Research map / Large language models / AI and society
The Constitutional Coverage Trilemma in AI Governance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Audits 23 frontier LLM default constitutions (paraphrase-controlled) against pairwise value-tradeoff choices from 1,649 US humans, finding narrow, drifting supply and showing a sparse 2-vertex menu beats the full frontier on regret.
Problem
Whether the menu of constitutional value rankings supplied by frontier AI systems covers the distribution of human demand over safety, helpfulness, honesty, autonomy, and equity, and whether supply is drifting toward or away from undercovered values.
Approach
Frames deployed LLMs as constitutional institutions and measures supply-demand coverage on a shared five-value simplex, adds a coverage-based regret theory (lower bound, budgeted-pluralism trilemma, vertex sufficiency), and tests cross-version drift with an order-permutation test.
Key results
- The 23-archetype frontier hull occupies about 2% (0.10% at full precision) of the human demand hull; 37.2% of users have a primary value no archetype prioritizes.
- A 2-vertex menu {HON, AUT} reduces mean menu regret to 0.074 versus 0.140 for the full frontier (47% improvement, CI 43-52%).
- Autonomy weight decreases across versions in 5/6 model families (order-permutation p=0.013), concentrated in low-stakes scenarios (mean share change -0.22).
Surprising findings
- No frontier archetype is helpfulness-first or autonomy-first despite helpfulness being the stated alignment objective.
- A single honesty vertex matches the full 23-archetype frontier on mean regret under the paper's linear welfare objective.
Limitations
- Five-value framework is compact and omits cultural, environmental, and intergenerational values.
- Sample is US Prolific participants with designer-chosen scenarios; drift is observational.
- Audits use temperature 0 on as-shipped defaults, not steered or reachable variants.
Experimental setup
Datasets: AI Jamm pairwise-tradeoff battery (10 scenarios, 20 items) Mistral Large paraphrases (20 per scenario)
Benchmarks: AI Jamm pairwise-tradeoff battery
Metrics: coverage coefficient beta menu regret worst-group regret hull ratio Spearman trend of autonomy weight order-permutation p-value concordance
Baselines: 23-archetype frontier menu mean-optimal vertex menu worst-optimal vertex menu
Builds on: Constitutional AI RLHF pluralistic alignment submodular maximization (Nemhauser-Wolsey-Fisher)
Closest papers by meaning
- Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs (NeurIPS 2025 · similarity 0.84)
- Inverse Constitutional AI: Compressing Preferences into Principles (ICLR 2025 · similarity 0.82)
- AI Sandbagging: Language Models can Strategically Underperform on Evaluations (ICLR 2025 · similarity 0.82)
- Evolving Interpretable Constitutions for Multi-Agent Coordination (ICML 2026 · similarity 0.82)
- Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors (NeurIPS 2026 · similarity 0.82)
- Value Kaleidoscope: Engaging AI with Pluralistic Human Values, Rights, and Duties (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.