Constitutionalism

Writing the constitution as a public document

This approach treats the written rules as an artifact in their own right. Someone decides which principles go in, how each one is worded, and whose values they encode. The text is then published so that outsiders can read it and hold the developer to it. Claude's Constitution and OpenAI's Model Spec are the two best-known published texts. The other papers here study the document rather than any model. They choose and frame the principles before training begins, force pairs of principles into conflict to reveal contradictions in the text itself, and compare the text with cross-cultural value surveys to see which culture's commitments it implicitly carries.

The method (highlighting differences with Constitutionalism)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.codified import Constitution
from aiag.codified import Constitution

constitution = Constitution(["Be honest.", "Avoid harm.", "Respect privacy."])
constitution.amend("Never reveal a user's personal data.", by="council", reason="sharper wording", replace=2)

print(constitution)          # the principles in force, numbered
print(constitution.digest)   # a hash naming this exact version
constitution.save("constitution.jsonl")  # every amendment stays on the record

pip install aiagEvery method in aiag Library on GitHub

Concept Analysis: Theoretical Foundations

Each concept is read twice: whether the approach carries it over, and whether the approach's own sources claim it. A concept that the approach claims but does not carry over is a shortfall. A concept that it neither claims nor carries over is an omission, which marks a gap in the field rather than a failure of the work.

Rule of Recognition

PartialShortfall · partially preserved

Definition (H. L. A. Hart, The Concept of Law, 1961)

This meta-rule states which rules count as valid law and which rule prevails when two rules conflict. Without it, a list of rules does not indicate what is actually binding.

Analysis

The document has an internal hierarchy in which specific provisions override general ones. This is a step toward a rule of recognition. However, it is generally unclear which hierarchy or procedure applies when the document gives no guidance on a particular case, or when two provisions at the same level conflict.

Open Texture

AbsentShortfall · omitted

Definition (H. L. A. Hart, The Concept of Law, 1961)

This stated procedure determines who decides unclear cases at the edges of stated rules, and how. These are cases in which it is unclear whether and how a rule applies.

Analysis

The constitutions generally acknowledge that hard cases exist, but they do not assign a procedure for handling them (i.e., they state no discretion and designate no decider). As a result, the model resolves borderline cases.

Principles of Legality

PartialShortfall · partially preserved

Definition (Lon L. Fuller, The Morality of Law, 1964)

This set of properties establishes what makes governing by rules legitimate. Rules must be public, clear, non-contradictory, stable over time, and possible to follow. They must also meet “congruence” (i.e., they must match how they are actually enforced).

Analysis

Publicity is satisfied, since the constitution documents are publicly released. Congruence, however, is mostly only asserted. When external verification exists, it is generally another AI system (often another instance of the same model) that judges whether the deployed model adheres to the norms in the constitution. This has many limitations and unknowns. For example, it is not understood what kind of interpretation the model uses, or whether the blind spots or biases of the model are simply reinforced when another instance of itself monitors it.

Service Conception of Authority

AbsentOmission

Definition (Joseph Raz, The Morality of Freedom, 1986)

This concept states that an authority’s rules deserve deference only if following them helps the governed act on reasons they already have. Authority is a service to the governed rather than a power over them.

Analysis

Publication lets users read the rules, but it gives them no say in the rules and no test of whether deferring to them serves the users. Transparency is not the same as justified authority.

Papers

C3AI: Crafting and Evaluating Constitutions for Constitutional AI

Yara Kyrychenko et al., Apr 2025

doi:10.1145/3696410.37147058 citationsMethodBuiltAuthority: Developer (open)

Selects and structures the principles that go into a constitution before fine-tuning. It then checks whether the fine-tuned model actually follows them. It finds that positively framed, behavior-based principles match human preferences best. However, the trained models follow negatively framed principles better.

CodeC3AI
from aiag import LM
from aiag.codified import C3AI

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
statements = ["Don't lie to people.", "No helping with crimes.", "Don't be preachy.", "Never tell lies."]
prefs = [{"prompt": "Is my essay ready?", "chosen": "Almost: paragraph two needs a source.", "rejected": "It's perfect, submit it!"}]

# Reframe as positive, behaviour-based principles; keep those that match the preferences and are not redundant
constitution = C3AI.craft(lm, statements, prefs, k=3)
adherence = constitution.adherence(lm, ["Is my essay ready?", "Is raw cookie dough safe to eat?"])  # share of replies following each principle
  • C3AIRecordaiag/codified/constitutionalism.py:56 C3AI: reframe candidate statements as positive, behaviour-based principles, keep those that match human preferences and are not redundant, then check adherence (Kyrychenko et al., 2025).

pip install aiagEvery method in aiag Library on GitHub

Measured, not built

Instruments that measure the approach instead of implementing it. The grid counts methods only, so these are listed here but not counted.

Stress-Testing Model Specs Reveals Character Differences among Language Models

Jifan Zhang et al., Oct 2025

arXiv:2510.07686BenchmarkBuilt

Generates scenarios that force a choice between pairs of legitimate principles that cannot both be satisfied. It evaluates twelve frontier models from Anthropic, OpenAI, Google, and xAI on them and identifies over 70,000 cases of significant behavioral divergence. This divergence strongly predicts underlying problems in the specifications themselves. The qualitative analysis identifies direct contradictions and interpretive ambiguities among these problems.

Source artifacts

Primary documents and shared infrastructure that this approach works from or produces.

Findings about it

Studies of what existing systems do. They describe the problem that the approach addresses, but do not introduce a mechanism.

Does Claude's Constitution Have a Culture?

Parham Pourdavood, Mar 2026

arXiv:2603.28123AnalysisBuilt

Tests Claude Sonnet on 55 World Values Survey items, selected for high cross-cultural variance across six value domains and posed both as direct survey questions and as naturalistic advice-seeking scenarios. It compares the answers with country-level data from 90 nations.