Code Is Law: Why Anthropic's Constitution Is Not the U.S. Constitution

In 1787, a group of men met in Philadelphia to solve a problem of power: how to constrain authority without destroying it. They wrote a constitution.

In 2023, a group of engineers wrote another one.

It did not govern a nation. It governed a model.

The U.S. Constitution and Anthropic's Constitution share a name, but they belong to different eras of power. One assumes law is interpreted by humans inside institutions. The other assumes law can be embedded directly into behavior.

That difference is the clearest modern expression of an old idea: code is law (Lawrence Lessig, Code and Other Laws of Cyberspace, 1999: https://codev2.cc/).

From Consent to Calibration

The U.S. Constitution begins with "We the People" because legitimacy flows upward. Authority is borrowed from citizens, temporarily delegated to institutions, and constantly contested. Even when the system fails, it fails under the theory that the governed remain sovereign (U.S. Constitution, Preamble and Article V: https://constitution.congress.gov/).

Anthropic's Constitution begins somewhere else entirely: with principles selected by designers.

Claude does not consent. It is calibrated.

Anthropic's "Constitutional AI" framework trains models using a written set of normative principles that guide self-critique and reinforcement learning (Anthropic, "Constitutional AI: Harmlessness from AI Feedback": https://www.anthropic.com/news/constitutional-ai-harmlessness-from-ai-feedback). Authority does not emerge from representation or deliberation; it emerges from architectural choice.

Political constitutions presume disagreement and manage it. AI constitutions presume optimization and eliminate it.

Enforcement Without Institutions

The U.S. Constitution is inert without institutions. Courts interpret it. Legislatures operationalize it. Enforcement relies on coercive power (Alexander Hamilton, James Madison, John Jay, The Federalist Papers: https://avalon.law.yale.edu/subject_menus/fed.asp).

Anthropic's Constitution is enforced mechanically.

During training, the model evaluates its own outputs against constitutional principles--such as harm avoidance, respect for autonomy, and refusal of dangerous requests--and is rewarded or penalized accordingly (Anthropic, "Constitutional AI: Harmlessness from AI Feedback": https://www.anthropic.com/news/constitutional-ai-harmlessness-from-ai-feedback). The constraint is not applied after the fact. It reshapes the output distribution itself.

There is no appeal because there is no runtime judgment. The judgment already happened during training.

This is where "Code is Law" stops being metaphor and becomes mechanism.

Amendments Without Friction

Amending the U.S. Constitution requires supermajorities, state ratification, and political risk by design. Friction protects legitimacy (U.S. Constitution, Article V: https://constitution.congress.gov/constitution/article-5/).

Anthropic can revise its Constitution in a product cycle.

Principles can be updated, retraining can occur, and behavior can shift globally without public deliberation. Speed is the advantage. Centralization is the cost.

Democracy treats friction as a safeguard. Machine governance treats friction as latency.

Rights Versus Constraints

The American constitutional tradition is defensive. It limits government power and protects individual liberties--speech, religion, due process (U.S. Constitution, Bill of Rights: https://constitution.congress.gov/browse/amendments-1-10/).

Anthropic's Constitution does not protect the model. It constrains it.

Claude has no speech rights. It has refusal policies.

Where the U.S. Constitution draws red lines around government action, the Anthropic Constitution draws red lines inside an output space. Law moves from shielding citizens to shaping behavior directly.

Interpretation Without Visibility

U.S. constitutional interpretation is public: opinions, dissents, precedent, contestation. Legitimacy emerges from visible disagreement (Alexander Hamilton, James Madison, John Jay, The Federalist Papers: https://avalon.law.yale.edu/subject_menus/fed.asp).

Anthropic's Constitution is interpreted during training, internally and opaquely. The user sees refusals and safe completions, not the interpretive process that produced them.

Opacity does not negate governance--but at scale, it changes accountability.

Governance Without Territory

The U.S. Constitution governs a bounded population within defined borders.

Anthropic's Constitution governs wherever the model is deployed.

A refusal applies globally, independent of local speech norms or democratic processes. This is not national law. It is infrastructural governance.

How Power Flows: Two Constitutional Models

The U.S. Constitution channels power from citizens through federated institutions. Anthropic's Constitution channels authority from designers through training into model behavior. Both constrain power--but through fundamentally different mechanisms.

When Code Becomes the First Constitution

Lawrence Lessig argued that software architectures regulate behavior as effectively as statutes (Lessig, Code and Other Laws of Cyberspace, 1999: https://codev2.cc/). Speed bumps enforce traffic law without police. Access controls regulate speech without courts.

Anthropic's Constitution completes that argument.

Instead of writing laws for institutions that manage behavior, the law is written into the system that produces behavior. Governance shifts from enforcement to prevention.

The U.S. Constitution assumes agents with power who must be constrained. The Anthropic Constitution assumes a system without agency that must be shaped.

Both respond to the same fear: unchecked power. They simply locate the solution in different places.

The Stakes

As AI systems become embedded in research, education, journalism, and governance itself, their internal constitutions will quietly shape public life--not through force, but through refusal and filtering.

The U.S. Constitution asked how humans should rule themselves.

The Anthropic Constitution asks how humans should rule their machines.

That question used to belong to engineers. It no longer does.

Because when governance moves from courts to code, legitimacy, accountability, and power move with it. The most consequential constitutional debates of the next decade may not happen in legislatures or supreme courts--but in training pipelines and principle lists most people will never read.

The first constitution said, "We the People."

The next ones say something quieter, and more consequential:

"Given these constraints, this is what will be possible."

That is what it now means for code to be law.