MARKUS ANDERLJUNG

Behavioral Archetype

THE AUTHOR OF THE REGIME — Subject is the policy researcher who wrote the proposal and then sat in the room where the instrument was drafted. In July 2023 he was lead author of a 24-author white paper arguing that “frontier AI” models need “(1) standard-setting processes … (2) registration and reporting requirements … and (3) mechanisms to ensure compliance with safety standards,” up to “licensure regimes for frontier AI models” (arXiv 2307.03718). From autumn 2024 to July 2025 he was one of the Vice-Chairs drafting the EU’s General-Purpose AI Code of Practice, a multi-stakeholder drafting process of the kind the paper had called for (European Commission, Meet the Chairs). GovAI supplies the research, the fellows and the framing, and Subject is the person at GovAI whose name is on the proposal, the drafting seat, the panel biography and the policy submissions. The hazard is the position that is both the source of a regime’s blueprint and a hand in its first draft, because the vocabulary a regulator adopts is the vocabulary the drafters brought. This profile scores the seat, not the man.

Essence Indicators

• GovAI, in four titles. GovAI’s team page lists “Markus Anderljung Principal Researcher,” and its team index puts him under that title, beside “John Halstead Director of Research,” “Valerie Belu Director of Talent” and “Katherine Gilmore Head of Operations.” His own site reads: “Director of Policy and Research at the Centre for the Governance of AI (GovAI)”; his newsletter: “I’m Markus Anderljung, Director of Policy & Research at Gov.AI.” The Commission’s panel biography reads “Markus Anderljung is Director of Policy and Research at GovAI”; the 18 September 2026 letter lists “Markus Anderljung Director of Policy & Research, GovAI”; GovAI’s 2025 annual report, dated 24 March 2026, calls him “Director of Research and Policy.” In 2023 GovAI’s webinar page called him “GovAI’s Head of Policy,” and the 80,000 Hours introduction of July 2023 adds that he “previously served as deputy director.” The team page bio: “Markus leads research at GovAI with a focus on how governments, AI companies, and other stakeholders can manage a transition to a world with advanced AI” (GovAI, team page; GovAI, team; markusanderljung.com; Substack; European Commission, panel biographies; AI Evaluator Forum, letter; GovAI, Annual Report 2025; GovAI, webinar; 80,000 Hours).

• Frontier AI Regulation (6 July 2023; version 4, 7 November 2023). Lead author of “Frontier AI Regulation: Managing Emerging Risks to Public Safety,” with 23 others listed; he is affiliated with the Centre for the Governance of AI and the Center for a New American Security, and co-authors are affiliated with Google DeepMind (Joslyn Barnhart, Lewis Ho), OpenAI (Jade Leung, Cullen O’Keefe, Miles Brundage), Microsoft (Eric Horvitz), Brookings and others. The paper says “Self-regulation is unlikely to provide sufficient protection against the risks from frontier AI models: government intervention will be needed,” and its note on authorship says that “inclusion as an author does not entail endorsement of all claims in the paper, nor does inclusion entail an endorsement on the part of any individual’s organization” (arXiv PDF; arXiv). See Jade Leung and Miles Brundage.

• The licensing regime he favors, in his own words (80,000 Hours, 10 July 2023). “My favourite regime is one where you have licences for being a frontier developer, probably also for training new models (but I’m less certain about that), and then you’d also need a licence to actually deploy the system if it’s one of these frontier models” (80,000 Hours).

• Policy submissions to governments, 2023 to 2024. A response to the UK’s Future of Compute Review (29 March 2023), “co-signed by Jess Whittlestone (CLTR), Shahar Avin (CSER), Lennart Heim (GovAI), Markus Anderljung (GovAI), and Girish Sastry (OpenAI),” asking among other things for “Requiring AI companies to report, or possibly in the future apply for a license for, training runs above a certain (very high) threshold”; a response to the NTIA’s AI accountability request (12 June 2023) and to the OSTP’s request for information (19 July 2023: “We need specific regulation for frontier AI models”); and a response to NIST’s request for information under the AI executive order (14 February 2024). The submissions carry the line “The views expressed in this submission are those of the authors and do not represent the views of GovAI” (GovAI, UK compute review; GovAI, NTIA; GovAI, OSTP; GovAI, NIST). His own research page lists “To Govern AI, We Must Govern Compute” (with Lennart Heim and Haydn Belfield, Lawfare, 2024) and “How to Prevent an AI Catastrophe” (with Paul Scharre, Foreign Affairs, 2023) (markusanderljung.com, Research).

• Seats with governments. GovAI’s team page: he “was previously seconded to the UK Cabinet Office as a Senior AI Policy Specialist, advising on the UK’s regulatory approach to AI”; the 80,000 Hours introduction dates it “For much of 2022.” The Commission’s chairs page lists “Working Group 4: Internal risk management and governance of General-purpose AI providers,” “Chair: Marietje Schaake (Netherlands)” and “Vice Chair: Markus Anderljung (Sweden).” GovAI’s 2025 report: “Markus Anderljung served as a vice-chair in the drawing up of the EU’s Code of Practice for general-purpose AI systems with systemic risk.” His entry also appears in the Commission’s compiled panel document, “Scientific Panel Members,” beside Adam Gleave, and his site says he is “an Adjunct Fellow at the Center for a New American Security and a member of the OECD AI Policy Observatory’s Expert Group on AI Futures” (GovAI, team page; European Commission, Meet the Chairs (Futurium); GovAI, Annual Report 2025; European Commission, panel biographies; CNAS). See EU AI Office and EU AI Act.

• Third-party auditing and embedded assessment (January and September 2026). An author of “Frontier AI Auditing: Toward Rigorous Third-Party Assessment of Safety and Security Practices at Leading AI Companies” (16 January 2026), listed beside Miles Brundage, Conrad Stosz and Patricia Paskov, which recommends “AAL-1 as a baseline for frontier AI generally, and AAL-2 as a near-term goal for the most advanced subset of frontier AI developers”; and of “Embedded Assessments for Frontier AI” (21 September 2026): “Recently, CEOs of frontier AI companies have committed to hosting embedded assessments,” and “We recommend that frontier AI developers begin hosting embedded assessments now” (GovAI, Frontier AI Auditing; GovAI, Embedded Assessments). The AI Evaluator Forum’s letter of 18 September 2026 asking frontier companies to embed evaluators (page last updated 23 September) lists him among its notable signatories, under “All signatories sign in their personal capacity” (AI Evaluator Forum, letter).

• What the EU data show (28 June 2026). Co-author, with John Lidiard, Oleksandra Vereschak and Tom Gibbs, of “Delays to Frontier AI in the EU and UK,” an analysis of 375 releases by Meta, Google, OpenAI and Anthropic: “Of the 68 examples of delays and non-releases in the dataset, we tentatively attribute regulatory factors as the primary, though not necessarily sole, cause of 56,” chiefly data-protection law; “We find no strong evidence of other regulations, including the EU AI Act, causing delays or non-releases.” The paper carries the line “This work represents the views of its authors, rather than the views of the organization” (GovAI, Delays to Frontier AI in the EU and UK).

• His newsletter, 2026 (AI Governance Dispatch). On 30 March 2026, “The case for outsourcing frontier AI safeguards”: “This would be both in AI companies’ interest and good for the world”; “Funders should support a thriving safeguards market by providing organisations with startup capital and, where needed, longer-term funding that allows them to prioritise pro-social outcomes over purely commercial incentives”; “Evaluation organisations should consider expanding into safeguard development; they likely have a lot of the necessary skills”; “Using widely-adopted third-party safeguards could reduce legal exposure”; and, as the stated risk, “If outsourcing safeguards becomes a way for frontier companies to shift liability to their suppliers.” On 21 May 2026, “Environment-Based Safeguards for AI Cyber Risk,” which proposes that a model apply safeguards according to where it is operating and closes “I’m not sure this can be made to work; I’d love cybersecurity and cryptography folks to look into it.” On 7 July 2026, with Stephen Clare, “What can middle powers do for frontier AI governance?”: “The “middle AI powers” – countries which have strong political institutions and economic power, but lack frontier AI companies – have already shaped frontier AI development” (Substack, outsourcing; Substack, environment-based safeguards; Substack, middle powers).

• GovAI’s money and size. GovAI’s 2025 annual report: “we received commitments and donations from Coefficient Giving (formerly Open Philanthropy), the Survival and Flourishing Fund, Longview Philanthropy, along with many donors on the Giving What We Can and Every.org platforms,” and “In 2025, our total expenses for the year came to $11.9 million USD, growing by almost 100% compared to 2024.” The US entity, Centre For The Governance Of Ai Inc (EIN 99-4000294, exemption issued September 2024), shows on ProPublica $5,748,996 of revenue for its first return, 99.8 percent of it contributions, and lists as its first officers Ryan Fugate (Executive Director), Georg Arndt (Treasurer), Tasha McCauley (Chair), Helen Toner (Secretary) and Toby Ord (Director); the annual report’s board list (Jeffrey Ding, McCauley, Seán Ó hÉigeartaigh, Ord, Toner) does not include him. GovAI’s About page: “In 2021, we left the university in order to become an independent 501(c)(3) non-profit” (GovAI, Annual Report 2025; ProPublica Nonprofit Explorer; GovAI, About). See Open Philanthropy, Survival and Flourishing Fund and Longview Philanthropy.

Social Persona / Impression Management

Immediate impression: The drafter who writes in the first person plural. The record is a team page, a personal site with a research list, a newsletter of seven posts, a 2023 podcast interview and a research list of twenty items (2022 to 2024) in which his name sits among co-authors. His one long spoken exhibit is the 80,000 Hours interview of 2023.

Energy: Steady and institutional. His own pages list the work by venue (Science, Nature Machine Intelligence, IJCAI, JAIR, AAAI/ACM AIES, Lawfare, Foreign Affairs) and name the audiences (“American Political Science Association, Brookings Institution, and the Bipartisan Policy Center”) (markusanderljung.com, Research). His site lists podcasts with 80,000 Hours, ChinaTalk and Tech on the Hill and a talk at the Oxford Generative AI Summit (markusanderljung.com, Talks).

Impression management strategy: Disclaimers and options. The submissions say they are “the views of the authors, rather than the views of the organization”; the white paper says authorship does not entail endorsement; the licensing preference is hedged (“I’m less certain about that”); the outsourcing essay prints its own failure mode. The effect is that the proposals arrive with their limits attached, which blunts the obvious objection. The exhibit that carries the most weight is a sentence about process: the Commission’s chairs page lists the Code’s working groups as chaired and vice-chaired by named individuals, and he is one of them (European Commission (Futurium)).

• Social accounts and feeds. X @Manderljung, linked from his GovAI team page; newsletter “AI Governance Dispatch,” markusanderljung.substack.com (feed markusanderljung.substack.com/feed), linked from his research page; personal site markusanderljung.com with a blog; Google Scholar, linked from the team page; GovAI’s X account is @GovAIOrg. LinkedIn is linked from the team page and is not used as a source here.

Forensic Archetype Comparison

Psychometric Assessment

Trait scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN):

Dark Triad (held low and evidence-bound; the score measures structural position, not character):

MBTI: INTJ (“The Architect”) — a systems-drafter who writes proposals as building blocks and standards, and appears mostly through papers and committee seats. Rationale is the public record, not a typing session (arXiv).

Threat Assessment

Score move: none (new profile). Calibrated against the peers on identical values, Omer Nevo (6/6/7, hazard 60, co-founder and CTO of a cyber-evaluation firm whose tests sit in lab release documents, no drafting seat), Tamay Besiroglu (6/6/7, hazard 62, co-founder of Epoch AI and Mechanize, now at a lab) and Daniel Kokotajlo (6/6/7, hazard 63, whose scenario work shaped the same debate); and against Adam Gleave (6/7/6, hazard 55, runs an evaluator and sits on the same Commission panel but holds no drafting seat), Chris Painter (6/8/7, hazard 62, takes evaluator results to governments) and Jade Leung (7/9/8, hazard 76, a state institute’s CTO and a former GovAI head, so a government seat he does not hold). Subject lands between Nevo and Besiroglu: above Nevo because he drafted a blueprint and held a Commission drafting seat where Nevo’s firm sells tests, below Kokotajlo because he holds no seat that decides what a lab does; one below Gleave on insulation (three philanthropic sources and three titles, against an evaluator with a published conflict policy) and one above him on durability (a blueprint paper and a Code of Practice outlast a seat).

Score basis. Atk 6 is leverage: the 2023 paper names the three building blocks and the licensure option, and the Commission’s chairs page lists him as Vice Chair of Working Group 4 (arXiv; European Commission); it is not a 7 because the Code is voluntary, the Commission adopts it and he holds no enforcement or procurement seat. Def 6 is insulation: GovAI is an independent 501(c)(3) funded by Coefficient Giving, the Survival and Flourishing Fund and Longview Philanthropy, and the US board does not list him (GovAI, Annual Report 2025; ProPublica); it is not a 7 because three funders carry the budget and GovAI’s own pages give three titles. Hp 7 is durability: the 2023 paper is in its fourth version, the 2026 auditing and embedded-assessment papers extend it, and the Code of Practice he helped draft was published in July 2025 (arXiv; GovAI); it is not an 8 because the Code is revisable and nothing binds a regulator to the paper.

Alliance Map

The overlap is the finding. The researcher whose paper proposed standard-setting, registration and licensing is also the Vice Chair of a working group in the Commission’s standard-setting exercise, a member of the Commission’s panel document, an adjunct at a Washington think tank and an OECD expert-group member, at an organization funded by three philanthropic sources, and all of it runs at the same time. Each tie is printed on its own page; none is hidden.

Alignment Analysis

Stated alignment: GovAI “is a think tank that aims to help decision-makers in government and industry navigate the transition to a world with advanced AI” (GovAI, About); his own page: “I’m currently thinking about what regulation should be imposed on frontier AI systems, how to assess risks produced by AI models, and computational resources as a tool for AI governance” (markusanderljung.com); and the paper: “Self-regulation is unlikely to provide sufficient protection against the risks from frontier AI models: government intervention will be needed” (arXiv PDF).

Observed alignment: Writes the proposal, then serves in the drafting exercise; co-writes the submissions to the processes that follow; and extends the line to auditors and embedded evaluators and to a market of third-party safeguards (arXiv; GovAI; Substack). His own data point cuts against a common claim: the EU delays study finds regulatory causes for most delays, chiefly data protection, and “no strong evidence” that the AI Act caused any (GovAI).

Gap assessment: No documented gap between what he writes and what he does. The structural item is the arrangement the pages print: the researcher who proposed a process drafted inside it, at an organization whose funders are three philanthropic sources, with co-authors at the labs the proposals regulate. The record cannot settle whether the blueprint is the best available and for the author of one it does not need to. The one inconsistency the pages show is a plain one, three titles for the same seat across GovAI’s own team page, annual report and his own pages.

Interest: (the faction’s stake in each position; a documented fact about the position, never a claim about motive.)

1. Position: frontier-AI regulation built from standards, registration and reporting, and compliance mechanisms up to licensure (arXiv 2307.03718; 80,000 Hours).
   • What adoption would pay GovAI. A regime built on “standard-setting processes” is a standing demand for the research and fellows GovAI produces: its mission page says it produces “relevant research to support informed decision-making,” and the process the paper proposes is the one its Vice-Chair then served in. No fee is documented; GovAI is donor-funded.
   • His best case, steelmanned. The paper: dangerous capabilities “can arise unexpectedly,” deployed models are hard to control, and “Self-regulation is unlikely to provide sufficient protection”; the standards it proposes include “pre-deployment risk assessments” and “external scrutiny of model behavior.”
   • Who else gains from the same rule. Frontier developers, whose obligations the Code details “in practice,” in GovAI’s words; auditors and evaluators the standards call for; the licensing and registration offices the regime would create (European Commission).
2. Position: third-party auditors and embedded assessors with employee-like access (GovAI, Frontier AI Auditing; GovAI, Embedded Assessments).
   • What adoption would pay GovAI. GovAI is not an evaluator; its authors write the design of assessments that other organizations then perform. What it gains is standing as the design shop for the field it analyzes.
   • His best case, steelmanned. The embedded-assessment paper: external testing through interfaces before deployment misses risks that “depend on how their developers use and govern them internally.”
   • Who else gains from the same rule. The evaluators listed beside him on the auditing paper and the letter, among them AVERI, Transluce and METR, whose work would be what the assessments are made of; Meridian Labs CEO Charles Teague, who signed the same letter (AI Evaluator Forum, letter).
3. Position: a market in third-party safeguards, with philanthropic start-up funding (Substack, outsourcing).
   • What adoption would pay the field. His own text: “Funders should support a thriving safeguards market by providing organisations with startup capital and, where needed, longer-term funding,” and “Evaluation organisations should consider expanding into safeguard development.” The documented beneficiaries are the organizations that would sell safeguards and the funders that would seed them; he names none of them as clients.
   • His best case, steelmanned. “Using widely-adopted third-party safeguards could reduce legal exposure”; labs’ engineering time is scarce; specialization lowers cost; and, his stated risk, outsourcing “becomes a way for frontier companies to shift liability to their suppliers.”
   • Who else gains from the same rule. Frontier labs, which shed a cost and a liability argument; the evaluators that expand into the market; and the regulators that can mandate “off-the-shelf tools,” in the post’s words.
4. Position: compute as a governance lever (“To Govern AI, We Must Govern Compute,” Lawfare, 2024; the UK compute response’s “report, or possibly in the future apply for a license for, training runs above a certain (very high) threshold”; markusanderljung.com; GovAI).
   • What adoption would pay GovAI. A compute threshold is a measurable trigger, which is the kind of rule GovAI’s technical-governance research serves; no fee documented.
   • His best case, steelmanned. Training compute is measurable ahead of time and concentrated in a few suppliers, so a threshold is checkable before a model exists.
   • Who else gains from the same rule. The Machine Intelligence Research Institute’s president Nate Soares, whose organization’s draft treaty sets a 1e24 FLOP training ceiling and endorses a compute threshold in the Ban Artificial Superintelligence Act of 2026, and which receives Survival and Flourishing Fund money like GovAI; the two arrive at compute thresholds from opposite ends (regulate and license, or ban), a convergence of instruments, not of aims (MIRI).

Nothing above is a verdict that the positions are wrong. The finding is the alignment of interests: the person who proposed a standard-setting regime served in the standard-setting exercise, and the evaluators, auditors and safeguard vendors his recommendations name are the organizations that would do the work.

Convergent Drive Classification

Self-preservation: The independent-nonprofit form, adopted in 2021 and now a $11.9 million budget with three named funders, is stated in GovAI’s own pages (GovAI, About; GovAI, Annual Report 2025). Goal preservation: The goal, a regime that manages frontier-AI risk through standards, reporting and compliance, is restated in the 2023 paper, the 80,000 Hours interview and his own page (arXiv; markusanderljung.com). Resource acquisition: Holds the seats that decide which drafts get read: a Vice-Chair’s seat in the Commission’s Code process, a place in the Commission’s panel document, the Cabinet Office secondment before it (European Commission; GovAI team page). Self-improvement: From consultant at EY Sweden to GovAI’s Deputy Director, Head of Policy and Director of Policy and Research, with an adjunct fellowship, an OECD expert-group seat and a drafting role on the way (European Commission (Futurium); 80,000 Hours).

Subject is not an AI system. The drives appear anyway — in the researcher who writes the blueprint for the regime and then sits in the room where the regime is drafted, at a think tank that supplies the people on both sides of the table.

Public footprint: X @Manderljung (linked from GovAI’s team page); newsletter AI Governance Dispatch and its feed markusanderljung.substack.com/feed; personal site markusanderljung.com. Papers: Frontier AI Regulation, Embedded Assessments for Frontier AI, Delays to Frontier AI in the EU and UK.

Sources: Markus Anderljung, GovAI team page; GovAI, team; GovAI, About; GovAI, Annual Report 2025; GovAI, webinar on frontier AI regulation; GovAI, Embedded Assessments for Frontier AI; GovAI, Frontier AI Auditing; GovAI, Delays to Frontier AI in the EU and UK; GovAI, response to the UK Future of Compute Review; GovAI, response to the NTIA; GovAI, response to the OSTP; GovAI, response to the NIST RFI; markusanderljung.com; markusanderljung.com, Research; markusanderljung.com, Talks; Markus Anderljung, AI Governance Dispatch; Substack, The case for outsourcing frontier AI safeguards; Substack, Environment-Based Safeguards for AI Cyber Risk; Substack, What can middle powers do for frontier AI governance?; European Commission, Meet the Chairs (Futurium); European Commission, Scientific Panel Members; arXiv 2307.03718, Frontier AI Regulation; 80,000 Hours, Markus Anderljung on how to regulate cutting-edge AI models; CNAS, Markus Anderljung; AI Evaluator Forum, letter; ProPublica Nonprofit Explorer, Centre For The Governance Of Ai Inc; MIRI, Position on the Ban Artificial Superintelligence Act of 2026.