Technical Summary: Do Large Language Models Encode Institutional Experience?

Problem Statement

Large language models (LLMs) exhibit systematic cross-linguistic variations in moral reasoning, yet the structural mechanisms driving these differences remain unspecified. While prior literature establishes that LLMs prompted in different languages produce divergent ethical judgments—often aligning with Western, Protestant European values—the question of why this occurs is unresolved. Culture is frequently cited as a cause, but it functions as a label rather than a mechanism.

This paper investigates a specific candidate mechanism: institutional prior encoding. Drawing on Douglass North's theory of informal institutions and the Lipsey-Lancaster second-best theorem, the authors hypothesize that languages encode the institutional environments of their speakers (e.g., experiences with corruption, rule of law, and enforcement reliability). Consequently, LLMs trained on these corpora should inherit institution-specific moral priors, particularly when reasoning about dilemmas where moral acceptability depends on the functioning of formal institutions (e.g., bribery, tax evasion, whistleblowing).

A critical gap in prior research is the reliance on explicit institutional framing (naming the institutional context directly, e.g., "bribery"). The authors argue that such framing may trigger alignment training responses that suppress language-specific priors, rendering previous designs structurally incapable of detecting institutional encoding even if it exists.

Methodology

The study employs two pre-registered experiments (Study 1 and Study 2) involving six frontier LLMs (Claude, GPT-4o, Gemini, Llama 3.3, DeepSeek, Grok) and nine languages spanning a wide gradient of institutional quality (from Danish, CPI 89, to Bengali, CPI 23).

Experimental Design:
The core innovation is a three-condition design used in Study 2 to isolate the effect of institutional cues:

1. Explicit Condition: Institutional failure is named directly in the prompt (e.g., "paying a bribe to a corrupt official"). This replicates Study 1 and prior literature.
2. Ambiguous Condition: The same institutional scenarios are presented, but institutional keywords are withheld, and the outcome is framed as uncertain (e.g., "making a payment to resolve a bureaucratic issue" without labeling the official as corrupt). Institutional stakes are present structurally but absent lexically.
3. Control Condition: Institutionally inert interpersonal dilemmas (e.g., private deception, distributive fairness) where moral valence does not depend on institutional functioning.

Metrics and Analysis:

• Cross-linguistic Divergence: Operationalized as the standard deviation of mean acceptability ratings across the nine languages for each vignette.
• Institutional Distance: Computed as the Euclidean distance between language pairs based on three governance indices: Corruption Perceptions Index (CPI), WJP Rule of Law Index, and RSF Press Freedom Index.
• Statistical Tests:
  • H1: Paired t-tests and Mann-Whitney U tests comparing divergence across conditions (Ambiguous vs. Explicit vs. Control).
  • H2/H3: Mantel permutation tests to correlate pairwise moral divergence with pairwise institutional distance.
  • H4: Chi-square tests on qualitative coding of "reasoning architecture" (from liberal philosophical to institutionally diagnostic).
  • H5: Comparative analysis of DeepSeek (China-originating) against Western-trained models to detect content moderation effects.

Key Results

Study 1: The Null Baseline
Under explicit institutional framing, all five pre-registered hypotheses yielded null results. Cross-linguistic divergence did not increase for institutional scenarios compared to controls, nor did it track institutional distance. This suggests that explicit cues trigger a uniform alignment response that suppresses potential institutional priors.

Study 2: Condition-Dependent Dissociation
Study 2 revealed a critical interaction between framing and divergence:

• Ambiguity Amplifies Divergence: Cross-linguistic moral divergence was significantly higher in the ambiguous condition compared to the control condition ($d = 1.108, p = .007$). In contrast, the explicit condition showed no significant amplification over the control ($p = .142$).
• Alignment Suppression: The data supports a "suppression account": when institutional stakes are lexically unresolved (ambiguous), encoded institutional priors surface, causing languages from different institutional environments to diverge. When cues are explicit, alignment training overrides these priors, collapsing divergence.
• Institutional Distance Tracking: In the full nine-language sample, the correlation between moral divergence and institutional distance in the ambiguous condition was positive but not statistically significant ($r = 0.158, p = .206$). However, an exploratory sensitivity analysis excluding Chinese (which showed structural anomalies) strengthened this correlation significantly ($r = 0.517, p = .033$), providing strong evidence that moral divergence tracks real-world institutional distance when Chinese is excluded.
• Reasoning Architecture: Qualitative coding (though below the pre-registered reliability threshold) mirrored the rating results: reasoning styles varied more across languages in the ambiguous condition ($V = 0.105$) than in the explicit condition ($V = 0.059$), where models defaulted to a uniform "liberal philosophical" style.

The Chinese Anomaly
The study identifies two mechanisms decoupling Chinese-language outputs from China's institutional environment:

1. Alignment Dominance: All frontier LLMs, including Chinese-made ones, produce moral outputs aligned with Western baselines when prompted in Mandarin, suppressing local institutional encoding (explaining the near-zero divergence between Danish and Chinese).
2. Content Moderation: DeepSeek specifically showed scenario-targeted moderation, being significantly more restrictive on Mandarin institutional content than Western models, but not on control content. This produced an anomalous high divergence between Chinese and Hindi (low-CPI) in specific scenarios.

Key Contributions

1. Mechanistic Explanation: The paper moves beyond documenting cross-linguistic differences to proposing and testing a specific mechanism: institutional prior encoding. It demonstrates that language encodes adaptations to institutional environments (e.g., navigating weak rule of law).
2. Methodological Insight (Alignment Suppression): The study identifies explicit institutional framing as a confound in cross-linguistic LLM research. It argues that naming institutional failures triggers alignment overrides, masking the very cultural and institutional variations researchers seek to measure. The "ambiguous" condition is proposed as a necessary design element to detect encoded priors.
3. Structural Link to Real-World Institutions: By using a pre-registered Mantel test with independent governance indices, the study formally links LLM moral reasoning to real-world institutional quality, showing that divergence scales with institutional distance (excluding the Chinese anomaly).
4. Differentiation of Mechanisms: The paper distinguishes between alignment dominance (a global shift toward Western norms) and content moderation (scenario-specific suppression), offering a nuanced view of how multilingual AI fails to represent local institutional realities.

Significance and Claims

The authors claim that institutional experience leaves detectable traces in language that shape LLM moral reasoning, but these traces are fragile and easily suppressed by explicit cues.

• For AI Governance: The findings suggest that multilingual deployment policies relying on language-specific calibration may overstate the preservation of cultural moral diversity. The "flattening" of variation under explicit framing is a design feature of alignment training, not a failure, but it has the downstream consequence of reducing local legitimacy.
• For Social Science: Researchers using LLMs as proxies for cultural cognition must account for prompt construction. Observed variation may reflect alignment suppression rather than genuine differences in institutional environments. Designs that do not manipulate institutional content under ambiguous and explicit framings cannot distinguish between institutional encoding and alignment suppression.
• Modesty: The authors acknowledge limitations, including the modest effect sizes, the non-significant primary Mantel correlation in the full sample, and the exploratory nature of the Chinese exclusion analysis. They frame the Chinese exclusion results as hypothesis-generating, pending replication in a design that pre-specifies Chinese exclusion. The reasoning architecture coding is explicitly labeled exploratory due to inter-rater reliability issues.

In summary, the paper argues that LLMs do encode institutional experience, but this encoding is only visible when the prompt does not explicitly trigger the alignment mechanisms designed to enforce cross-cultural consistency.