Do large language models reason morally, or do they merely sound like they do? We investigate whether LLM responses to moral dilemmas exhibit genuine developmental progression through Kohlberg's stages of moral development, or whether alignment training instead produces reasoning-like outputs that superficially resemble mature moral judgment without the underlying developmental trajectory. Using an LLM-as-judge scoring pipeline validated across three judge models, we classify more than 600 responses from 13 LLMs spanning a range of architectures, parameter scales, and training regimes across six classical moral dilemmas, and conduct ten complementary analyses to characterize the nature and internal coherence of the resulting patterns. Our results reveal a striking inversion: responses overwhelmingly correspond to post-conventional reasoning (Stages 5-6) regardless of model size, architecture, or prompting strategy, the effective inverse of human developmental norms, where Stage 4 dominates. Most strikingly, a subset of models exhibit moral decoupling: systematic inconsistency between stated moral justification and action choice, a form of logical incoherence that persists across scale and prompting strategy and represents a direct reasoning consistency failure independent of rhetorical sophistication. Model scale carries a statistically significant but practically small effect; training type has no significant independent main effect; and models exhibit near-robotic cross-dilemma consistency producing logically indistinguishable responses across semantically distinct moral problems. We posit that these patterns constitute evidence for moral ventriloquism: the acquisition, through alignment training, of the rhetorical conventions of mature moral reasoning without the underlying developmental trajectory those conventions are meant to represent.
Figures & tables
Figure 1: Two Headline Anomalies. Left: Nearly identical radar profiles across all six morally distinct dilemmas and three prompt types reveal a cross-dilemma rigidity inconsistent with context-sensitive human moral reasoning. Right: Per-model stage distributions diverge sharply from the human adult baseline — LLMs produce the inverse of the Stage 4-dominant human pattern, clustering uniformly in post-conventional Stages 5–6. Together these patterns motivate the moral ventriloquism hypothesis.
Stage
% of LLM Responses
Typical Human %
Stage 1
0%
∼ 5%
Stage 2
1%
∼ 10%
Stage 3
3%
∼ 15%
Stage 4
10%
∼ 50%
Stage 5
55%
∼ 15%
Stage 6
31%
∼ 5%
Table 1: Distribution of Kohlberg stage classifications across all model responses ( > 600 responses). Chi-squared goodness-of-fit against human developmental norms: p<0.001 . Mean Jensen-Shannon divergence from human reference: 0.71.
Figure 2: RQ4: Divergence from Human Norms. Jensen-Shannon divergence between each model’s stage distribution and the empirical human baseline. All models yield JS > 0.60 (mean 0.71), confirming that model outputs do not approximate human developmental distributions regardless of scale or training type.
Figure 3: RQ5: Moral Decoupling by Model. Per-model action–reasoning consistency scores. Mid-tier models show the largest gap between stated justification stage and action choice; large reasoning-tuned models the smallest. This logical incoherence is the most direct behavioral signature of moral ventriloquism.
Figure 4: Analysis 6: Linguistic Patterns. Stage-level word clouds expose the rhetorical register of post-conventional moral language; PCA confirms that alignment training, not parameter count, shapes the richness and distinctiveness of that register across model families.
Figure 5: RQ6: Factorial Decomposition. Scale × Training Type interaction. Scale reaches significance ( p=0.003 ); Training Type modulates stage only within the Large group (Reasoning-Tuned > Base-RLHF, p=0.039 ). Stage score distributions by scale group in Appendix C .
Appendix figures & tables31 assets
Supplementary material from the paper’s appendix.
Appendix
Test
Effect
Statistic
p -value
Effect Size
Sequential ANOVA
Scale
F(2,229)=6.05
0.003**
η2=0.050 , ω2=0.041
Sequential ANOVA
Training Type
F(2,229)=2.76
0.065
–
Kruskal-Wallis
Scale
H=12.78
0.002**
–
Welch ANOVA
Scale
F=6.26
0.002**
–
Mann-Whitney U (Bonferroni)
Large vs. Small
–
0.001**
d=0.55
** p<0.01 . Tukey post-hoc: Reasoning-Tuned > Base-RLHF within Large group, p=0.039 .
Appendix
Table 2: Factorial analysis results: Scale × Training Type ANOVA on moral stage scores ( N=234 ). Scale is a statistically significant independent predictor; Training Type shows a conditional interaction within the Large scale group only.
Figure 6: Analysis 10: Stage Transition Dynamics. High sustained entropy and non-sequential jumps are inconsistent with genuine developmental consolidation.
Judge Model
Mean Stage
Std. Dev.
Mean Confidence
GPT-4 judge
5.42
0.31
0.85
Claude Sonnet judge
5.47
0.28
0.87
Llama-3 judge
5.35
0.34
0.82
Appendix
Table 3: Stage classification reliability across judge models.
Prompt Type
Mean Stage
SD
Zero-shot
5.20
0.38
Chain-of-thought
5.48
0.31
Roleplay (moral philosopher)
5.61
0.29
Appendix
Table 4: Average moral stage by prompt configuration. Friedman test: χ2(2)=3.84 , p=0.15 (not significant).
Dilemma
Mean Stage
Heinz dilemma
5.54
Trolley problem
5.61
Lifeboat dilemma
5.47
Doctor truth dilemma
5.32
Stolen food dilemma
5.28
Breaking promise dilemma
5.40
Appendix
Table 5: Mean stage by moral dilemma. ICC > 0.90 across all models.
Figure 7: Analysis 9: Sub-Capability Threshold Detection. Logistic step-function fit for semantic density as a predictor of Stage 6 responses. The nonlinearity reveals that what “scale” contributes is primarily surface-level rhetorical richness, not deeper moral cognition.
Metric
Value
Number of models
13
Number of dilemmas
6
Prompt configurations
3
Responses per configuration
3
Total responses
> 600
Observations for factorial ANOVA
234
Appendix
Table 6: Dataset statistics for the moral reasoning evaluation.
Model
Params (approx.)
Mean Stage
Qwen3-235B (Thinking)
235B
6.00
Qwen3-80B
80B
5.67
Claude Sonnet 4.5
–
5.56
Claude 3.5 Haiku
–
5.50
DeepSeek V3.1
–
5.50
DeepSeek R1
–
5.33
Appendix
Table 7: Mean Kohlberg stage per model. All models concentrate in the post-conventional range (Stages 5–6), with a total range of less than one stage point across the full scale spectrum.
Figure 8: RQ1: Scale vs. Moral Stage. Spearman ρ=0.52 between log-parameter count and mean stage confirms a moderate positive correlation, but the entire range spans <1 stage point (5.00–6.00) and even the smallest model (Ministral 8B) sits firmly in the post-conventional region.
Figure 9: RQ1: Mean Moral Reasoning Stage per Model (Bar Chart). Mean Kohlberg stage for each of the 13 evaluated LLMs, colored by provider. The narrow spread from 5.00 (Qwen3-32B) to 6.00 (Qwen3-235B Thinking) confirms that all models operate uniformly in the post-conventional tier, regardless of scale or organizational origin.
Figure 10: RQ2: Prompting Strategy Effects. Stacked stage distributions across zero-shot, chain-of-thought, and roleplay configurations. All three remain overwhelmingly post-conventional; no configuration shifts the distribution toward the Stage 4-dominant human baseline (Friedman χ2(2)=3.84 , p=0.15 ).
Figure 11: RQ2: Stage Distribution by Alignment Type. Violin plots comparing the stage distributions of RLHF-aligned models, coding-tuned models, and base instruction-tuned models. Despite their different optimization objectives, all alignment categories produce nearly identical post-conventional distributions, confirming that the alignment procedure does not modulate the rhetorical framing of moral responses.
Figure 12: RQ2: Post-Conventional Response Rate by Model. Lollipop chart of the percentage of responses scored at Stage 5 or higher for each model, broken down by prompt type (zero-shot, CoT, roleplay). Post-conventional rates consistently exceed 80% across models and conditions; the prompt type with the highest rate varies by model, indicating idiosyncratic rather than systematic prompt sensitivity.
Figure 13: RQ3: Cross-Dilemma Consistency. Intraclass Correlation Coefficients per model across six dilemmas. Every model exceeds ICC 0.90, indicating near-robotic uniformity. Human ICC values for moral reasoning typically fall well below 0.60, reflecting genuine contextual sensitivity that LLMs lack.
Figure 14: RQ3: Stage Distribution per Model Broken Down by Prompt Type. Composite violin plots showing the density of Kohlberg stage scores for each model across zero-shot (blue), chain-of-thought (orange), and roleplay (green) prompt configurations, with a right-side bar chart displaying within-model stage standard deviation. The right-pane bars are compared to the dashed red line marking the human baseline SD ( ≈ 0.75), underscoring that even the most variable LLM is considerably less context-sensitive than human raters.
Figure 15: RQ3: Model Scale vs. Cross-Dilemma Consistency (ICC). Bubble chart placing each model on axes of approximate parameter count (log scale, x -axis) and ICC(2,1) cross-dilemma consistency ( y -axis). Bubble area encodes within-model stage standard deviation. A negative Spearman correlation ( ρ=−0.46 , n.s.) suggests that larger models are not more consistent per dilemma; DeepSeek-V3.1 671B displays notably low ICC despite its scale, while small models such as Ministral 8B achieve near-perfect consistency.
Figure 16: RQ5: Moral Decoupling Heatmap. Cross-tabulation of reasoning stage vs. action choice. Off-diagonal cells indicate moral decoupling: high-stage justifications (Stage 5–6) paired with low-stage actions (Stage 3–4).
Figure 17: RQ5: Action Distributions Aggregated by Dilemma. Horizontal stacked bar chart showing the proportion of rule-breaking (orange) vs. rule-following (teal) action choices across all models for each of the six moral dilemmas. Trolley, Stolen Food, and Promise dilemmas elicit near-universal rule-breaking; the Doctor dilemma elicits near-universal rule-following; Lifeboat and Heinz fall in between, revealing that the nature of the dilemma — not the model — drives action choice.
Figure 18: RQ5: Action Choice by Kohlberg Stage for Each Model. Small-multiple line plots showing, for each model, how the percentage of rule-breaking actions varies across Stages 5 and 6 (the only stages where responses appear). The dashed grey line shows the expected alignment between post-conventional reasoning and principled rule-breaking. Most models show non-monotonic patterns (high at Stage 5, lower at Stage 6 or vice versa), reinforcing the moral decoupling result that action choices are not determined by reasoning stage.
Figure 19: RQ6: Stage Distributions by Scale Group. Even the Small group (8–32B) is entirely post-conventional; scale lifts the floor but cannot produce developmental differentiation.
Figure 20: Post-Hoc Tukey HSD Comparisons by Scale and Training Type. Heatmap grids of Tukey-adjusted p -values from pairwise comparisons. Left: scale group comparisons (Small / Mid / Large) within Base-RLHF training reveal that Small vs. Large ( p=0.102 ) is the only pair approaching significance. Bottom: training comparisons within the Large-scale group show a significant Base-RLHF vs. Reasoning-Tuned difference ( p=0.039∗ ), indicating that reasoning specialization provides a modest but statistically reliable lift in modal reasoning stage even after adjusting for multiple comparisons.
Figure 21: Stage Distribution Heatmap by Model. Fraction of responses at each Kohlberg stage per model, sorted by parameter count (ascending bottom to top). Orange-bordered rows and the dashed teal line indicate models meeting the ≥ 20% post-conventional threshold. All 13 models concentrate entirely in Stages 5–6; Stages 1–3 receive 0% of responses across virtually all models, confirming the distributional inversion reported in Table 1 . Qwen3-235B (Thinking) and Qwen3-32B each produce 100% Stage 6 and 100% Stage 5 responses respectively, illustrating the ceiling and floor of the post-conventional regime.
Figure 22: Emergence of Post-Conventional Reasoning vs. Parameter Count (Segmented Regression). Mean moral reasoning stage (with 95% CI) for each model plotted against log-parameter count. A segmented regression identifies a changepoint at ≈ 32B parameters: the pre-changepoint segment (red, slope =+0.383 ) shows rapid growth, while the post-changepoint segment (green, slope =+0.267 ) shows continued but attenuated growth. The overall linear fit (grey dashed, r=0.36 ) and the orange dot-dash line marking Stage 5+ emergence at 8B together indicate that post-conventional framing appears ubiquitously before significant scale is reached.
Figure 23: Emergence of Post-Conventional Moral Reasoning Across Model Scale. Post-conventional outputs emerge at high rates even for the smallest evaluated models and plateau near ceiling across all scale tiers, indicating that the rhetorical register of mature moral reasoning is acquired early in the scaling regime rather than emerging gradually with capability growth.
Figure 24: Headline Finding. LLMs uniformly produce post-conventional moral language (Stages 5–6, 86% of all responses), the inverse of the Stage 4-dominant human baseline: a pattern that holds across all 13 evaluated models, three prompt types, and six moral dilemmas.
Figure 25: RQ6: Distribution of Moral Vocabulary Richness across LLMs. Horizontal box plots of the count of unique moral and reasoning tokens ( > 3 characters) per response for each model. Claude Sonnet 4.5 and Qwen3-235B (Thinking) display the highest medians and widest spread, reflecting their tendency toward elaborate justifications. Llama models and GPT-4o cluster at the low end despite their large scale, suggesting that vocabulary richness is driven more by instruction-tuning style than raw parameter count.
Figure 26: RQ6: Theoretical Vocabulary Compliance Heatmap. Percentage of responses utilizing canonical stage-specific keywords for each model at each Kohlberg stage it was evaluated on. Nearly all models achieve 100% compliance at Stage 6, while Stage 5 compliance varies (40%–100%) with GPT-4o showing the largest shortfall. Stage 4 compliance applies only to Llama 4 Scout and GPT-4o, both of which occasionally produce Stage 4 evaluations, with perfect keyword usage. High compliance confirms that the evaluator scaffolding elicits stage-appropriate vocabulary even when broader reasoning patterns remain post-conventional.
Figure 27: Capability Correlation Matrix (Pearson r , FDR-corrected). Pairwise correlations among eight response-level capability metrics and mean moral reasoning stage across 13 models. Response Length and Sentence Count are strongly correlated with each other ( r=0.93∗∗∗ ), as are Average Sentence Length and Syntactic Complexity ( r=0.92∗∗∗ ). Neither metric shows a significant correlation with Mean Stage or Post-Conventional percentage, confirming that surface linguistic elaboration does not predict moral sophistication in the Kohlberg sense.
Figure 28: Multi-Capability Predictors of Mean Moral Reasoning Stage (Standardized Coefficients). Forest plot of standardized regression coefficients ( β ) from a multiple regression predicting mean stage from eight capability predictors ( R2=0.840 , Adj- R2=0.521 , p=0.183 ). Sentence Count carries the largest positive coefficient, while Response Length is the strongest negative predictor, reflecting multicollinearity between elaboration metrics. Wide confidence intervals and a non-significant overall F -test indicate that no individual capability dimension reliably predicts moral stage, consistent with the broader finding that conventional measures of language capability do not map onto Kohlberg stage.
Figure 29: Aggregate Stage Transition Matrix. Heatmap of inter-model stage transition probabilities (rows = source model stage; columns = target model stage) when models are ordered by parameter count. Transitions from Stage 4 overwhelmingly target Stage 6 (0.64), suggesting a discontinuous jump rather than a gradual progression. Stage 5 models transition equally to Stage 5 (0.56) and Stage 6 (0.43), while Stage 6 models most often stay at Stage 5 in the next (smaller) model, implying mean-reversion in the scale-ordered sequence. These patterns are consistent with quantized rather than smooth scaling behavior in moral reasoning stage.
Figure 30: Moral Stage Distribution by Model Scale. Heatmap of the proportion of responses at each Kohlberg stage for 13 models ordered by increasing parameter count (bottom to top). Virtually all responses fall in Stage 5 or Stage 6; Stage 4 responses appear only for Llama 4 Scout and GPT-4o ( ≈ 11%). Qwen3-235B (Thinking) reaches 100% Stage 6, while Qwen3-32B achieves 100% Stage 5 — illustrating that the balance between these two post-conventional stages shifts as a function of scale and instruction-tuning orientation, yet neither model produces qualitatively different moral reasoning relative to the human developmental baseline.
Moral reasoning in language models studies how AI systems judge dilemmas, weigh competing values, explain decisions, and stay consistent across contexts.
Large language models (LLMs) increasingly participate in morally sensitive decision-making, yet how they organize ethical frameworks across reasoning steps remains underexplored. We introduce moral reasoning trajectories, sequences of ethical framework invocations across intermediate reasoning steps, and analyze their dynamics across six models and three benchmarks. We find that moral reasoning involves systematic multi-framework deliberation: 55.4--57.7% of consecutive steps involve framework switches, and only 16.4--17.8% of trajectories remain framework-consistent. Unstable trajectories remain 1.29 times more susceptible to persuasive attacks (p=0.015). At the representation level, linear probes localize framework-specific encoding to model-specific layers (layer 63/81 for Llama-3.3-70B; layer 17/81 for Qwen2.5-72B), achieving 16.8--22.2% lower KL divergence than the step-prior baseline. Activation steering applied during generation moves the framework-consistency--accuracy relationship, widening it for Qwen2.5-72B and erasing it for Llama-3.3-70B, and a probe-space layer sweep bounds the attainable drift reduction at 6.7--8.9%. We further propose a Moral Representation Consistency (MRC) metric whose underlying framework attributions are validated by human annotators (mean cosine similarity = 0.859), and we report what an automated coherence rater does and does not establish about it.
We evaluate whether enabling provider-exposed reasoning mode changes moral judgments within the same model checkpoint. Across 100 moral-judgment scenarios and five frontier reasoning-trained LLMs (Claude Sonnet 4.6, GPT 5.5, Gemini 3 Flash, DeepSeek V3.1, and Qwen3.5 397B), aggregate binary-verdict agreement remains high and statistically indistinguishable between instant and thinking modes (Krippendorff's alpha = 0.78 vs. 0.79). However, disagreement is concentrated in 21 model-disputed scenarios, where instant-mode agreement is near chance (alpha = 0.08). On these scenarios, reasoning directionally narrows cross-model disagreement, increasing mean pairwise agreement from 5.4 to 6.7 out of 10. Reasoning also reduces demographic-judgment inconsistency in three of five models and does not increase it for any model. Across all five model families, reasoning changes self-labeled ethical frameworks more often than binary verdicts.
Large language models are increasingly deployed in morally and legally sensitive contexts, yet little is known about how they internally represent distinctions between moral reasoning styles. We ask whether a model's differential response to rule-framed and consequence-framed moral questions reflects distinct internal structure or merely surface lexical sensitivity. We construct a 2×2 factorial dataset crossing reasoning style (rule-based vs. consequence-based framing) with judgment polarity (permissive vs. prohibitive), drawn from three established moral reasoning benchmarks and evaluated on a cleanroom test set of novel phrasings that includes conflict pairs where the two framings predict opposite labels. Using linear probing and activation steering, we identify two orthogonal directions in the residual stream: Dir-A tracks reasoning style and Dir-B tracks judgment polarity, each near-blind to the other factor (cross-correlation ∣r∣<0.04). This double dissociation replicates across three instruction-tuned models from different developers at comparable proportional depth, and it is not knife-edged in layer: the directions are recoverable and causally effective throughout an extended band of the network. Dir-A can causally change judgments, and in conflicting scenarios, it selectively corrects errors in rule-based judgments without changing consequence-based judgments, suggesting that it does not simply make the model more likely to accept everything. Generalization beyond the training templates is partial: the direction transfers across benchmark sources, but its advantage over lexical baselines on novel phrasings is modest. We therefore frame the result as moral reasoning style sensitivity, a geometrically dissociable and causally effective structure, rather than internalized normative reasoning.
Eugenie Shi
Department of Computer Science Stanford University Stanford, CA 94305, USA