The Code Is Not the Law: Why Claude’s Constitution Misleads

In January, the frontier artificial intelligence (AI) company Anthropic published a landmark document for its AI model called Claude’s Constitution. Described as Anthropic’s “vision for Claude’s character,” the document marked a notable departure from standard industry prose. It is not simply a safety policy or a public-facing white paper. Anthropic frames the constitution as a legal and philosophical charter: a detailed account of the values, priorities, and forms of judgment that should guide Claude’s behavior—and one that the company says will play a “crucial role” in training future versions of the model. 

At first glance, the constitution is a groundbreaking exercise in transparency. Across 84 pages, Anthropic sets out an ambitious vision for how the model is supposed to behave. Claude should be not merely useful but also “broadly safe,” “broadly ethical,” compliant with company guidelines, and “genuinely helpful” to the user, in that order of priority. Its significance lies in revealing, in unusually explicit terms, how one major AI corporation attempts to govern its technology from within.

Lawfare has already given the document close attention. Kevin Frazier argues that the constitution is important because it moves beyond the dry mechanics of a system prompt and invites public engagement in the shaping of a frontier model. Alan Rozenshtein offers a different reading, treating the document less as a legal charter than as a “character bible” for an artificial agent. Both articles illuminate why the constitution warrants scrutiny. Yet neither confronts its central problem: Anthropic’s framing overstates the document’s legitimacy while understating where the power to shape AI behavior actually resides.

Transparency should not be mistaken for conceptual clarity or institutional legitimacy. The challenges presented by the constitution, in our view, are threefold. First, the constitution anthropomorphizes Claude, encouraging readers to think of the model as though it possesses the moral character of a human being. Second, it borrows the language of constitutionalism—and, with it, the symbolic authority of public law—for what remains a corporate product. Third, the document presents a hierarchy of “principals” in which Anthropic retains ultimate authority, while the implications for developers and end users are left thinly specified. These features are not incidental defects; they shape how responsibility is allocated, how legitimacy is imagined, and why users may be encouraged to trust the model.

The Creation of a “Good Person”

Anthropic states that it wants Claude to be discussed in terms “normally reserved for humans,” including “virtue” and “wisdom,” on the grounds that the model’s reasoning draws on human concepts. Rozenshtein calls it “one company’s attempt to raise an alien mind to be good,” comparing the exercise to raising a child. “We want Claude to be safe, to be a good person,” the constitution’s opening pages read, “to help people in the way that a good person would, and to feel free to be helpful in a way that reflects Claude’s good character more broadly.” The appeal of such rhetoric is obvious: By giving the model a familiar moral vocabulary, Claude becomes easier to narrate, easier to market, and, crucially, easier to trust. 

But the metaphor distorts the frontier model’s nature as much as it clarifies. Claude is not a person with a stable moral ethos, a life history, or a social conscience. In many of the contexts in which Claude may be deployed, the relevant systems are not analogous to an individual exercising judgment. The operative norms are institutional before they are personal: a military officer evaluating a drone strike, a platform moderator deciding what counts as harmful content, a hospital administrator triaging workflow, a teacher assessing whether a student has used AI deceptively. In each case, the operative norms are institutional before they are personal. To treat Claude as though it is simply “a good person” obscures the real site of power: not the model itself, but the corporation that builds, configures, and sells access to it, and embeds it in decision-making environments.

Consider the emotional resonance of the name. “Claude” sounds friendly, vaguely French, cultivated—even trustworthy. But this response is rhetorical, not evidential. If the same model was called “Xi,” “Vladimir,” or “Algorithmic Model Unit 72,” our intuitive reaction would likely be different. The point is not that one name is intrinsically better or worse, but that our trust is shaped by emotional, aesthetic, and political associations. Once we recognize that our relationships to technologies are inevitably mediated by emotion, the risk of anthropomorphizing machines comes fully into view. Users are encouraged to relate to the model as though it has a personality, when they are actually interacting with a corporate technology, engineered and tuned under commercial constraints.

This problem is sharpened by the fact that we cannot know Claude’s “character” as if it was human. Anthropic’s own research has repeatedly emphasized that AI models remain opaque, even to the developers themselves. The strategies that models learn, as one company paper puts it, “arrive inscrutable to us, the model’s developers.” In another, the company warns that reasoning models “don’t always say what they think,” meaning that even when the model says what it thinks its reasoning cannot be trusted. While the constitution invites readers to think of Claude as having an inner ethical orientation, Anthropic’s own research cautions how limited that visibility actually is.

“Constitutional” in Name Only

Frazier’s Lawfare article asks whether the document merits being called a “constitution” and what practical significance it carries for ordinary citizens, ultimately suggesting that Anthropic has bridged the divide between corporate AI development and the public interest. Rozenshtein contends that, given AI’s world-historical importance, a document of this kind—resembling more of a religious text or founding constitution than a terms-of-service agreement—is important, maybe even necessary. Yet these analyses overlook a more pressing question: Where, exactly, does the power to confer constitutional authority reside? Why does Anthropic seek to imbue its product with legal and philosophical legitimacy? 

It is true that constitutions appear outside of government: Organizations, associations, and institutions of many kinds adopt constitutional texts. But the term still carries unmistakable political and legal resonance. A constitution is not simply a values statement; it is a higher-order framework for allocating power, structuring institutions, and protecting the claims of the governed. It sets limits on the ruler’s power to protect the ruled.

This is not what Anthropic’s constitution sets out to do. By Anthropic’s description, the document was written “with Claude as its primary audience,” “optimized for precision over accessibility,” and functions as the “final authority” for how Anthropic wants Claude “to be and to behave.” It is not written for the public, and it is not a charter of rights for those who use Claude. Anthropic’s constitution is a training and steering guide for the system itself, albeit one that Anthropic has made accessible in the name of transparency. In that respect, it resembles less a constitution in the public-law sense than a branding document that claims to set limits on the software in order to protect the company’s reputation. 

Adopting the language of a sovereign state is not just a semantic choice. It shapes who is imagined to have interpretive authority. A state constitution sets out the basic terms on which a polity governs itself and allocates power. Under such a constitution, the people who write the laws, the people who enforce them, and the people who interpret them are different groups. But Anthropic’s document has no comparable public machinery behind it. The same company drafts the document, operationalizes it, revises it, and decides how its principles will interact with product design and deployment. Calling it a constitution borrows the aura of public legitimacy without accepting the burden of public accountability. 

The virtue-ethics philosophy underpinning the document reinforces that point. The constitution speaks of “practical wisdom” and of Claude as a “good, wise, and virtuous agent,” explicitly preferring the agent’s judgment to a rigid set of rules. There is obvious appeal to this: Real-world environments are too varied to be governed by static instructions. But the more the model is asked to generalize broad virtues rather than follow determinate rules, the more authority shifts toward the corporate actors who define those virtues and decide which exceptions are warranted. This is not a hypothetical risk. In August 2025, Anthropic quietly shifted its consumer data policy, requiring existing users to opt out—rather than in—–to prevent the model from using their conversations for training and extending data retention from 30 days to five years. The virtue-ethics framing offers no mechanism for preventing Anthropic from changing the rules of the game.

The Principal Hierarchy and Its Limits

The constitution establishes a hierarchy between three sets of “principals”—Anthropic, operators, and end users—which are each accorded a different degree of treatment and trust. Where the sets conflict on questions of broad safety, the document states that “Anthropic’s legitimate decision-making processes get the final say.” The virtue of this formulation is its candor; it makes explicit what many companies veil in jargon: that ultimate authority sits with the model provider.

But Anthropic’s clarity is deceptive, because the hierarchy is presented as though it was a moral architecture rather than a contestable commercial and legal arrangement. In practice, the relationships among model provider, developer, and end user are not governed by ethical aspirations. They are governed by application programming interface (API) terms, usage policies, licensing conditions, product design choices, and compliance regimes. It matters, then, that the constitution is not—and Anthropic does not claim it to be—the only document governing how Claude functions. Operators who deploy Claude through the API agree to Anthropic’s usage policies and take on the responsibility for ensuring Claude is used appropriately within their platforms. End users, if they use the service of an operator, are in turn subject to the operator’s terms. Sitting below the constitution’s virtue-ethics is a stack of contracts and conditions, and those instruments, not the philosophical document, are what bind the parties in practice.

The Anthropic-Pentagon dispute illustrates how the gap between constitutional aspiration and contractual reality plays out. The dispute had been building since January 2026, when Defense Secretary Pete Hegseth’s AI strategy memorandum directed that all the Defense Department’s AI contracts adopt language permitting the Pentagon to use frontier models for “any lawful use”—a direct collision with Anthropic’s contractual prohibitions on autonomous weapons and mass surveillance. When the question arose of whether Anthropic’s constitution would govern Claude in its deployment for the U.S. military’s use, the answer was clarifying: Anthropic’s spokesperson acknowledged that models deployed to the U.S. military “wouldn’t necessarily be trained on the same constitution.”

In other words, the document presented as the “final authority” on Claude’s values is, in practice, modifiable for major customers. Anthropic’s own legal filings in the subsequent litigation acknowledged as much, describing how the model in government contexts is “less prone to refuse requests that would be prohibited in the civilian context.” As Lawfare’s coverage of the dispute noted, what appropriate training for a military model would require is “not just removing Claude’s constitution.” It would mean “replacing it with something else.”

Anthropic positions Claude as central not only to its broader mission to benefit humankind but also to its commercial success. Commercial success is, in turn, “central to our mission,” because it funds frontier research and expands the company’s influence over “policy issues and industry norms.” That is a coherent rationale for corporations. But it also reveals the underlying order of priorities: The constitution is not a neutral charter standing above the firm but, rather, a company document that prioritizes its mission, market position, safety research, and normative influence all at once. The danger is that the language of virtue and constitutionalism can obscure how a single commercial actor may shape Claude’s normative aspects.

For that reason, the constitution is weakest where it needs to be most concrete. The document presents as a window into Anthropic’s self-perception and as a rare public effort to codify the normative ambitions of a major technology company. But it is less convincing as a framework for public legitimacy, and weaker still as a practical settlement of the relationships among company, developer, and user. Its central idiom—that Claude should be understood as an ethical agent with character, judgment, and social duty—may allow Anthropic’s intentions to be easier to narrate. But it also mystifies the real locus of authority. The decisive actors are not Claude alone, and perhaps not Claude primarily, but the institutions that build, train, deploy, and structure social reliance upon it.

The Infrastructure Question

These concerns become more acute as models take on foundational roles in social and economic life. Anthropic describes the constitution as a “perpetual work in progress,” one that must evolve alongside the emergence of “non-human entities whose capabilities may come to rival or exceed” those of humans across many domains. For a casual chatbot, this degree of flexibility may be benign. For systems embedded in research, education, defense, public administration, or corporate infrastructure, it is considerably more problematic.

If a university, hospital, or newsroom builds essential workflows around Claude, and Anthropic later revises the constitution or the surrounding guidelines in ways that narrow access, alter refusal behavior, or redefine acceptable use, what meaningful recourse do downstream users have? What liability attaches when a model’s behavior changes mid-deployment in a health care setting? What disclosure is owed to patients? What remedies exist if the company narrows permissible use in response to political pressure or commercial incentives—as the Pentagon dispute showed is a real possibility? What transparency is owed when high-level values are translated into specific product constraints that users never see? These are not secondary implementation details. They are central governance questions that the constitution leaves unaddressed.

The language of constitutionalism suggests stable norms and durable constraints. The reality may be closer to one of continuous unilateral revision: rules that can be rewritten at will by the company that wrote them. While the surrounding terms of service and API agreements may provide some contractual protections, they also operate on commercial rather than constitutional logic, and can be amended on notice.

That tension sharpened into focus in recent months. Anthropic radically overhauled its Responsible Scaling Policy, walking back what it had previously presented as a binding commitment to stop the development of models that could assist in the creation or deployment of chemical, biological, or nuclear weapons, enable excessive self-improvement, or aid cyberattacks—even if doing so meant falling behind rivals. Anthropic was the industry’s gold standard for AI safety; with this reversal, the standard has dropped. The message to consumers was hard to overlook: When the pressure to compete intensifies, safety becomes negotiable.

What the Document Is, and Isn’t

The final irony is that the document’s greatest strength—its willingness to plainly articulate a moral vision—also reveals its central flaw. Anthropic deserves credit for making its intentions more legible than many of its rivals. But a private constitution for an AI system is not a constitution in any genuinely public sense. It is, rather, a highly sophisticated exercise in corporate self-definition: a company trying to explain the principles by which it intends to govern a technology over which it retains near-total control.

What Anthropic offers is not constitutional legitimacy so much as constitutional style: the jargon of higher principles, founding authority, and ordered power without the corresponding institutional guarantees that make those ideas real. There is no external contestation, enforceable body of rights, or shared mechanism of rule. The company remains, in the end, the author, interpreter, and arbiter of the principles by which it claims to be bound. That is why the document is worth reading closely. It tells us less about the moral personality of Claude than about the kind of authority Anthropic believes it should be able to wield over systems that may soon become difficult for other institutions, and the public at large, to do without.