The Unstable Object: Why "The Global Governance of AI" Cannot Be Defined Broadly Enough to Justify Itself and Narrowly Enough to Exclude What It Successfully Governs

Not a Divisible Good tests whether Erman and Furendal's authorization-delegation-revocability chain, granted their own stated domain, transfers to a case — longevity allocation — their published work never examined. This paper asks a prior and more exposed question: whether "the global governance of artificial intelligence," as Erman and Furendal themselves define and defend it, is a stable object at all, or whether the same reasoning that justifies the term's global scope already dissolves any principled boundary the term might otherwise hold at the edge of the AI sector itself. The paper's evidence is Erman and Furendal's own text, checked directly against the primary sources rather than against a prior draft's paraphrase of them. Their argument for why AI governance must be global rather than national rests explicitly on an externality claim, stated in their 2025 Current History essay: AI's disruptive effects "are not confined by geographical boundaries," a condition they compare directly to "air pollution and the release of greenhouse gases," where AI models "create substantial value for some, but serious and tangible problems for many others." This is an argument from the technology's own general-purpose character. The paper's central claim is that this argument, applied consistently rather than selectively, does not stop at the jurisdictional boundary; a technology whose disruptive effects are not confined by geography, because it is general-purpose, is for the same reason not confined by sector. Erman and Furendal's own definition of their object of study — stated in their 2022 Moral Philosophy and Politics paper and repeated nearly verbatim in their own 2024 Nature Machine Intelligence essay — is, read on its own terms, already sectorally unbounded: "the rules, processes and decision procedures... established by governments, international and intergovernmental organizations, non-state and private actors to regulate the development and deployment of those systems," aimed at solving problems or providing "goods associated with AI technologies." Nothing in this wording excludes a cellular reprogramming protocol an AI system was built to accelerate. This paper's second and more concrete contribution builds directly on No New Actor To Name's finding that OpenAI, Anthropic, and Alphabet — actors Erman and Furendal do not merely gesture at but name specifically, and analyze using a precise, three-part authority typology of their own construction — are now also longevity-research actors, holding market and, in at least two cases, epistemic authority in that domain by the identical operational criteria their own published typology already applies to these same organizations' AI-governance conduct. Erman and Furendal's own epistemic-authority criterion for OpenAI and DeepMind is that they "partly function as research institutes, probing the limits of AI technology's capabilities in research papers that inform regulatory efforts" — a criterion that transfers by its own wording, not by analogy, the moment the research papers in question probe the limits of cellular reprogramming rather than language modeling. The paper's evidentiary base is strengthened further by two facts Erman and Furendal themselves report rather than this paper importing from outside: that fourteen of twenty-two members of the US Department of Homeland Security's 2024 AI safety board were sitting tech-company CEOs, and that Anthropic, OpenAI, and Google are named founding participants, alongside Amazon, IBM, Meta, and Nvidia, in the US State Department's Partnership for Global Inclusivity on AI. The actors this paper needs are not supplied by extension. They are supplied by Erman and Furendal's own reporting. The paper concludes that this is a scope challenge, not a refutation: Erman and Furendal's apparatus, wherever it is finally judged to apply, remains the correct account of what legitimate non-state authority requires. What this paper contests is the cla