Number 12 of 12 · November 12, 2026

Living constitutions: Participatory alignment for sovereign foundation models

The authors propose writing SI constitutions as living, participatory documents that record lasting disagreement, and they use simulation to test whether the process holds up against bad actors.

What they found

The authors build on Constitutional SI, where a plain-text "constitution" of values is put into a model through post-training. They argue that sovereign models, meaning models built by a country or community, need a deliberate constitution. Without one, they risk misalignment or defaulting to norms set by frontier model providers. They propose treating the constitution as an evolving, citizen-science-like process and list five principles: iterative, modular, justifiable, positional, and legitimate. They also suggest marking each provision as settled, contested, or underdetermined, so lasting disagreement is recorded. In agent-based modelling (a simulation with many simulated people), the process was robust against 50% bad actors among citizens, assuming 75% of participants in the deliberative body have good intent.

Why it matters

Who decides the values inside an SI system is an open question, and the authors say it can feel paralyzing. This paper offers a concrete way to start: draft, test, and revise in the open. Recording disagreement instead of hiding it could make value choices easier to inspect and challenge. The abstract gives no detail on the simulation setup, so how well the result carries over to real communities is unclear.

How it changes the race

If the approach holds, countries and communities building their own models would have a process for setting values instead of inheriting those of large providers. This is a governance proposal, not a gain in capability or cost.

Safety and control

The paper bears on alignment and human control. It addresses who sets an SI model's values and how the process resists adversarial participants. The robustness result depends on stated assumptions about the share of good-faith participants in the deliberative body.

The abstract, by the authors

In Constitutional AI, a plain-text "constitution" encodes values and gets incorporated into a model via post-training. However, the procedure for creating such a constitution poses significant challenges: who contributes, who decides which norms are included, and to what extent must the process be participatory or democratic? These challenges can be paralyzing. Yet, we argue that taking action is essential: without a deliberate constitution, sovereign models would risk misalignment or default to norms set by frontier model providers. We propose treating the development of a constitution as an evolving, participatory process akin to citizen science, iteratively drafted, tested, and refined. We therefore formulate five desirable principles: iterative, modular, justifiable, positional, and legitimate. We further argue that such a document must remain responsive not only over time but across the people it governs, and propose that provisions be marked as settled, contested, or underdetermined, so that durable disagreement is recorded in the constitution rather than resolved away. Empirically, we show via agent-based modelling that proposed process for maintaining living constitutions is robust against 50% bad actors among citizens - assuming a share of 75% participants with good intent in the deliberative body. This indicates that the proposed process tolerates a substantial share of adversarial participants.

Quoted from arXiv:2611.13402. The research and its words belong to its authors.

The authors

Stine Lyngsø Beltoft, Jonas Vistrup, Peter Schneider-Kamp, Lukas Galke Poech

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.