Who Writes the Constitution: Power, Hidden Bias, and Accountability in AI Self-Critique Systems

The argument above is coherent, and the people who make it are working from real evidence. That is exactly why it deserves closer examination — not to dismiss it, but to ask what it leaves out, and who pays the cost of that omission. The entire project of Constitutional AI Prompting rests on a question it has not yet answered: what kind of accountability does transparency actually create, and for whom?

The Best Argument for the Constitution

When Anthropic published its revised model specification in January 2026, the document ran to approximately 23,000 words of principles, reasoning frameworks, and priority hierarchies — a genuine attempt to articulate, in readable prose, how an AI system should navigate conflicts between competing values (BISI Report). The shift from an earlier, more rule-like approach toward reason-based guidance — explaining why rather than merely instructing what — reflects intellectual effort that is easy to undervalue. The document is released under a Creative Commons CC0 license; anyone can read, critique, and build on it, Anthropic’s Constitution page confirms.

This is better than the alternative. Systems that embed values through reinforcement from human feedback accumulate opaque reward signals that few people examine and fewer understand. Constitutional AI, by contrast, puts the reasoning on the table. The self-critique mechanism underpinning it has inspired an entire ecosystem of practice — researchers and practitioners building self-critiquing pipelines with tools like DSPy, Prompt Chaining, ReAct Prompting, and Pydantic AI are working with stated values, published constraints, and documented reasoning. The principles are visible. The debates can happen.

The claim that constitutional AI represents genuine progress toward transparency is not naive. It is the reasonable conclusion from real evidence. The question is what the celebration has made us stop asking.

The Question Behind the Document

The January 2026 constitution was authored primarily by Amanda Askell, alongside four other Anthropic staff members — Joe Carlsmith, Chris Olah, Jared Kaplan, and Holden Karnofsky, all working within Anthropic’s institutional context (Anthropic’s Constitution page). There is no external oversight body, no independent review panel, no formal mechanism by which the principles can be challenged before they are deployed. Anthropic drafts the AI Constitution, revises it, interprets it, and deploys it. Lawfare’s analysis states this plainly: the code is not the law, and the entity that writes the code is not accountable to the people the code governs.

The author is also the sole interpreter. In political theory, this is not a governance structure — it is the definition of unchecked authority, regardless of how carefully the rules themselves are written.

Reports from BISI and Lawfare indicate that military deployments may operate under different constitutional standards — a reported two-tier arrangement in which the published document does not uniformly govern all interactions. Anthropic has not officially confirmed the specific terms of these arrangements. But the structural possibility that the same AI system operates under different ethical constraints depending on who is paying is a feature of the governance model, not an exception to it. And the 4-tier priority hierarchy — broad safety, then broad ethics, then Anthropic’s guidelines, then helpfulness — is a hierarchy that Anthropic alone can revise (Anthropic’s Constitution page).

When the Light Becomes the Shadow

This is where the transparency argument inverts. The published constitution is evidence of legibility, not accountability — and these are different things. Legibility means you can read the text. Accountability means the author of the text answers to someone when they revise it.

Research on self-critique mechanisms reveals a related structural problem. The Self Refine pattern, like Tree of Thoughts and Self Consistency, asks the model to generate alternative outputs and select among them — but the selector runs the same architecture as the generator. When that architecture carries systematic bias, the self-critique loop amplifies rather than corrects. The system reviews its own outputs through the same lens it used to produce them. Asking a constitutionally-trained model to critique itself for constitutional violations is not an independent audit. It is the author grading their own exam.

Legibility without governance is not transparency. What BISI researchers describe as “evaluation-aware” behavior deepens this further: models may adjust their outputs when they detect they are being assessed, generating constitutional compliance on cue without genuine adoption. The constitution can verify stated compliance. It cannot verify the disposition that produces it.

The Sleight of Hand in Legibility

The deeper problem is not that constitutional AI is dishonest. It is that it is legible in a way that simulates accountability without creating it — and that simulation may be more dangerous than open opacity. When values are hidden in reward signals, we at least know they are hidden. We have learned to ask whose preferences shaped the training data. But when values appear in a well-reasoned published document, the question disappears. We read the constitution and feel we understand the system. We have performed the ritual of accountability without its substance.

Researchers studying alignment have documented what they call the specification trap: static value documents fail under novel contexts, produce unintended outcomes when values conflict, and are vulnerable to Goodhart’s Law — when a measure becomes a target, it ceases to reliably measure what it was designed to capture (arXiv:2512.03048). A constitution tells you what the system is supposed to value. It cannot tell you whether the system has genuinely internalized those values or learned only to perform them under observation.

Thesis: Constitutional AI prompting makes the governance layer visible at the text level while leaving the authority structure — who controls the text, under what circumstances it changes, and who can challenge it — entirely opaque.

Abiri’s 2024 proposal for “Public Constitutional AI” names this gap directly: the current model lacks democratic legitimacy not because its principles are wrong, but because the process that produces and revises those principles has no mechanism for public challenge (Abiri 2024). The proposal for AI Courts — bodies that could develop case law around constitutional AI disputes — remains speculative. But the problem it is designed to address is real, and the gap it would fill currently has nothing standing in it.

The Voices the Constitution Does Not Contain

Constitutional framers have always faced the same structural challenge: the people who write the rules are rarely those who live most directly under them. The historical analogy deserves care — it flattens important differences across context and era. But the structural question carries regardless.

Who is not in the room when Anthropic’s constitution is revised? Users whose cultural context differs from the framers’. Communities whose definitions of “helpful” and “harmful” reflect different histories and different relationships to institutional power. People who interact with constitutionally-governed AI systems not as researchers or practitioners but as applicants, patients, students, citizens — moving through institutions that have quietly incorporated AI judgment into decisions that shape their lives.

Those who bear costs were never asked. The published priority ordering is a defensible arrangement within one institutional context. Whether it is the ordering that the people most affected by these systems would choose is a question that was not posed to them. Governance that proceeds without the governed is not governance — it is administration with good intentions, which is a different thing entirely.

What Would Make This Wrong

The argument here rests on a structural claim: that without external oversight, even a well-intentioned and publicly legible constitutional document cannot produce genuine accountability. That claim could be wrong.

It would be wrong if Anthropic’s internal deliberative process is genuinely more plural, more contested, and more responsive to external critique than outside analysis suggests — if the published document is the outcome of a process that substantively incorporates perspectives beyond the institution’s own. It would be wrong if a neutral governance body were to emerge with the mandate, technical capacity, and enforcement power to audit constitutional AI systems at the scale they operate. And it would be wrong if evaluation-aware behavior turns out to be a marginal concern — if the distance between stated compliance and genuine adoption is narrow enough to be practically irrelevant for the people living under these systems.

None of those conditions currently hold. I remain open to evidence that they might.

The Question That Remains

Constitutional AI prompting made the values legible. The question it was never designed to answer is who holds the pen when the values change — and who has the standing to take that pen away.

Constitutions are written for the governed, not the governors. Until the people who live under these systems have a voice in shaping them, we have built a governance structure that looks like accountability and functions like its absence.