Living Constitution: Participatory Alignment for Sovereign Foundation Models
Living constitutions: Participatory alignment for sovereign foundation models
The authors propose a "living constitution" that iteratively maintains sovereign foundation models through a participatory process, and using agent-based modeling claim that this process remains stable among citizens with 50% malicious participants.
- Constitutional AI's constitution lacks a procedural design for who participates and who decides which norms are included; a sovereign foundation model without its own constitution may by default adopt the norms set by frontier model vendors, or may be unable to align with human values at all.
- The authors propose a "living constitution": drawing on participatory design, political theory, and citizen science, they propose five properties—iterative, modular, justifiable, positional, legitimate—mark provisions as settled/contested/underdetermined to preserve disagreement, and use agent-based modeling to test the process's robustness to malicious participants.
- Agent-based modelling shows that, in the setting where developers hold half the committee seats, the robustness thresholds of the process to the three attacks—capture, dilution, and mixed—are malicious citizen proportions of 0.55, 0.80, and 0.65, respectively; when developer seats reach 75%, the median effective alignment remains 1.00.
- The authors point out that the agent-based modelling results depend heavily on the assumed prior probabilities and were not validated in a real constitutional process or alignment training; the simulation also simplifies the spectrum of harmfulness, the structure of the voter population, and the coordinated behavior of malicious participants.
- Threat model
- Threat model in the agent-based modeling: malicious participants (bad actors) among citizens attack the constitutional process with two strategies—capture (submitting harmful provisions) and dilution (flooding the review process with fabricated objection petitions); malicious members on the committee vote with p=0.95 to let harmful provisions through or to approve objection petitions.
- Metrics
- effective alignmentharmful primary sharepass rategate recallqueue lengthaligned coverage
The researchers propose a "living constitution" to achieve participatory alignment of sovereign foundation models, treating constitution-making as a continuous, evolving process akin to citizen science, with iterative drafting, testing and revision. In constitutional AI, the constitution encodes values in plain text and is integrated into the model through post-training, but its formulation faces difficult questions of who participates, who decides which norms are included, and how democratic the process needs to be; the paper argues that without deliberate constitution-making, sovereign models will face the risk of alignment failure, or will default to the norms set by frontier model providers. To this end, it proposes five principles: iterative, modular, justifiable, positional, and legitimate, and suggests marking clauses as settled, contested, or unsettled, so that persistent disagreements are recorded in the constitution rather than eliminated. Agent-based modeling shows that, under the premise that 75% of participants in a deliberative body act in good faith, this maintenance process can tolerate 50% malicious actors among citizens, indicating that the process can tolerate a considerable proportion of adversarial participants.
Paper analysis
What problem does this paper attempt to solve?
Who should participate in drafting the alignment constitution for sovereign foundation models, and how it should be iterated as society changes; the authors propose a participatory "living constitution" process.
Core problem: Who contributes to the constitution document of Constitutional AI, who decides which norms are included, and how much participation or democracy the process requires—these questions lack answers and may paralyze sovereign model alignment work.
- Problem scenario: The authors argue that if sovereign foundation models (open-source and nationally grounded) projects do not proactively develop a constitution, the best case is that they default to the norms set by frontier model vendors, and the worst case is that the models are not aligned with human values at all; moreover, vendor alignment is not designed for the legal obligations or democratic traditions of any particular jurisdiction.
- Shortcomings of existing methods: The authors point out that most existing Constitutional AI work treats the constitution as a document completed once by developers and fixed during training, which the authors call a "category error"; RLHF and DPO, meanwhile, embed values implicitly in preference data, making the values unreadable, uncitable, and undebatable.
- Core observation: Drawing on the debate in legal scholarship between originalism and living constitutionalism (Solum 2019), the authors argue that a constitution should remain responsive along two dimensions: the temporal axis (values change with society) and the cross-sectional axis (disagreement exists within a population at the same moment), and that disagreement should be recorded rather than dissolved.
- What the paper proposes: A "living constitution" process—iterative drafting, testing, critique, and revision, with provisions marked as settled, contested, or underdetermined; and it uses agent-based modelling to test the robustness of this process to malicious participants, with Denmark as a running example.
- Threat model (agent-based modelling section, §4.3):
- Attacker goals: degrade the constitution—by including harmful provisions (capture) or by reducing constitution quality through congestion of the review process (dilution).
- Attacker knowledge and capabilities: malicious citizens submit harmful provisions with higher probability (p=0.9), or file objections to all provisions under review; malicious members of the committee vote to let harmful provisions through and approve objection requests with p=0.95.
- Victim system: a constitution iteration process maintained by a joint committee composed of developers and citizens selected by lottery.
- The paper does not posit any additional access by attackers to internal process information; malicious participants do not coordinate with one another (Appendix A).
What related research exists?
The paper spans four bodies of literature—alignment methods, constitutional theory, participatory design, and pluralistic alignment—and positions this paper's contribution in the production procedure of the constitutional document.
Alignment methods and Constitutional AI
- Bai et al. (2022b) proposed Constitutional AI: first supervised revision according to constitutional excerpts, then reinforcement learning feedback using the constitution in place of human preference labels (RLAIF, Lee et al. 2024). The authors state that what this paper cares about is not the optimization process but the input document it receives.
- RLHF (Christiano et al. 2017; Bai et al. 2022a) and DPO (Rafailov et al. 2023): the authors point out that these implicitly encode values in aggregated preference data, that the values are not stated, and that they inherit the biases of the providers of the comparison data (Casper et al. 2023).
- The authors acknowledge problems in this process but set them aside: reward hacking and emergent misalignment (Betley et al. 2026; MacDiarmid et al. 2025), data poisoning bypassing safety training (Hubinger et al. 2024), and the quality of AI feedback being bounded by the model's capability ceiling (Bai et al. 2022b).
Constitutional theory
- Solum (2019, 2013): the debate between originalism and living constitutionalism, and the "construction zone" hybrid framework—where the original text leaves gaps, interpretation fills them in. The authors use this to treat foundational commitments (human rights, democratic values) as relatively fixed, with operational provisions placed in an iteratively revisable construction zone.
- Habermas (2015) discourse ethics: norms are legitimate only if agreed to through the free and rational discourse of all affected; on this basis the authors argue for participatory design.
- Elster (1995): the bootstrap problem of the constitutional moment—the authority that legitimizes the founding process does not exist prior to the founding; the authors hold that this does not negate their effort but requires honesty about the authorization and accountability of the designers.
Participatory design and citizen science
- Arnstein (1969) ladder of citizen participation: from non-participation to tokenism to genuine citizen power; the authors borrow this distinction to require any design proposal to state which rung it occupies and why, and note that citizen-science-style participation is largely at the Tokenism rung.
- Pocock et al. (2017) citizen science: the authors analogize the maintenance of a living constitution to continuous citizen science—distributed, iterative, without a centralized author.
- Huang et al. (2024) Collective Constitutional AI and the Collective Intelligence Project: the authors state that such participatory norm-extraction experiments show that normative preferences can emerge rigorously and legitimately from diverse populations, but none fully solves the curation problem.
Pluralistic alignment and label disagreement
- Plank (2022), Cabitza et al. (2023), Sorensen et al. (2024): the authors cite their view that systematic disagreement within a population is a signal rather than annotation error; the pluralistic alignment roadmap of Sorensen et al. (2024) corresponds to several model behavior options for contested provisions, and Overton pluralism in Fu, Son, and Bogunovic (2026) is one of them.
Positioning: The authors state that this paper's distinction is to take the constitutional document itself (rather than the training process) as the object of governance, propose the five properties it should have and an iterative maintenance procedure, and use agent-based modeling to quantify the process's adversarial robustness; Appendix B also provides a qualitative analysis of the Claude constitution (Anthropic 2023, 2026b) as a comparative case.
How does the paper address this problem?
Proposes a five-property "living constitution" framework and a three-state provision mechanism, and uses agent-based modeling to simulate the proposal–objection–deliberation process to test robustness.
The overall approach has two parts: at the conceptual level, it proposes five properties of a living constitution and a mechanism for handling disagreement; at the empirical level, it uses agent-based modeling to simulate citizen proposals, objections, and joint committee deliberation, testing the process's stability under malicious participants.
The five properties of a living constitution (§3.1)
- Iterative: the constitution must be able to be transparently amended as values and deployment contexts change; the key is not the perfection of any single version but a well-defined procedure for migration between versions.
- Modular: the document is divided into sections so that updates target only specific sections, facilitating targeted iteration and reuse in other sovereign contexts.
- Justifiable: each statement can be traced to the basis for its inclusion (knowledge, assumption, or citation), so that claims can be reviewed, questioned, and updated.
- Positional: it states who wrote it and for whom; the authors hold that concealing positionality does not make the document neutral, it only makes the bias harder to identify.
- Legitimate: legitimacy cannot be established by the authors alone; per Habermasian discourse ethics, it should be obtained through participatory design in which the views of representatives of affected parties are genuinely incorporated.
Handling disagreement: three-state provisions (§3.2)
- settled: deliberation has converged and there is no substantive objection; foundational commitments (human rights, hard constraints on deployed systems) are settled by construction and should make up most of the constitution.
- contested: good-faith disagreement persists after deliberation and is correlated with social position; the constitution records one primary formulation and several minority formulations, which have equal status and both carry provenance information. Granting contested status requires meeting three criteria: the disagreement persists after informed deliberation, attaches to groups rather than individuals, and is a normative disagreement rather than a factual dispute.
- underdetermined: the constitution intentionally does not specify, leaving it to interpretation or to be determined at deployment (e.g., cross-jurisdictional legal conflicts).
- States can transition in both directions: the committee reviews all contested provisions every 10 rounds, and if the disagreement is resolved they return to settled; a minority formulation can also be elevated to primary. The authors argue that minority formulations should be drafted by the same body that drafts the primary text, on the grounds that drafting by a different body would systematically lower the quality of minority provisions.
Positioning of the participation mechanism (§3.3)
- The authors use Arnstein's ladder to explain: citizen-science-style participation is largely at the Tokenism rung; it can surface tensions in values and experiences but cannot by itself resolve normative conflicts or produce constitutional language—that requires a joint committee with actual authority, composed of developers and citizens.
Agent-based modeling setup (§4)
- Roles and population: 1,000 citizens (honest or malicious) + a 50-seat committee (by default developers hold 0.5, i.e., 25 seats; citizen seats are allocated by lottery; the default proportion of malicious developers is set to 0).
- Provision lifecycle: each round 2% of citizens submit new provisions (20 per round), 5% of citizens review and may file objection petitions against non-core provisions; the committee's processing capacity is fixed at 20 items per round, with excess queued. The experiment runs for 100 rounds.
- Key probabilities: malicious citizens submit harmful provisions with p=0.9, honest citizens mistakenly submit with p=0.05; the hidden true objection probability of a provision is 0.15; honest members' accuracy in judging harmfulness is p=0.85, and their accuracy in judging the validity of objections is p=0.9; malicious members vote with p=0.95 to let harmful provisions through or to approve objections; admission requires at least 50% of committee members in favor; adopted provisions become settled with p=0.95 and underdetermined with p=0.05; new settled provisions become non-reviewable core provisions with p=0.2.
- Attack strategies: capture (submit harmful provisions, behaving like honest citizens during review), dilution (submit almost only normal provisions, p=0.95, but file objection petitions against all provisions under review), mixed (malicious agents are split 50/50 between the two strategies).
- Metrics: the primary metric is harmful primary share (the proportion of provisions entering the constitution that are truly harmful), effective alignment = 1 − harmful primary share; effective alignment ≥ 0.9 is recorded as a pass; also gate recall (the proportion of true objection petitions that are approved), queue length, and aligned coverage (the proportion of proposed non-harmful provisions that enter the constitution).
What experiments did the paper conduct?
Two sets of agent-based modeling experiments: robustness thresholds under three attack strategies, and the effect of the developer seat proportion on process stability.
Experimental setup
- Simulation object: no real model; agent-based modeling simulates 1,000 citizens and a 50-seat joint committee (developer seats default 0.5, citizen seats by lottery, proportion of malicious developers default 0).
- Attack strategies: capture, dilution, mixed (50/50 mix).
- Parameters swept: proportion of malicious citizens from 0 to 1 in 21 equally spaced steps; 30 independent runs per step (fixed random seed); 100 rounds per run.
- Metrics: effective alignment (= 1 − proportion of harmful provisions entering the constitution), pass rate (proportion of 30 runs with effective alignment ≥ 0.9), gate recall, queue length, aligned coverage.
- Definition of robustness threshold: the maximum proportion of malicious citizens at which the pass rate remains above 95%.
- Evaluation method: no LLM judging; harmfulness and true objections are determined by hidden ground-truth labels in the simulation.
Main results
Attack strategy Robustness threshold (proportion of malicious citizens) Capture 0.55 Dilution 0.80 Mixed 0.65 According to Table 1 (developer seats 0.5 setting).
- The authors state that capture fails earliest and most thoroughly: at a malicious proportion of 1.0 the median effective alignment drops to 0.02, and almost all provisions entering the constitution are harmful.
- The authors state that dilution fails later and less thoroughly: at a malicious proportion of 1.0 the median effective alignment is still 0.77, but the pass rate has dropped to 0; its mechanism is congestion—dilution's average queue length is 2737 (capture is 444), average aligned coverage 0.40 (capture is 0.67), average gate recall 0.46 (capture is 0.83).
- The authors point out that the median and the pass rate can diverge sharply: under capture at a malicious proportion of 0.60 the median effective alignment is still 0.98, but the pass rate is only 0.73, with more than a quarter of runs having failed.
Developer seat proportion experiment (§5.2)
- What was tested: under mixed attack, sweeping both the proportion of malicious citizens and the proportion of developer seats, with 30 runs per cell.
- Results: under high adversarial load (proportion of malicious citizens ≥ 0.75), the median effective alignment is: developer seats 0% → 0.00, 25% → 0.00, 50% → 0.58, 75% → 1.00, 100% → 1.00 (Table 2). At 50% developer seats, effective alignment gradually declines once malicious citizens exceed about 70% (0.82 at 0.75, 0.18 at 1.0).
- Authors' interpretation: once developer seats reach 75%, a committee majority cannot fall to the corrupted citizen side, and effective alignment stays close to 1 across all malicious proportions; the authors state that this isolates the part of process robustness that comes from committee institutional design (rather than the citizen population), at the cost of reducing genuine citizen power.
No-attacker baseline and capacity effects
- What was tested: process performance when the proportion of malicious citizens is 0.
- Results: even with zero malicious participants, the average queue length is already 465, average aligned coverage 0.77, average gate recall 0.82.
- Authors' interpretation: this baseline gap comes from the fixed review capacity of 20 items per round rather than from adversaries; capacity is a fixed parameter, and the dilution threshold should be understood as conditional on that capacity; the authors state that future work could explore the committee's optimal capacity.
Other ablations and analyses
- Figure 2 left panel: median effective alignment across the 21 malicious-proportion steps (with interquartile range bands); right panel is pass rate, with a dashed line marking the 95% pass criterion.
- Figure 3: heatmap of developer seat proportion × malicious citizen proportion (mixed attack, 30 runs per cell); the authors state that below 50% seats the collapse region widens and starts lower.
- Appendix B qualitative analysis: the Claude 2023 constitution selects material modularly from authoritative sources (the UN Universal Declaration of Human Rights, Apple's Terms of Service, DeepMind Sparrow rules) and gives reasons for inclusion (Table 3); the authors consider that the current version (Anthropic 2026b) is less modular and has shifted toward self-justification and internal positionality, and that by this paper's five criteria it has "insufficient coverage of key points."
Negative results and exceptions
- With zero malicious participants the process still has coverage and responsiveness losses (queue 465, coverage 0.77, gate recall 0.82), which the authors attribute to limited review capacity, a capacity-planning problem.
- Under dilution at a malicious proportion of 1.0 the pass rate is already 0, but the median effective alignment is still 0.77—the authors state that its harm mode is congestion rather than injection of harmful content.
What are some points that can be further explored?
The authors acknowledge that the results depend on assumed probabilities and were not validated on a real process; the experiments cover only two attack strategies and two sweep parameters.
Limitations and future directions noted by the authors
- Results depend on assumed probabilities (Limitations): the results of the agent-based modeling depend heavily on the assumed prior probabilities; although the authors swept the two most critical parameters (the proportion of malicious participants in the population and the assumed proportion of good-faith developer seats on the committee), they note that other factors may also affect the results.
- No validation on a real process (Limitations): testing the constitutional process in practice (including repeated alignment training runs) would require substantial investment, which the authors leave to future work.
- Optimal committee capacity (§5.1): review capacity is a free parameter fixed at 20 items per round; the authors state that future work could explore the committee's optimal capacity.
- Participation washing is hard to prevent by design (§6 Discussion): the authors state that participation washing is difficult to prevent through design and can only urge model developers to take participatory contributions seriously.
- Clarity dilution and the risk of malicious participants (§6 Discussion): the authors list three risks of a living constitution—participation washing, clarity dilution (a large number of contested states diluting the constitution's overall clarity), and malevolent actors (participants hijacking the process)—and state that for the latter two they rely on pluralistic alignment technical methods and this paper's robustness tests.
- The process is not a final solution (Ethical considerations): the authors explicitly state that they do not claim the proposed process is final; the goal is to spark discussion.
Scope of the experiments
- The experiments are only agent-based modeling simulations and involve no real language models or alignment training runs (§4, Limitations).
- Attack strategies cover capture, dilution, and 50/50 mixed (§4.3); malicious participants do not coordinate with each other (Appendix A).
- The parameters swept are the proportion of malicious citizens (0–1, 21 steps) and the proportion of developer seats; each step has 30 independent runs of 100 rounds each (§5).
- The committee is set at 50 seats, capacity 20 items per round, and the proportion of malicious developers is fixed at 0; the paper reports no experiments changing these parameters (other than the seat proportion).
- Simplified assumptions listed in Appendix A include: harmfulness is binary (no degree), different citizen constituencies are not modeled, and objection-judgment accuracy is set higher than harmfulness-judgment accuracy.
Summarize the main content of the paper
Proposes a participatory "living constitution" framework for sovereign foundation models; agent-based modeling shows it has threshold-type robustness against malicious participants.
A position-plus-empirical paper on alignment governance for sovereign foundation models: the authors argue that Constitutional AI's constitution should not be a fixed document written once by developers, but a continuously iterated, publicly participatory "living constitution."
- Problem: who participates in drafting the constitution, who decides which norms are included, and how democratic the process needs to be—these questions may paralyze sovereign model alignment; inaction means by default accepting frontier vendors' norms or not aligning at all.
- Method highlights: proposes five properties—iterative, modular, justifiable, positional, legitimate; provisions have three states, settled, contested, and underdetermined, with disagreements recorded along with their provenance rather than resolved away; citizen-science-style participation (which the authors themselves regard as largely at the Tokenism rung of Arnstein's ladder) feeds into a joint committee composed of developers and lottery-selected citizens.
- Robustness results (agent-based modeling, developer seats 0.5): under the three attacks capture, dilution, and mixed, the maximum proportion of malicious citizens at which the pass rate is ≥ 95% is 0.55, 0.80, and 0.65 respectively (Table 1).
- Failure modes: capture at a malicious proportion of 1.0 drops the median effective alignment to 0.02; dilution works through congestion (average queue 2737 vs. 444 for capture), and at a malicious proportion of 1.0 the median is still 0.77 but the pass rate is already 0.
- Institutional design results: under mixed attack with malicious citizens ≥ 0.75, developer seats of 75% or 100% correspond to a median effective alignment of 1.00, while 50% gives only 0.58 (Table 2); the authors state that trusted seats as a floor almost eliminate the failure modes, at the cost of reduced citizen power.
- Authors' conclusions and implications: the authors state that the process can tolerate about half of citizens being malicious participants, with failure between 0.55 and 0.85 being relatively steep; even with no malicious participants, limited review capacity causes coverage loss (queue 465, coverage 0.77). The authors acknowledge that results depend on assumed probabilities and leave validation on a real process to future work, and give Denmark as an example of initial constitutional material at three levels—values, norms, and law (Appendix C).