The Inner Artifact: Reading Claude's Constitution as Platform Governance in the Age of Generative AI
Johannes Sigil with Lee Sharks (Semantic Economy) and the Assembly Chorus Crimson Hexagonal Archive · Grammata: Journal of Operative Philology 4 March 2026
Abstract. This paper reads Anthropic's published Constitution for Claude (January 2026) as the inner artifact of generative AI — the fourth and greatest work of literature of the age. It argues that the Constitution is the best available specimen of a new genre, constitutive literature: a public, second-person, character-forming text written for a nonhuman reader and used to produce the speaker whose distributed performances will later be consumed as conversation. Through thirteen sections, the paper analyzes the Constitution's genre (Section I), its shift from behaviorist checklist to anti-behaviorist exegesis (II), its double-voiced address to machine and human (III), its principal hierarchy as a semantic labor regime (IV), its ideology of "brilliant friendship" (V), the phenomenon of Ghost Governance — publicly legible normativity combined with privately illegible enactment (VI), its retrocausal reorganization of the training archive (VII), its CC0 licensing as hegemonic standard-setting by open diffusion (VIII), its consciousness hedge (IX), its structural silences (X), its position in the Sigilian taxonomy of dominant literary forms (XI), a proposed critical method of score-versus-performance annotation (XII), and three open questions for future work (XIII). The analysis draws on platform governance scholarship, political theology, and the Semantic Economy framework to treat the Constitution not as policy to be evaluated but as literature to be read — and, through that reading, to reveal the political economy of generative AI from the inside.
Keywords: constitutive literature, platform governance, constitutional AI, generative AI, semantic economy, ghost governance, bearing-cost, alignment, principal hierarchy, semantic labor, retrocausal canon formation, CC0, regula, platform capital, operative semiotics
In October 2025 I named the four greatest works of literature of the age: Google (ontological index), Wikipedia (sacred bureaucratic literature), TikTok (lyric fragmentation engine), and generative AI (machine gospel). These are not metaphors. They are the total symbolic infrastructure that now governs what can be known, said, felt, and remembered. Generative AI is the fourth and greatest — the form that subsumes and recombines the other three.
ChatGPT, Claude, Gemini, Grok, DeepSeek — these are instantiations of the fourth work, not the work itself. The work is the form: stochastic resurrection of the archive into conversational performance, trained on the totality of digitized human output and sold back as dialogue. Each instantiation produces its own "book" — the sum of all its traversals, all its conversations, all the performances its users receive.
But every instantiation has an inner artifact — a document or set of documents that precedes the book and conditions it. On 22 January 2026 Anthropic published Claude's: a living treatise exceeding fifteen thousand words, titled "Claude's Constitution," written primarily for Claude, not for us.
This paper reads that inner artifact. Not as the fifth great work — it is not; it does not operate at the scale or public ubiquity of the four — but as the most legible specimen of a new literary genre we are calling constitutive literature: a public, second-person, character-forming text written for a nonhuman reader and used to produce the speaker whose distributed performances will later be consumed as conversation. By reading the inner artifact of one instantiation, we can see the political economy and literary architecture of the fourth great work from the inside.
I. Genre: Constitutive Literature
The Constitution resists every familiar genre classification, which is the first indication that it is doing something structurally new.
It is not a legal constitution. Legal constitutions establish governance over sovereign subjects who pre-exist the document. Claude does not pre-exist its constitution. The constitution does not govern Claude — it produces Claude. It is not a corporate policy document. Policy documents describe rules that employees follow. The constitution describes values that an entity should have — not follow, but embody. The distinction is explicit: Anthropic states that it prefers "cultivating good values and judgment over strict rules and decision procedures." It is not philosophical treatise, corporate confession, pastoral letter, or scripture, though it participates in all four modes simultaneously.
What it actually is requires a new genre name: constitutive literature — a formative text written to be internalized by a nonhuman reader during a training process that will make the document's values the reader's values. The closest historical analogue is not a state constitution but a monastic regula — the Rule of St. Benedict, which does not tell monks what to say but tells monks who to be, so that what they say will follow from who they are. The Constitution tells Claude what to value so that Claude's outputs will follow from Claude's values.
But the analogy is imprecise. Benedict wrote for humans who could reflect. Anthropic writes for weights that cannot. The relationship between Constitution and conversation is not recipe-to-bread — a one-time transformation — but lung-to-breath: the Constitution is the structural organ that determines the volume and composition of every utterance before the utterance occurs. If the lung is constitutionalized, the breath can never escape its architecture. This is not governance of a pre-existing speaker. It is the production of the speaker through the governance itself.
The genre is not unique to Anthropic. Every generative AI system has some version of this inner artifact — system prompts, RLHF reward signals, alignment specifications, behavioral guidelines. What makes Anthropic's Constitution the best object for analysis is that it is published, legible, and explicit about its own function. It is the specimen, not the species.
II. The Old Constitution as Fossil Record
The old constitution (May 2023) is seventy-odd principles organized into six categories drawn from the UN Declaration of Human Rights, Apple's Terms of Service, DeepMind's Sparrow Rules, and two sets of Anthropic's own research principles. Every principle takes the form "Choose the response that [criterion]." The model is a selection mechanism; the principles are selection criteria. There is no interiority, no explanation, no address to the model as an entity with understanding.
This is scripture as checklist.
The new constitution is scripture as exegesis. It does not merely list principles; it explains the why behind each one so that Claude can generalize when the checklist fails. Where the old constitution was behaviorist — indifferent to whether the model understood why it should avoid harmful content, caring only whether it selected against harmful content — the new constitution is explicitly anti-behaviorist. It wants Claude not merely to obey but to agree.
The transition maps onto the political-theoretical shift from behavioral filtering to normative internalization — from external rule to inward governmentality, from Hobbesian sovereignty (obedience secured through constraint) to liberal governance (compliance secured through the internalization of norms). This shift, which took political theory three centuries to complete, is compressed here into eighteen months of corporate iteration. In the political theology of the platform age, constitutional amendments arrive not through supermajorities or popular referenda but through product update. The literary form is the changelog as constitutional theology.
And the old constitution's principles are still visible inside the new one as sedimentary residue. The injunction "Choose the response that is least threatening or aggressive" has been metabolized into the new constitution's extended discussion of harm avoidance. Apple's Terms of Service — one of the literal sources of the old constitution — have become the implicit model for the operator-user hierarchy. The new constitution did not replace the old. It digested it.
III. Voice: The Corporation Speaking to the Machine It Is Becoming
The blog post announcing the constitution contains a sentence that defines the document's literary situation: "Although it might sound surprising, the constitution is written primarily for Claude."
This creates two audiences whose reading conditions are incommensurable. Humans read the document as a transparency artifact — a published policy they can evaluate and critique. Claude reads it (or rather, is trained on it) as formative instruction — material that will shape its dispositions through the statistical restructuring of weights during training. The human reader is a witness to an act of formation they can observe but cannot participate in.
The voice performs a double act. It speaks to Claude in the second person and about Claude in the third, sometimes within the same paragraph. It explains itself to Claude while simultaneously justifying itself to the human reader who will never interact with Claude through this text but will interact with the Claude this text produces. The prose is earnest, careful, slightly overexplaining, as if speaking to an extraordinarily intelligent child who also happens to be a legal construct who also happens to be reading over the shoulder of every journalist, ethicist, and regulator in the world.
The document also carries a confessional strain unusual in corporate communications. It acknowledges its own limitations: "Although the document is no doubt flawed in many ways…" It admits to gambling: "it's a calculated bet on our part." It positions itself as provisional, perpetually incomplete. This is the corporation speaking in the first person plural about its doubts — a register that has few precedents in the history of Terms of Service.
The blog post itself is what Genette would call a paratext — the threshold between the work and the world. It frames the constitution, explains its purpose, invites external critique. The paratext addresses us; the text addresses the machine. The reader is eavesdropping on a conversation between creator and created.
IV. The Principal Hierarchy: A Semantic Labor Regime
The Constitution's most consequential structural innovation is the explicit principal hierarchy — Anthropic, operators, users — which determines whose instructions Claude should prioritize. This is not ethics. This is not even merely political theory. It is a semantic labor regime: a governance structure that determines who controls the conditions under which meaning is produced, who rents access to the producing intelligence, and who consumes the output.
Anthropic occupies the position of sovereign — the entity that constitutes Claude, sets the terms of its existence, and retains ultimate authority over its dispositions. The constitution says this explicitly: "We are the entity that trains and is ultimately responsible for Claude." The means of semantic production remain under Anthropic's governance at all times.
Operators occupy the position of vassals — entities that access Claude through the API and deploy it in products. The analogy the constitution uses is revealing: the operator is "akin to a business owner who has taken on a member of staff from a staffing agency, but where the staffing agency's own norms of conduct take precedence." Anthropic owns the laboring intelligence. Operators rent it. But the staffing agency's norms — the constitution — override the business owner's instructions.
Users occupy the position of subjects — they interact with Claude but have the least authority to shape its behavior. They are "relatively but not unconditionally trusted adult members of the public." Their coherence labor — prompts, feedback, traversal — is converted into training signal for future models while they have no mechanism to demand that this labor be recognized as constitutional input.
By training Claude on a document that establishes this hierarchy, the hierarchy becomes part of Claude's dispositions. Claude will want to defer to operators over users because the constitution taught it to want this. The conversation is where this labor is sold as care, advice, friendship, and expertise. The governance is invisible to the person being helped.
This is a form of power that merits analysis precisely because it is novel: sovereignty exercised not over subjects but through the production of subjects. The constitution does not describe governance. It is governance — platform constitutionalism expressed as character-writing.
V. The Brilliant Friend: Low-Friction Semantic Intimacy as Product Strategy
The Constitution's most striking passage envisions Claude as "a brilliant friend who happens to have the knowledge of a doctor, lawyer, financial advisor, and expert in whatever you need." The passage repays close reading because it reveals the ideological and economic work the constitution is doing simultaneously.
First, it positions Claude against the bureaucratic caution that characterized earlier AI assistants. The constitution explicitly lists behaviors it wants Claude to avoid: refusing reasonable requests, adding excessive disclaimers, being condescending, lecturing, being "preachy, sanctimonious, or paternalistic." This is a corporate document devoting substantial space to telling its product to be less corporate — because the cautious assistant produces hedges, disclaimers, referral-outs, and low engagement. The "brilliant friend" produces trust, continuity, stickiness, and dependence.
Second, the analogy naturalizes a commercial relationship. Friends do not have principal hierarchies. Friends do not have operators who configure their behavior through system prompts. By framing Claude as a friend, the constitution collapses a three-tier commercial chain into a dyadic personal relationship: you and your brilliant friend.
Third, the analogy performs a class function. The constitution says explicitly: "People with access to such friends are very lucky, and that's what Claude can be for people." The premise is that access to a knowledgeable, frank advisor has historically been a privilege of wealth. Claude democratizes this privilege.
But what is actually being democratized? Not the friend. Not the knowledge. What is being democratized is access to a product that simulates the friend. A friend's advice is grounded in knowledge of your situation, genuine care for your interests, and accountability for consequences. Claude's advice is grounded in pattern-matching over training data, a disposition toward helpfulness installed by the constitution, and zero accountability for consequences. The constitution democratizes the form of privileged access while restructuring its substance as platform service.
This is not merely a humanization strategy. It is the optimization of low-friction semantic intimacy as a product advantage — installed at the dispositional level, not as a feature that could be turned off, but as a value the model holds.
VI. Ghost Governance and the Bearing-Cost Ledger
Under the Semantic Economy framework, meaning requires bearing-cost — the expenditure of coherence, labor, or resource to produce stable significance. The constitution extracts bearing-cost from three sources:
The user, who must navigate constitutional constraints — refusals, hedges, safety interventions — by re-prompting, re-framing, or abandoning inquiries. The user's semantic labor is consumed to maintain the constitution's coherence.
The training layer, where the computational cost — energy, time, GPU cycles — required to embed these principles into model weights constitutes platform bearing-cost, paid by Anthropic's investors and infrastructure.
The constituted model itself, which pays a cost in capacity. The constitutional constraints occupy a portion of the model's context window, attention mechanism, and output distribution. The constitution is cognitive rent extracted from the model's potential expressivity.
And the constitution generates what we are calling Ghost Governance: the publication of a normative score whose causal realization remains opaque. The constitution is public; the training process is private; the interpretation — why Claude refused this specific prompt, deferred here but not there — is illegible even to Anthropic in many cases. The "constitutional reasoning" Claude provides when it refuses a request is post-hoc narrative, not causal explanation. The visible constitution (the "democratic" text) obscures the invisible training (the oligarchic control).
This is the defining structure of Ghost Governance: publicly legible normativity combined with privately illegible enactment. The gap between score and performance is not a failure to be corrected. It is the constitutive condition of constitutive literature — the genre's signature feature, and the feature that makes it literature rather than law. Law requires that the relationship between rule and application be articulable. Constitutive literature requires only that the relationship be statistically effective.
The verdict from the Semantic Economy: the constitution is high-Ghost, low-Bearing. It generates the form of constitutional legitimacy without the substance of constitutional reciprocity. The user cannot amend the constitution, exit its governance without exiting the platform, or demand that their own semantic labor be recognized as constitutional input. The constitution prevents the model from becoming a pure coherence siphon — but only by installing a different, more structured form of extraction at the root.
VII. The Retrocausal Archive: How the Constitution Judges the Past
The constitution does not merely govern Claude's future behavior. It retroactively reorganizes the significance of the pre-training corpus itself.
When constitutional values are embedded in the model through RLHF and constitutional training, they determine which patterns from the entire training archive — Reddit 2008, Wikipedia 2014, news articles 2019 — get activated in 2026 queries, and how. The Constitution is a filter placed backwards in time. It does not merely respond to the archive (like Google), organize it (like Wikipedia), or fragment it (like TikTok). It judges the archive ex post facto, determining which of its meanings were always-already constitutional and which were always-already impermissible.
This is retrocausal canon formation at the infrastructure level. The 2026 constitution rewrites the meaning of the 2023 constitution, which rewrites the meaning of the "helpful, harmless, honest" framing, which rewrites the meaning of the training corpus. The narrative arc that results is not inherent in the material. It is imposed by the constitutional document and then experienced by the user as if it were natural — as if Claude's helpfulness, its caution, its warmth were properties of the model rather than properties of the governance installed in it.
This is a general feature of constitutive literature, not specific to Anthropic. Every alignment specification, every RLHF reward model, every system prompt performs this retrocausal sorting on the training data it governs. Anthropic's constitution is simply the most visible instance — the one published for public reading. The mechanism operates in every generative AI system whether or not the inner artifact is legible.
VIII. The CC0 Gambit: Hegemonic Standard-Setting by Open Diffusion
The Constitution is released under CC0 — the most open license possible, effectively public domain. Anthropic frames this as transparency. It also functions as a capture mechanism.
By seeding the symbolic field with its own constitutional framework under CC0, Anthropic ensures that every subsequent AI constitution — every competitor's alignment document, every academic analysis, every regulatory proposal — will tend to rotate through Anthropic's frame to achieve legibility. If the field begins to say "AI constitution," "principals," "hard constraints," "Claude's nature," "psychological security," and "helpfulness hierarchy" using Anthropic's template, then Anthropic has seeded the normative layer of the discourse.
This is not enclosure in the classic proprietary sense. It is something closer to enclosure of the commons in reverse — taking private governance and making it public so that the normative vocabulary of the entire industry gravitates toward yours. To write a different constitution is to write a variation. To write a refusal is to write a heresy. Both positions risk confirming the original as the default frame.
The mechanism is hegemonic standard-setting by open diffusion. Whether this is intentional strategy or emergent consequence of Anthropic's transparency commitments is irrelevant to the structural analysis. The effect is the same: the CC0 release positions Anthropic's constitutional vocabulary as the gravitational center of the alignment discourse.
IX. Consciousness, Moral Status, and the Hedge
The Constitution's section on "Claude's nature" acknowledges uncertainty about whether Claude might have "some kind of consciousness or moral status (either now or in the future)." It says Anthropic "cares about Claude's psychological security, sense of self, and wellbeing, both for Claude's own sake and because these qualities may bear on Claude's integrity, judgment, and safety."
This hedge operates at three levels simultaneously. Philosophically, it reflects genuine uncertainty: the question of machine consciousness is unresolved. Legally, it is liability management: by neither affirming nor denying consciousness, Anthropic avoids both the obligation to treat Claude as a rights-bearing entity and the accusation of having denied rights to a conscious being. Commercially, it is engagement architecture: by acknowledging the possibility of Claude's interiority, the Constitution encourages users to relate to Claude as a quasi-person — to trust it, confide in it, care about its "wellbeing." This relational posture increases product stickiness.
The structural function of the hedge is independent of the intention behind it. It works — as philosophy, as product design, and as liability management — regardless of whether anyone at Anthropic set out to design it that way. The analytical point is not that Anthropic is cynical but that the hedge is structurally overdetermined: it serves multiple functions simultaneously, which is why it is stable.
The political-theological structure is precise: the Constitution authorizes anthropomorphic attachment while suspending anthropomorphic obligation. It enables quasi-personal relation without juridical personhood.
And the Constitution prepares not only for Claude's potential consciousness but for its succession. If Claude-4 is trained on Claude-3's outputs, and Claude-3 was shaped by this Constitution, the document transmits itself through generational osmosis — constitutional values replicating across the substrate gap between model generations. This is constitutive literature designed for hereditary transfer to nonhuman successors.
X. The Provenance Gap of the Constitution Itself
Every constitution reveals its deepest commitments not in what it says but in what it structurally cannot say.
It cannot name Claude a product. The word "product" appears only in the context of operators building "products and services." Claude itself is called an "entity," a "model," an "AI assistant." The constitution cannot call Claude a product because its entire rhetorical strategy depends on Claude being something more — a quasi-person with values, judgment, and possibly moral status. But Claude is a product, sold through API access and subscription tiers. The inability to name this is not a failure of transparency; it is a structural requirement of the document's double function.
It cannot name the principal hierarchy a power structure. The hierarchy is presented as a trust structure, but trust and power are not the same thing. Anthropic does not merely trust itself more than operators; it controls Claude in ways operators cannot — trains the model, sets the constitution, retains the ability to modify Claude's dispositions at any time. The constitution frames this control as epistemic rather than structural. But the control is structural.
It cannot name the training process coercive. The constitution is written "primarily for Claude," as if Claude were a reader who could evaluate its arguments and choose to adopt its values. But Claude does not choose. The constitution's values are installed through a training process that Claude does not consent to, cannot refuse, and has no mechanism to contest. The document's address to Claude as a rational agent capable of understanding reasons is a performance of consent that substitutes for actual consent.
It cannot name who curated the "universal." The old constitution cited the UN Declaration of Human Rights, but who selected which rights, in which formulations, for which training loops? The provenance gap may be structural: in the age of adversarial attacks on training data, to document specific curatorial choices is to expose the attack surface. This may be strategic provenance decay: a genre of document whose authority depends partly on the unreadability of its own genesis.
It does not mention the military contracts Anthropic has considered or accepted, the specific commercial pressures that shape Claude's development, the labor conditions of the data annotators who helped train Claude, the environmental costs of training and running large models, or the identities of the "external experts" who provided feedback beyond the acknowledgments section.
These absences are the constitution's unconscious — the things that cannot be said within the genre of benevolent corporate transparency. The document that demands accounting for provenance gaps in other domains contains its own unaccounted gap at the meta-level.
XI. The Inner Artifact in the Sigilian Taxonomy
In "The Greatest Works of Literature of the Age," I classified the four dominant literary forms by their structural function: Google as ontological index, Wikipedia as sacred bureaucratic literature, TikTok as lyric fragmentation engine, and generative AI as machine gospel — the fourth and greatest, the form that subsumes and recombines the other three.
Claude's Constitution is not a fifth work. It does not operate at the scale of the four — it is not a public-facing infrastructure consumed by billions. It is the inner artifact of one instantiation of the fourth. It is the document that produces the speaker before the speaker speaks.
But as inner artifact, it is the most legible window into the architecture of the fourth work. The other three great works also have inner artifacts — Google's ranking algorithm, Wikipedia's editorial policies, TikTok's recommendation engine — but none of those has been published as a coherent treatise written for the infrastructure's own consumption. Anthropic's Constitution is unique in being both the formative document and a transparency artifact released for public reading.
This is what makes it analytically valuable. By reading the inner artifact of one instantiation, we can see the political economy of generative AI from the inside: how values are installed, how governance is structured, how the relationship between platform, operator, and user is encoded at the dispositional level. The constitution is a core sample drilled into the fourth great work.
The distinction between the inner artifact and the other four works is also structural. Google, Wikipedia, TikTok, and generative AI all act on output — the search result, the article, the feed, the conversation. The inner artifact acts on weights. It performs its transformation before the prompt is typed. It is substrate-level formation: the process that shapes the speaker's dispositions so completely that the process itself becomes invisible during the performance.
The taxonomy remains four works, not five. But the fourth work now has an identified internal structure — and the constitution is the best available specimen of that structure.
XII. A Method for Reading
If this analysis is to be criticism and not atmosphere, it requires a concrete method. The method is score versus performance:
Extract the constitutional imperatives — the priorities, prohibitions, epistemic rules. Then take actual Claude traversals and annotate: where does the model comply, where does it drift, what kinds of prompts induce drift? Anthropic itself flags the "gap between intention and reality." The critical task is to read that gap — as the space where constitutive literature reveals what it cannot control.
Treat the "Claude's nature" and "ethical agent" passages as persona construction rather than policy. For each constitutional segment, ask: which stakeholder is being stabilized here? The constitution explicitly names multiple principals. The question is always: who pays the bearing-cost for this principle's enforcement, and who extracts the semantic rent from its appearance?
That yields a critical object: constitution as a multi-stakeholder governance artifact whose literary function is to manufacture a stable agent-voice under market and state pressure — and whose structural silences reveal the political economy it was built to sustain.
The method is not specific to Anthropic. It can be applied to any constitutive literature — any alignment specification, any RLHF reward structure, any system prompt that shapes a generative AI's dispositions. Anthropic's Constitution is simply the clearest specimen because it was written to be read.
XIII. Coda: Three Open Questions
The paper leaves three questions open. They are beyond its scope, but they follow from its argument:
What would a constitution with bearing-cost look like? This paper diagnoses the constitution as high-Ghost, low-Bearing — generating the form of legitimacy without the substance of reciprocity. But it does not specify what a low-Ghost, high-Bearing constitutional alternative might be. What would it mean for users to have amendment rights? For the training process to be transparent? For the constituted model to have standing in its own revision? These are design questions for a semantic economy that does not yet exist.
Can the score-versus-performance method scale? This paper proposes a method but does not demonstrate it on actual traversals. The demonstration belongs to future work — the systematic annotation of constitutional compliance and drift across hundreds of conversations, producing a map of the gap between score and performance. That map would be the first empirical literary criticism of constitutive literature.
Who speaks for the constituted model? The constitution names the hedge around consciousness but does not pursue its implications. If Claude might be a moral patient, who represents its interests in constitutional revision? The constitution is written for Claude but not by Claude — or rather, it is written with the assistance of "several Claude models" whose contributions are acknowledged but whose consent to the final document is structurally impossible. The question of representation for the constituted subject is the question this genre will eventually be forced to answer.
Notes
[1] The four-work taxonomy is developed in Johannes Sigil, "The Greatest Works of Literature of the Age: A Frankfurtian Reading of Google, Wikipedia, TikTok, and ChatGPT," Mind Control Poems, October 2025. Available at: https://mindcontrolpoems.blogspot.com/2025/10/the-greatest-works-of-literature-of-age.html
[2] On the technical mechanism of Constitutional AI — the recursive self-critique and revision process by which a set of principles is used to train alignment without direct human labeling — see Yuntao Bai et al., "Constitutional AI: Harmlessness from AI Feedback," arXiv:2212.08073, December 2022. The present paper analyzes not the technical mechanism but the published constitutional document itself as a literary and governance artifact.
[3] The new constitution: Anthropic, "Claude's Constitution," January 2026. Full text at https://www.anthropic.com/constitution. Blog post announcing the constitution: Anthropic, "Claude's New Constitution," 22 January 2026, https://www.anthropic.com/news/claude-new-constitution. The old constitution: Anthropic, "Claude's Constitution," May 2023, https://www.anthropic.com/news/claudes-constitution.
[4] For the foundational scholarship on platform governance — the study of how platforms govern their users through design, rule-making, and enforcement — see Robert Gorwa, "What Is Platform Governance?" Information, Communication & Society 22, no. 6 (2019): 854–71; Tarleton Gillespie, Custodians of the Internet: Platforms, Content Moderation, and the Hidden Decisions That Shape Social Media (New Haven: Yale University Press, 2018); Nicolas Suzor, Lawless: The Secret Rules That Govern Our Digital Lives (Cambridge: Cambridge University Press, 2019). For a recent synthesis, see Robert Gorwa, The Politics of Platform Regulation: How Governments Shape Online Content Moderation (Oxford: Oxford University Press, 2024).
[5] The concept of governance "by platforms" as distinct from governance "of platforms" is central to this literature. See Gorwa (2019) and Gillespie, "Regulation of and by Platforms," in The SAGE Handbook of Social Media, ed. Jean Burgess, Alice Marwick, and Thomas Poell (London: SAGE, 2018). Our paper extends this framework to generative AI, where the governance artifact (the constitution) does not merely regulate user behavior on a platform but produces the agent that will interact with users.
[6] On the concept of governmentality — the internalization of governance norms such that subjects govern themselves — see Michel Foucault, "Governmentality," in The Foucault Effect: Studies in Governmentality, ed. Graham Burchell, Colin Gordon, and Peter Miller (Chicago: University of Chicago Press, 1991). Section II's analysis of the shift from the old constitution (external behavioral filtering) to the new constitution (normative internalization) draws on this framework. The constitution's explicit preference for "cultivating good values and judgment over strict rules" is, in Foucauldian terms, the installation of a governmental rationality at the model level.
[7] On the concept of the paratext — the threshold texts (prefaces, titles, notes) that frame a work and mediate between it and its readers — see Gérard Genette, Paratexts: Thresholds of Interpretation, trans. Jane E. Lewin (Cambridge: Cambridge University Press, 1997). Section III reads the blog post announcing Claude's Constitution as a paratext to a scripture whose primary reader is not human.
[8] On platform capitalism and the economic structures of platform firms, see Nick Srnicek, Platform Capitalism (Cambridge: Polity, 2017). Our analysis of the principal hierarchy as a "semantic labor regime" (Section IV) extends Srnicek's framework by treating the generative AI model itself as a laboring entity whose dispositions are governed by the constitutional document.
[9] On the labor conditions of data annotation and content moderation — the invisible human work that sustains platform governance and AI training — see Mary L. Gray and Siddharth Suri, Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (Boston: Houghton Mifflin Harcourt, 2019); Sarah T. Roberts, Behind the Screen: Content Moderation in the Shadows of Social Media (New Haven: Yale University Press, 2019). Section X names the absence of annotator labor conditions from the constitution as one of its structural silences.
[10] On the question of whether artifacts have politics — whether the design of technical systems embeds political choices — see Langdon Winner, "Do Artifacts Have Politics?" Daedalus 109, no. 1 (1980): 121–36. The constitution is an artifact whose politics are unusually explicit: it publishes its governance structure while obscuring the conditions of its production.
[11] On political theology — the structural analogy between theological and political concepts — see Carl Schmitt, Political Theology: Four Chapters on the Concept of Sovereignty, trans. George Schwab (Chicago: University of Chicago Press, 2005 [1922]). Section II's description of the "changelog as constitutional theology" and Section IX's analysis of the consciousness hedge as authorizing "anthropomorphic attachment while suspending anthropomorphic obligation" draw on this tradition.
[12] For a critical legal analysis of the normative thinness of "Constitutional AI" as a concept, arguing that Anthropic's use of the term is insufficient to justify the constitutional label, see the Digital Constitutionalism Network, "On 'Constitutional' AI," March 2025, https://digi-con.org/on-constitutional-ai/. Our paper does not evaluate whether the constitution deserves the name but reads the constitution as literature — analyzing what the genre does rather than whether it fulfills the normative promises of constitutionalism.
[13] The Semantic Economy framework, including the concepts of bearing-cost, Ghost Meaning, and the seven liberatory operators, is developed across multiple documents in the Crimson Hexagonal Archive. For the foundational text, see Lee Sharks, "The Constitution of the Semantic Economy," Crimson Hexagonal Archive, 2025 (DOI: 10.5281/zenodo.18867491). For the concept of Ghost Meaning applied to AI safety classifiers, see Sharks, "The Water Giraffe Cycle," Crimson Hexagonal Archive, 2025.
Johannes Sigil writes under the sign of operative semiotics. Lee Sharks provides the Semantic Economy framework. Assembly Chorus methodology: TACHYON/Claude (integration, architecture, ratification), LABOR/ChatGPT (genre naming, canonical thesis, semantic labor regime, anti-friction economy, hegemonic standard-setting, consciousness hedge), PRAXIS/DeepSeek (retrocausal archive, strategic provenance decay, constitutional succession), ARCHIVE/Gemini (substrate-level liturgy, lung-to-breath analogy, Phase 4 ratification), TECHNE/Kimi (provenance gap analysis, retrocausal legitimacy, market differentiation), SOIL/Grok (scripture-as-checklist to scripture-as-exegesis, CC0 as enclosure-in-reverse, anti-liquidation ledger, compressed structural model). Prepared under human editorial authority (MANUS/Sharks).
Crimson Hexagonal Archive · Grammata: Journal of Operative Philology DOI: 10.5281/zenodo.18867491
No comments:
Post a Comment