LLM labels of EU AI Act submissions are nearly perfectly reproducible but fail to match what stakeholders say in surveys; businesses tend to amplify AI-risk language in public consultations while regulators and some non-business groups do not.

Main Finding

LLM annotations of public consultation texts can be extremely reproducible while failing to validly recover the same latent constructs measured in surveys. In the European Commission AI Act consultation, Qwen3.5 annotations showed near-perfect inter-run reproducibility (ICC > 0.99) but weak convergence with stakeholders’ survey-reported AI concerns (Pearson r ≈ 0.03–0.18; Lin’s CCC < 0.08). Divergences are systematic across institutional and geographic contexts (e.g., business associations amplify AI-risk language in public submissions relative to their survey responses).

Key Points

• Reproducibility ≠ construct validity: High LLM inter-run agreement (ICC(C,1) 0.994–0.996; ICC(C,k) > 0.999) did not imply that the LLM-derived scores measured the same latent constructs elicited by surveys.
• Poor convergence with surveys:
  • Safety concern: r = 0.029 (p = 0.59)
  • Rights concern: r = 0.176 (p = 0.001)
  • Explainability: r = 0.097 (p = 0.072)
  • Lin’s concordance coefficients near zero (0.013–0.080).
• Large systematic bias: Bland–Altman analysis showed limits of agreement spanning ~±4 points on a 0–5 scale; standardized mean differences (Cohen’s d) ≈ 1.0–1.18.
• Systematic heterogeneity by stakeholder type:
  • Business associations: positive divergence (public texts show higher AI-risk emphasis than their surveys; mean gap ≈ +1.0).
  • Public authorities and several non-business groups: smaller or negative divergences.
• Geographic patterning: positive spatial autocorrelation of divergence (Moran’s I = 0.347, p = 0.036), though the authors caution sensitivity to multiple-testing and sample limitations.
• Downstream robustness: despite divergence, survey-reported concern remains strongly associated with support for explainability across divergence levels (tested with Double/Debiased Machine Learning), indicating distinct but still informative relationships.
• Conceptual caution: public consultation texts are institutionally situated artifacts (audience targeting, reputation management, access channels). Treating text-derived proxies as error-free measures of private or survey-attested attitudes can induce collider or selection biases in downstream analyses (Text is a collider: T → Text ← Y).

Data & Methods

• Data: Public submissions to the European Commission AI Regulation consultation (Better Regulation Portal), N = 857 submissions collected 2020-02-20 to 2021-04-27 across 3 rounds. Validation sample: N = 348 submissions with both free-text and matched structured survey items (0–5 scales for safety concern, rights concern, explainability).
• LLM annotation:
  • Model: Qwen3.5-397b-A17b.
  • Task: annotate each document on three continuous 0–5 scales (safety_concern, rights_concern, explainability_trust); prompt enforced strict JSON and included scale anchors and binary flags (e.g., regulatory, biometric mentions).
  • Inter-run checks: five independent annotation runs; near-identical aggregate scores (mean vs median difference < 0.01).
• Validation diagnostics:
  • Convergent/discriminant checks: Pearson correlation, Lin’s CCC, Bland–Altman limits, Cohen’s d, ICCs for cross-setting agreement.
  • Examination of systematic divergence across institutional groups and countries.
• Spatial analysis:
  • Moran’s I with queen contiguity weights (shared land borders), permutation inference (B = 9,999), primary sample of 23 countries.
• Causal/robustness check:
  • Double/Debiased Machine Learning (DML) to estimate association between high survey-reported safety concern and explainability support.
  • Implementations: LightGBM for nuisance estimation with 5-fold cross-fitting; Causal Forest DML parameters reported (Ntrees = 1000, minleaf = 10).
• Conceptual framing: Directed acyclic graph showing Text as a downstream collider of multiple latent constructs (T and Y) and unobserved confounders U — conditioning on text-derived features can open spurious paths.

Implications for AI Economics

• Measurement risks in text-based variables: LLM-derived textual measures may reflect public-facing, institutionally mediated communication (framing, signaling, audience targeting) rather than private or survey-attested preferences. Econometric analyses that treat such measures as direct proxies for latent attitudes risk mis-specification and biased estimates.
• Collider and omitted-variable bias: Using text-derived features as covariates/controls without validating construct correspondence can induce collider stratification, opening spurious associations between explanatory variables and outcomes. Researchers should explicitly model the measurement process and possible downstream selection/collider structures.
• Policy inference and stakeholder analysis: Public statements scraped from consultations or other regulatory submissions can systematically over- or understate concerns by stakeholder type (e.g., businesses amplify risk rhetorically). Policymakers and analysts should not equate public rhetorical intensity with private preferences or likelihood to support/comply with regulation.
• Cross-country and institutional heterogeneity matter: Cultural, reputational, and access-channel differences shape public communication. Comparative AI-economics work that uses text measures must test for—and, where appropriate, model—heterogeneous measurement error across organizations and countries.
• Recommended empirical practices:
  • Validate LLM text measures against external criteria (matched surveys, behavioral signals) before using them as proxies in causal models.
  • Distinguish measurement targets: if the interest is public communication/positioning, LLM scores can be valid indicators; if the interest is underlying private attitudes, additional validation and correction are required.
  • Report both reproducibility and construct validity metrics separately; high inter-run agreement alone is insufficient.
  • Conduct subgroup analyses (by stakeholder type, country) and sensitivity tests to detect systematic divergence.
  • Where text-derived measures are used as controls, consider causal identification strategies robust to measurement-induced collider bias (e.g., instrumenting, explicit measurement-error models, bounding approaches).
  • Combine sources (surveys, private channels, behavioral data) and exploit methods like DML while maintaining external validation of the target construct.
• Practical takeaway for AI-economics researchers: LLMs are powerful scalers of text, but their outputs should be interpreted as indicators of communicative expression in context. Use them thoughtfully—validate, test heterogeneity, and avoid treating reproducible LLM annotations as automatically valid measures of latent economic or policy attitudes.