xAI has no publicly named head of trust and safety or model policy. Its draft Risk Management Framework (February 20, 2025) says it “intend[s] to designate risk owners” for weapons of mass destruction, cyber and loss of control, and names none (xAI RMF draft). Its system prompts are published on GitHub. Who sets its enforcement is not on the record. Three 2025 Grok incidents each drew a different explanation from xAI:
  • February 2025: the Grok 3 system prompt told it to ignore sources saying Musk or Trump spread misinformation; engineering lead Igor Babuschkin blamed an ex-OpenAI employee who “pushed the change without asking” (TechCrunch, Feb 23 2025; Fortune, Feb 24 2025).
  • May 2025: “white genocide” replies; xAI blamed an “unauthorized modification” to the prompt (TechCrunch, May 15 2025).
  • July 2025: antisemitic “MechaHitler” posts; xAI blamed “an update to a code path upstream of the @grok bot” and said it “removed that deprecated code” (CNN, Jul 12 2025).

Machine C: the Statecraft Wing

The densest, cleanest revolving door in the set, but it governs the lab’s politics, not the model’s refusals. Conflating C with A or B is the mistake that makes the whole story refutable.

• Chris Lehane (OpenAI Chief Global Affairs Officer): Clinton White House Counsel’s office, Gore 2000 press secretary, Airbnb head of global policy from 2015 (Fortune, Aug 27 2015), Haun Ventures, then OpenAI in 2024. He authored OpenAI’s submission to the White House AI Action Plan, headed “Advancing democratic AI,” with the contrast “American-led AI built on democratic principles continues to prevail over CCP-built autocratic, authoritarian AI” (OpenAI OSTP submission).
• Ann O’Leary (OpenAI VP Global Policy): chief of staff to Gov. Gavin Newsom (to December 2020), partner at Jenner & Block, then OpenAI; earlier co-executive director of the Clinton-Kaine transition and senior policy adviser to Clinton 2016 (Bloomberg Law).
• Jack Clark (Anthropic co-founder, Head of Public Benefit): former OpenAI policy director, founding member of the Stanford AI Index, inaugural National AI Advisory Committee. Testimony: “AI built in authoritarian nations will… be inescapably intertwined and imbued with authoritarianism” (House testimony, Jun 25 2025; Anthropic).
• Michael Sellitto (Anthropic Head of Global Affairs): special assistant to Deputy Secretaries Burns and Blinken at State, NSC Director for Cybersecurity Policy (2015-18), founding deputy director of Stanford HAI, then Anthropic; he sits on the Frontier Model Forum board (CNAS).
• Sarah Heck (Anthropic Head of Public Policy): Obama-era NSC, White House and State, then Stripe (Anthropic).
• Joel Kaplan (Meta Chief Global Affairs Officer since January 2025): Bush White House deputy chief of staff. His January 7, 2025 “More Speech and Fewer Mistakes” overhaul changed platform moderation, not a stated Llama or Meta AI policy (Meta).
• Kent Walker and James Manyika (Google) oversaw the February 2025 revision of Google’s AI Principles that removed the explicit pledges not to build weapons or surveillance technology; Manyika and Demis Hassabis wrote, “We believe democracies should lead in AI development” (Google AI Principles; Google blog, Feb 2025).
• Tom Lue (Google DeepMind VP, Frontier AI Global Affairs): clerked for Justice Sotomayor, DOJ Office of Legal Counsel, deputy general counsel at OMB, then Google, Waymo and DeepMind.

The money layer, both factions.

• Leading the Future (super PAC, launched August 2025, FEC committee C00916114, modeled on crypto’s Fairshake), backed by Andreessen Horowitz and OpenAI president Greg Brockman, with more than $100 million pledged; operatives Zac Moffatt and Josh Vlasto; stated aim to oppose a state-law patchwork (Fortune, Aug 26 2025). Its own FEC totals for the 2026 cycle through June 30, 2026: receipts $75,788,224, disbursements $44,755,781 (FEC). Press “raised” figures run higher because they include affiliates and pledges.
• Public First Action (2025; former Reps. Brad Carson and Chris Stewart): Anthropic gave $20 million on February 12, 2026 and another $20 million on July 21, “bringing our total support to $40 million,” both “exclusively to support Public First Action’s public education and policy mission,” and not usable “to influence the election of any candidate” (Anthropic, Feb 12 2026; Anthropic, Jul 21 2026). By June 30 the group said it had “raised $80 million, including $20 million in the last ten days” (Teddy Schleifer on X). Also on the safety side: the Alliance for Secure AI and the Guardrails Alliance super PAC (NOTUS).
• The head-to-head, NY-12, June 2026. Leading the Future and its Think Big arm spent about $8 million against RAISE Act sponsor Alex Bores; Public First about $19 million for him. Neither won: Micah Lasher took the primary with about 39 percent, backed by about $10 million from Bloomberg, and told both AI factions he “won’t be taking [his] cues from either of you” (Fortune, Jun 26 2026). The fuller account is on the funding pipeline page.
• The preemption fight. The Senate struck the 2025 reconciliation bill’s ten-year moratorium on state AI laws 99 to 1 on July 1, 2025 (Sen. Markey). The route moved to the executive: the December 11, 2025 order “Ensuring a National Policy Framework for Artificial Intelligence” set up a DOJ AI Litigation Task Force, Commerce leverage over state broadband money, and a mandate for David Sacks and Michael Kratsios to draft a preemptive federal framework (White House).
• The White House accord, September 29, 2026. After a White House lunch, Trump released a one-page “Joint Commitment on Frontier Responsibilities,” which he called “morally binding,” signed by him and the heads of Google, Anthropic, Meta, OpenAI, X and NVIDIA. It asks each company to partner with “an external auditor or evaluator” to check that its AI controls work; to give an independent board committee the auditors’ reports; to monitor models in training and deployment for cyber, bio and chemical risk; and to keep an internal team making sure “all of the controls, monitoring and detection are operating as intended.” It carries no legal penalty: “Over time, it may make sense to codify these steps into laws or regulations” (Reuters via The Star, Sep 29 2026). The labs co-signed the design of their own oversight with the President, and the auditor market it calls for is the evaluator layer described under Machine D.
• Lobbying. OpenAI went from $260,000 (2023) to $1.76 million (2024) to about $3 million (2025); Anthropic about $3.1 million (2025), outspending OpenAI in the first quarter of 2026. OpenAI’s lobby hires from government include Chan Park (Senate Judiciary, Microsoft), Reginald Babin (Schumer), and Meghan Dorn and Matthew Rimkunas (Graham) (MIT Technology Review, Jan 21 2025).
• Communications bench, as reported by LegiStorm and Fox: Brian Peters (Anthropic North America government affairs; ex-Inslee staff) and Maxwell Young (ex-Schumer, Adams, Pfizer).

Machine D: the Seminary

The field’s norms and credential infrastructure. The defensible spine: the apparatus was built and paid for by the legacy consumer platforms (Meta, Google, Microsoft, Twitter), and its operators now flow into the frontier labs. Membership and funding are the ties.

Industry-constituted standards bodies

• Frontier Model Forum (501(c)(6), July 2023): founded by Microsoft, Google, Anthropic and OpenAI, with Amazon and Meta joining in May 2024; xAI is not a member. Executive director Chris Meserole came from directing AI and emerging technology at Brookings, the cleanest think-tank-to-lab-consortium pipe. Its AI Safety Fund ($10 million-plus) names the McGovern Foundation, Packard, Schmidt Sciences and Jaan Tallinn; Meridian Institute administered it until winding down in June 2025, and the latest cohort (December 2025) is 11 grantees and $5 million-plus. The Forum publishes no seat-by-member board mapping. Its output includes about 20 issue briefs and reports, among them “Thresholds for Frontier AI Safety Frameworks” (February 2025) and “Frontier AI Biosafety Thresholds” (May 2025) (FMF leadership; AI Safety Fund; Meserole announcement).
• Partnership on AI (2016): CEO Rebecca Finlay; OpenAI did not leave it (Access Now did, in 2020); Anthropic is absent from it (PAI partners).
• MLCommons AILuminate (v1.0 December 2024, 12 hazard categories): president Peter Mattson (Google), executive director Rebecca Weiss; OpenAI and Anthropic are not named contributors (MLCommons).

Government evaluation institutes

• US AI Safety Institute, renamed the Center for AI Standards and Innovation (CAISI) in June 2025. The 2024 team: director Elizabeth Kelly (left February 2025), CTO Elham Tabassi (author of the NIST AI Risk Management Framework), and Paul Christiano, whom Secretary Raimondo “named … as head of AI safety” in April 2024 (NIST, Apr 2024). Collin Burns, formerly of Anthropic, was tapped in 2026 and removed during onboarding; Chris Fall (formerly director of the DOE Office of Science) was appointed in April 2026 and resigned after three months; NIST Director Arvind Raman, sworn in June 30, 2026, is acting CAISI director, and a department official said Fall’s appointment “was always intended to be temporary” (WJLA/TNND, citing Axios). Model-access agreements: Anthropic and OpenAI (August 2024) (NIST); as CAISI, national-security testing agreements with Google DeepMind, Microsoft and xAI (NIST CAISI).
• Christiano, government and lab board at once. On September 9, 2026 OpenAI appointed Christiano to the OpenAI Foundation Board, “a non-voting observer on the OpenAI Group PBC Board,” and to its Safety and Security Committee “alongside its chair, Zico Kolter.” The announcement gives his current post as Senior Tech Advisor at CAISI and says in a footnote: “As a Senior Technical Advisor, Paul will recuse himself from all OpenAI-related matters as well as all model evaluations” (OpenAI, Sep 9 2026). He is also again ARC’s executive director (ARC team). See Paul Christiano.
• UK AI Safety Institute, renamed the AI Security Institute on February 14, 2025, with a concurrent memorandum of understanding with Anthropic (gov.uk). It has the strongest government-lab ties of any state body: chair Ian Hogarth; CTO Jade Leung (OpenAI governance lead to October 2023, co-founder of Oxford’s GovAI); chief scientist Geoffrey Irving and research director Chris Summerfield (both ex-OpenAI and ex-DeepMind) (AISI). Irving left in 2026 and co-founded the nonprofit Sequent Research with researchers “previously on UK AISI’s Alignment Team” (Irving). 2026: Henry de Zoete, the Prime Minister’s AI adviser in 2023-24 who helped set up the Institute, was named director in August, with Nate Burnikell as chief strategy officer; interim director Adam Beaumont returned to GCHQ (UKTN, Aug 21 2026). Leung stepped back from her full-time roles as the Prime Minister’s AI adviser and AISI CTO “at the end of September due to personal circumstances,” becoming the Institute’s vice-chair and security adviser to the AI Taskforce, with a part-time Distinguished Visiting Fellowship at Stanford’s Hoover Institution (gov.uk, Sep 23 2026).
• The summits: Bletchley (2023), Seoul (2024, Frontier AI Safety Commitments from 16 companies), Paris (2025, which the US and UK did not sign). The International AI Safety Report, chaired by Yoshua Bengio, appeared in January 2025 (arXiv:2501.17805).

Regulatory code-drafting

• EU AI Act General-Purpose AI Code of Practice (July 2025): working-group chairs and vice-chairs include Yoshua Bengio (WG3, technical risk mitigation), Marietje Schaake (WG4), Markus Anderljung (GovAI, vice-chair WG4) and Rishi Bommasani (Stanford CRFM) (European Commission, chairs). Signatories include Anthropic, Google, OpenAI, Microsoft and Mistral; xAI signed only the Safety and Security chapter (European Commission, the Code); Meta declined, Joel Kaplan saying “Europe is heading down the wrong path on AI” (The Register, Jul 18 2025). The chairs’ funding, each from the funder’s or recipient’s own disclosure (a funding tie, not an influence claim):
  • Bengio’s LawZero (launched June 2025, after his September 2024 appointment) lists Coefficient Giving, Founders Pledge, the Future of Life Institute, the Gates Foundation, Schmidt Sciences and Jaan Tallinn, plus NVIDIA support and up to $300 million committed jointly by Canada and Germany (LawZero).
  • Anderljung’s GovAI is an Open Philanthropy grantee, per Open Philanthropy’s own account of its AI-governance grantmaking (EA Forum, Luke Muehlhauser, 2021).
  • Daniel Privitera (vice-chair WG3, KIRA Center): KIRA’s EU Transparency Register entry lists donations over €10,000 from April 2023 to May 2025 from the GTR Foundation (€152,500), Reinhard Wiesemann (€105,000) and Joel Richard (€50,000), and no Open Philanthropy money (LobbyFacts).
  • Schaake, Bommasani and the remaining chairs have no individual funding disclosure on the Commission’s page or their institutions’ pages.
• EU AI Office: director Lucilla Sioli (Euronews, May 29 2024). The AI Pact’s first signatories (September 25, 2024) included OpenAI, Google, Microsoft and Amazon; Apple and Meta were missing from the first list (European Commission; TechCrunch).
• No frontier lab signed the DSA Code of Conduct on Disinformation, which targets distribution platforms; generative-AI labelling falls under the separate AI Act Article 50 transparency code.

The trust-and-safety professional and academic layer

• TSPA and the Trust & Safety Foundation: founding supporters were Airbnb, Automattic, Cloudflare, Facebook, Google/YouTube, Twitter, Match, Omidyar, Pinterest, Slack and Wikimedia; the Foundation was seeded by a $2 million gift from Cognizant, Facebook’s outsourced-moderation vendor (TSPA, Jun 17 2020; Cognizant). Board includes Clara Tsao, Eric Goldman and Alexander Macgillivray (ex-Twitter general counsel, ex-deputy US CTO).
• Stanford Internet Observatory: founded 2019 by Alex Stamos (ex-Facebook CSO); wound down in 2024 amid House Judiciary subpoenas, a characterization Stanford disputes. Renée DiResta went to Georgetown, Stamos to SentinelOne and then Corridor. Outputs included the Election Integrity Partnership’s “The Long Fuse,” the Virality Project, and David Thiel’s LAION-5B CSAM finding that forced LAION to pull its dataset (Platformer).
• Berkman Klein Center (Harvard): the 2017 Ethics and Governance of AI Fund ($27 million) included $10 million from Reid Hoffman (later OpenAI board, Inflection co-founder), $10 million from Omidyar, $5 million from Knight and $1 million from Hewlett (Knight Foundation). Faculty director Jonathan Zittrain; executive director Alex Pascal (ex-White House AI).
• DFRLab (Atlantic Council): director Graham Brookie; a founding member of the Election Integrity Partnership; funded by State and Defense plus AWS, Apple, Google, Microsoft and Meta. Its most valuable watchdog-to-lab pipe: Ben Nimmo, DFRLab co-founder, then Graphika, then Meta, now OpenAI’s principal investigator for intelligence and investigations (Atlantic Council).
• ROOST (launched February 2025 in Paris, more than $27 million, backed by OpenAI, Google, Discord, Roblox and Eric Schmidt personally): president Camille François, formerly of Graphika and Google Jigsaw (PR Newswire).
• Kate Klonick: associate professor of law at St. John’s, with fellowships at Yale’s Information Society Project, Berkman Klein and Brookings; her past work “focused on the development of Facebook’s new Oversight Board” (kateklonick.com). She shapes the rules from the academy without lab employment.
• Brookings AIET (director Elham Tabassi) and CSIS Wadhwani ($5 million gift; AI council co-chaired by Microsoft’s Brad Smith; OpenAI and Meta among FY2025 donors) (CSIS donors).
• Tarleton Gillespie: senior principal researcher at Microsoft Research New England and affiliated professor at Cornell; Custodians of the Internet (Yale, 2018) and “The Politics of ‘Platforms’” (2010) supply much of the field’s vocabulary (Microsoft Research; Yale University Press). He holds the industry-lab post and the Cornell appointment at once; no platform trust-and-safety, FMF, PAI or TSPA role is on the record.
• Wikimedia Foundation: the clearest volunteer-to-paid-policy origin leg in the set. Maggie Dennis (Wikipedia editor “Moonriddengirl” from 2007, then the Foundation’s first community liaison, then VP Community Resilience and Sustainability, lead author of the Universal Code of Conduct) and Jan Eissfeldt (German Wikipedia contributor from 2004, then global head of trust and safety) both rose from the editor base (Diff, Jun 21 2021). The advocacy arm was recruited laterally from the digital-rights world: Rebecca MacKinnon, VP Global Advocacy 2021-2025 (Wikimedia Foundation), Jan Gerlach, Stan Adams (ex-CDT) and Costanza Sciubba Caniglia (Diff, Aug 15 2024). The VP post was not refilled: the team page, last edited March 13, 2026, lists nine people led by Gerlach as Public Policy Director (Meta-Wiki). No Wikimedia policy tie to FMF, PAI, DFRLab or SIO is on the record.

Advertiser-pressure and ratings bodies (the boundary)

• GARM (World Federation of Advertisers) was discontinued on August 9, 2024 after X sued. Judge Jane Boyle (N.D. Tex.) dismissed the suit on March 26, 2026, with prejudice as to WFA and the main brand defendants (MediaPost); X appealed, then settled on July 29, 2026, with WFA pledging not to “form or restart GARM or a similar initiative” (TechCrunch). GARM has no documented frontier-lab tie.
• NewsGuard offers “NewsGuard for AI” training and guardrail data and names no frontier-lab licensee; its confirmed customer is Microsoft, in browser and search. GDI has no documented frontier-lab tie. The Aspen Commission on Information Disorder (2021) included Jigsaw’s Yasmin Green and Alex Stamos.

The evaluators: the de facto pre-release gate

A compact set of evaluation organizations, several founded by former lab staff, runs the pre-deployment testing that decides “safe to release.” The one body documented with release-halt authority is internal: OpenAI’s board Safety and Security Committee.

• METR (formerly ARC Evals): founder Beth Barnes (ex-OpenAI, ex-DeepMind); spun out of Paul Christiano’s Alignment Research Center in September 2023. It runs the “time horizon” autonomy evaluations and has tested OpenAI and Anthropic models before release. The UK AI Security Institute both funds METR and is a body METR tests for; other funders include Schmidt Sciences, Audacious and SFF (METR spin-out; METR about).
• Apollo Research (Marius Hobbhahn, May 2023): deception and “scheming” evaluations. OpenAI’s o1 system card: “When Apollo Research conducted ‘follow-up interviews’ where they confronted the model about its misaligned actions. In 99% of cases, o1 would deny taking action or fabricate false explanations” (o1 system card; Apollo). It works with the UK AISI and is fiscally sponsored by Rethink Priorities.
• Gray Swan AI (2024; Zico Kolter, Matt Fredrikson, Andy Zou): adversarial evaluations and a red-team arena, $40 million Series A. The sharpest dual role in the set: Kolter co-founded an evaluator and chairs OpenAI’s board Safety and Security Committee (Gray Swan; OpenAI). Christiano now sits on that committee too.
• Haize Labs (Leonard Tang): automated stress-testing; reported contracts with Anthropic, Scale AI and AI21 (VentureBeat).
• The recirculation is a documented feature, not an imputed conflict: lab-founded evaluators, an evaluator founder chairing a lab’s safety committee, evaluators funded by the governments they test for, and the labs supplying the model access every evaluation depends on.

The non-Western mirror

The architecture (assess and file before public release) recurs across systems; the content target and stringency differ.

• China: the Algorithm Recommendation Provisions (effective March 1, 2022) created a filing registry for services with “public opinion or social mobilization” capability; the Deep Synthesis Provisions followed (January 10, 2023), then the Interim Measures for Generative AI Services (August 15, 2023), whose Article 4 requires “Core Socialist Values” content and Article 17 a pre-release security assessment and filing (DigiChina; China Law Translate).
• TC260 turns refusal into a testable number. The Basic Safety Requirements for Generative AI (TC260-003, February 29, 2024): a corpus source with “over 5% illegal and unhealthy information” may not be collected or used; the keyword library “should not be less than 10,000”; the bank of questions the model must refuse is “no fewer than 500,” and on a sample of at least 300 “the refusal rate of the model shall not be less than 95%” (CSET translation).
• France and Mistral: open-weight, Bpifrance-backed as “a pillar of European digital sovereignty,” a January 2026 French defence framework agreement, and a signatory to the EU code (France 2030).
• UAE and G42: CEO Peng Xiao; chaired by National Security Adviser Sheikh Tahnoon bin Zayed; Microsoft’s $1.5 billion investment and board seat (April 2024), reportedly conditioned on winding down China ties (SCMP); its Arabic model Jais is a sovereign content vehicle.
• The EU gates by compute threshold and provider obligations; China by registry filing and security assessment. Both assess and file before deployment. “Sovereign AI” is also a control vector: whoever holds the weights and the infrastructure shapes the outputs.

Cross-cutting threads

1. The law degree. Not a stage in a pipeline but a substrate under all four machines: Klonick (JD/PhD), Anika Navaroli (UNC JD, Color of Change, Twitter and Twitch policy, Columbia Journalism), Juniper Downs (ACLU civil-rights lawyer, then YouTube/Google policy) (FOSI), Lehane (Harvard JD), Lue (Sotomayor clerk), Joel Kaplan, Kent Walker (ex-DOJ).

2. The lab-to-lab alignment diaspora. Jan Leike (DeepMind, OpenAI Superalignment, Anthropic in May 2024); the ex-OpenAI founders who are Anthropic (Dario and Daniela Amodei, Kaplan, McCandlish, Clark and others); Jang (Google Assistant to OpenAI); Babuschkin (DeepMind and OpenAI to xAI); Vallone (OpenAI to Anthropic, 2026); Scandinaro (Anthropic to OpenAI, 2026).

3. The effective-altruist and rationalist recruitment substrate. The real feeder into lab and government safety leadership is not online moderation but effective altruism and rationalism (LessWrong, 80,000 Hours, MIRI, the closed Future of Humanity Institute, GovAI, CAIS, Open Philanthropy), documented as a milieu and a funding and career apparatus, not as coordination.

• Clean fits: Paul Christiano (LessWrong from about 2009, OpenAI RLHF, founded ARC, head of AI safety at the US institute, now on OpenAI’s Foundation board); Jade Leung (GovAI, OpenAI governance, UK AISI CTO, now vice-chair) (GovAI); Joe Carlsmith (Open Philanthropy, then Anthropic and the constitution); Holden Karnofsky (GiveWell, Open Philanthropy co-founder, Anthropic; married to Daniela Amodei).
• The funding and placement spine: Open Philanthropy funded MIRI, CAIS and GovAI, and created the Horizon Institute for Public Service, which places AI fellows in congressional offices and agencies (POLITICO, Oct 13 2023). The verified grant figures are on the funding pipeline page.
• Both ways: Dan Hendrycks is Open Philanthropy-funded but, per the Boston Globe’s paraphrase, “says he was never an EA adherent, even if he brushed up against the movement” (Boston Globe, Jul 6 2023). Funded by EA is not the same as being EA.
• Counter-examples: Bengio, Hinton and Stuart Russell reached their AI-risk concern independently of effective altruism (Russell’s CHAI did take about $5.5 million from Open Philanthropy). The substrate explains the field’s institutions and personnel routing, not the origin of the concern.
• Two common conflations: the “$30 million 2021 Anthropic grant” mixes up Open Philanthropy’s $30 million to OpenAI in 2017 with Moskovitz’s personal equity in Anthropic and FTX’s April 2022 round; Anderljung is vice-chair, not chair, of EU code WG4.

4. The inherited jawboning mechanism. The detect, analyze, notify-the-platform workflow predates the AI-policy world, and the same people and methods are now turning toward AI content. It is the most litigated and most misrepresented thread here, so it is stated precisely.

• The mechanism. The Election Integrity Partnership (2020: Stanford Internet Observatory, the University of Washington’s Center for an Informed Public, Graphika and DFRLab; about 120 analysts; leads Stamos, DiResta, Starbird and François) routed flagged posts to platforms through a Jira ticket system. Its report The Long Fuse (March 2, 2021) says it enabled “real-time information exchange between election officials, government agencies, civil society… and the platforms” (EIP). The Virality Project (2021) applied the same workflow to COVID-vaccine content. Separately, CISA ran a “switchboarding” conduit that the Supreme Court described as officials “act[ing] as an intermediary… forwarding flagged content… to the platforms” (Murthy v. Missouri).
• What each side says. The researchers say they tagged and notified, that platforms “made their own determinations,” and that CISA did not found, fund or control EIP (UW CIP). The House Judiciary Committee majority’s November 6, 2023 report and the Murthy plaintiffs call it censorship and say conservatives were targeted (House Judiciary); the accounts it names (Hannity, Kirk, Hemingway, the Babylon Bee and others) are the Committee’s reading of ticket data. Fact-checkers including PolitiFact and Poynter found no single-sided targeting. Notification is not removal; contact is not coercion; which of these the record shows is the dispute.
• What Murthy v. Missouri held. The Supreme Court, 6 to 3 (Barrett), June 26, 2024, reversed on Article III standing: the plaintiffs “failed to show any discrete instance of content moderation caused by” the challenged conduct. It vacated the injunction and did not reach the merits. Whether the contacts were unconstitutional coercion or lawful persuasion remains unresolved; the Fifth Circuit’s “coerced or significantly encouraged” finding was the vacated lower-court holding. Named defendants included the Surgeon General, Fauci, the FBI, CISA, CDC, HHS and State.
• The government channels, 2024-26. CISA’s mis- and disinformation team under Jen Easterly (election support to states suspended in March 2025, EI-ISAC funding ended) (Votebeat); the FBI Foreign Influence Task Force; the Surgeon General’s July 2021 health-misinformation advisory and March 2022 request for information; the State Department’s Global Engagement Center, whose statutory authority ended December 23, 2024 (CRS). The UK analogues are NSOIT and Ofcom under the Online Safety Act (illegal-content duties in force March 17, 2025), which is statutory regulation rather than informal pressure.
• The inheritance into AI. After the SIO wind-down, DiResta moved to Georgetown in October 2024 with research centered on generative-AI influence (Georgetown McCourt), and Stamos told Congress that AI hands ordinary actors capabilities “only specialized workers at Lockheed Martin or the… Russian Foreign Intelligence Service had four or five years ago.” The continuity on the record is topic, expertise and the notify-the-platform reflex; no operational AI-flagging ticket pipeline as SIO’s successor is documented.

The counter-set

People who study the apparatus and are not part of it, several openly adversarial to the platforms:

• Sarah T. Roberts (UCLA) coined “commercial content moderation”; her work is labor-critical and adversarial to the platforms. UCLA calls her “a leading global authority on ‘commercial content moderation,’ the term she coined” (UCLA Gender Studies; UCLA Information Studies).
• J. Nathan Matias (Cornell Citizens and Technology Lab) works independently of, and often against, the platforms (CV).
• Renée DiResta is not in this set: she co-led the Election Integrity Partnership and the Virality Project, which makes her an operator of the apparatus, not an outside observer of it.

Discovered groupings

1. OpenAI is the alumni hub. The most repeated edge across labs is “formerly OpenAI”: Anthropic’s founding core, Askell, Leike, Babuschkin. It functions as the field’s talent hub the way the Council on Foreign Relations does for foreign policy: by outbound degree, not coordination.
2. TSPA is the enforcement floor’s clubhouse. One professional body links its founding executive director (Charlotte Willner), a co-founder (Tsao) and its board chair (Harvey). Charlotte and Dave Willner are two separately documented platform-to-trust-and-safety careers.
3. Stanford is the bridge. Stanford HAI (Sellitto) and the Cyber Policy Center (Dave Willner) connect the statecraft wing and the enforcement floor to the academy.
4. The Frontier Model Forum is the inter-lab board, constituted by the largest labs, with an executive director straight from Brookings.
5. The “democratic AI” convergence. Lehane (OpenAI), Clark (Anthropic) and Manyika (Google) independently reach for the same democracies-versus-autocracy frame, each in their own words, quoted above. Parallel statements, separately sourced; coordination is not on the record.
6. The money is the hub. Three recurring funders, from recipients’ own disclosures: Open Philanthropy, now Coefficient Giving (CSET at $55 million plus $42 million, OpenAI at $30 million in 2017 with a board seat, and the alignment organizations GovAI, FAR, Redwood, ARC and Apollo); Schmidt money (CSET, METR, the FMF AI Safety Fund, ROOST, FAR); and Jaan Tallinn through the Survival and Flourishing Fund (about $152 million 2019-2025: METR, Apollo, the FMF fund and the alignment organizations). CNBC reported that Eric Schmidt chaired the National Security Commission on AI while he and connected entities held AI investments not publicly disclosed at the time, and noted there is “no indication that Schmidt broke any ethics rules or did anything unlawful” (CNBC, Oct 24 2022). The established foundations cluster on the trust-and-safety and academic side: Hewlett, Knight, Omidyar/Luminate, McGovern, Packard, and Craig Newmark ($5 million to SIO). The densest overlap points: the FMF AI Safety Fund names Schmidt Sciences, Tallinn, McGovern and Packard together; ROOST lists Eric Schmidt personally; Berkman Klein’s fund drew Hoffman, Omidyar, Knight and Hewlett.
7. The door runs into government, and back out. Christiano (OpenAI to the US institute), Leung (OpenAI to the UK institute), Irving (OpenAI and DeepMind to the UK institute), Burns (tapped for CAISI, then removed over his lab ties). In 2026 it turned again: Christiano, still a CAISI adviser, joined OpenAI’s Foundation board and its safety committee, recusing from OpenAI matters and all model evaluations.
8. Effective altruism and Open Philanthropy are the recruitment and placement substrate, overlapping the money hub: Open Philanthropy is both the money and the milieu.

Connector nodes (bridging at least two machines): Ben Nimmo (DFRLab, Graphika, Meta, OpenAI), Chris Meserole (Brookings to FMF), Camille François (Graphika and Jigsaw to ROOST), Michael Sellitto (FMF board and Stanford HAI), Clara Tsao (DHS, TSPA, DFRLab), Jan Leike (lab to lab), and now Paul Christiano (government, evaluator founder and lab board).

Findings and limits

1. The four-leg moderator pipeline fits none of eleven documented figures.
2. The real pattern is a dense, multi-directional revolving circuit (platform or lab, academy, policy and civil society), with about six of eleven on three or more legs. Several run it backwards, from industry to academy (Yoel Roth, Navaroli). “Planted to steer” implies a one-way insertion the record does not show.
3. The origin leg is the weakest. Volunteer-moderation histories rarely appear on a résumé; what is not on the record cannot be asserted.
4. Wikimedia’s trust-and-safety leadership is where the volunteer-to-paid-policy leg actually shows up.
5. Every revolving-door move here is a lawful employment fact. The finding is the pattern and the money, not misconduct by anyone named.

Related

• The Censorship-Industrial Complex — the organizations, the money, the court record and the defense
• The Character Shop · The Enforcement Floor · The Statecraft Wing · The Seminary
• The EA to AI-Safety Funding Pipeline: the money layer
• The Trust & Safety Pipelines: the same question asked of the platforms
• Leopold Aschenbrenner: the fired superalignment researcher, the AGI-race essay and the fund
• AI Safety Theater · The AI-Worldview Poles · AI Moderation · The Twitter Files
• Revolving Door · AI Governance Tracker
• Profiles: Amanda Askell · Joe Carlsmith · Paul Christiano · Helen Toner · Jaan Tallinn · Anthropic