An AI constitution that keeps disagreement on the record

Governing the rulebook

Constitutional AI uses written principles to guide a language model’s post-training, including AI-generated criticism and revision of responses. Making those principles explicit creates something people can inspect and challenge. It does not, by itself, establish who has authority to choose them.

Stine Lyngsø Beltoft and colleagues focus on sovereign foundation models: nationally situated AI initiatives, with Denmark as their running example. They argue that leaving constitutional design untouched risks inheriting commercial providers’ priorities rather than deliberately considering the people a model will serve.

Their proposal is a “living constitution”: a document repeatedly drafted, tested, criticized and revised through citizen input and a deliberative body with decision-making authority. The authors acknowledge that collecting opinions is not the same as sharing power. Public participation can become window dressing if the committee is free to ignore it.

Disagreement becomes part of the document

The framework sets out five desirable properties. A constitution should be iterative, modular, justifiable, positional and legitimate. In practical terms, changes should be traceable; sections should be independently revisable; provisions should explain their grounding; authorship and intended constituencies should be visible; and affected people should have meaningful representation.

Provisions would also carry one of three statuses:

• Settled: deliberation produces a commitment without material objection.
• Contested: informed, good-faith disagreement persists between constituencies. The document preserves both a primary formulation and minority alternatives, with their reasoning and provenance.
• Underdetermined: the constitution deliberately leaves a question for interpretation or deployment, such as reconciling legal requirements once the applicable jurisdiction is known.

Not every objection qualifies as constitutional dissent. The proposed committee would check that disagreement survives informed discussion, concerns a constituency rather than an isolated individual, and is normative rather than a factual question evidence could resolve.

The proposal does not prescribe one model behavior for every contested rule. Depending on the context, a system might explain several reasonable positions, support different legitimate configurations, or defer to a human. Those possibilities are design options, not demonstrated capabilities of a model trained in this study.

Stress-testing participation with simulated actors

The empirical work is an agent-based simulation, not a trial with real citizens or a newly aligned language model.

Each run models 1,000 citizens and a fixed committee of 50 members over 100 rounds. In the baseline arrangement, half the committee seats belong to developers, who are assumed never to be malicious; the remaining seats are randomly selected citizens. Citizens submit proposals and claims that existing provisions should be marked contested. The committee can process only 20 proposals or claims per round.

The researchers examine three attack conditions. “Capture” actors try to introduce harmful provisions. “Dilution” actors flood the process with disagreement claims. A mixed condition combines both strategies.

The simulation assigns provisions hidden harmfulness labels and gives honest committee members fixed probabilities of judging correctly. Its main outcome, “effective alignment,” is the non-harmful share of settled and contested provisions—not a measure of an actual AI system’s behavior.

For each condition and malicious-citizen fraction, the researchers conduct 30 runs. Their robustness criterion requires at least 95% of runs to achieve effective alignment of at least 90%. With the baseline committee, the reported malicious-citizen thresholds are 55% for capture, 65% for mixed attacks and 80% for dilution.

A clean document can still have a clogged process

Those thresholds do not mean the process remains healthy in every respect. Dilution produces larger backlogs and prevents more legitimate proposals from reaching the document within the simulation. Even with no malicious participants, the committee ends with an average queue of 465 items: its fixed capacity cannot keep up.

Committee composition is another major qualification. Giving assumed-benign developers 75% or more of the seats keeps median effective alignment close to one throughout the tested citizen-corruption range. That protection comes with an explicit political cost: less citizen power.

What remains unproven

The preprint, posted on November 10, 2026, offers a governance proposal and conditional simulation evidence, not real-world validation. Attackers act independently, harmfulness is binary, and voting accuracy is supplied as an assumption. The simulation does not model distinct social constituencies, despite their importance to the dissent framework.

No repeated model-training experiment establishes whether these living documents produce better AI behavior. The contribution is narrower: a concrete way to preserve disagreement, alongside a demonstration that trusted decision-makers and adequate review capacity strongly shape its simulated resilience.