Scaling Laws for Moral Machine Judgment in Large Language Models Kazuhiro Takemoto Abstract Autonomous systems increasingly require moral judgment capabilities, yet whether these capabilities scale predictably with model size remains unexplored. We systematically evaluate 75 large language model configurations (0.27B–1000B parameters) using the Moral Machine framework, measuring alignment with human preferences in life-death dilemmas. We observe a consistent power-law relationship with distance from human preferences ( D D ) decreasing as D ∝ S − 0.10 ± 0.01 D\propto S^{-0.10\pm 0.01} ( R 2 = 0.50 R^{2}=0.50 , p < 0.001 p<0.001 ) where S S is model size. Mixed-effects models confirm this relationship persists after controlling for model family and reasoning capabilities. Extended reasoning models show significantly better alignment, with this effect being more pronounced in smaller models (size × \times reasoning interaction: p = 0.024 p=0.024 ). The relationship holds across diverse architectures, while variance decreases at larger scales, indicating systematic emergence of more reliable moral judgment with computational scale. These findings extend scaling law research to value-based judgments and provide empirical foundations for artificial intelligence governance. 1 Introduction Autonomous systems increasingly face ethical dilemmas requiring value-based judgments [ 2 ] . Self-driving vehicles must decide how to allocate risk in unavoidable collision scenarios [ 6 ] , while medical artificial intelligence (AI) systems navigate complex tradeoffs between patient autonomy and beneficence [ 24 ] . The rapid advancement of large language models (LLMs) has accelerated their integration into such safety-critical decision-making roles [ 11 , 32 , 20 , 27 , 33 ] , making their moral judgment capabilities a central concern for AI safety and governance. Recent work has established scaling laws across diverse cognitive capabilities. Model performance improves with increased parameters in language modeling and understanding [ 16 , 13 ] . Many sophisticated capabilities emerge spontaneously through scale alone, without explicit training [ 30 ] . These observations raise the question of whether moral judgment capabilities similarly scale with model size, following patterns analogous to other cognitive domains. The Moral Machine experiment [ 3 ] provides a framework for investigating this question. This study collected 40 million decisions from participants across 233 countries on autonomous vehicle dilemmas, measuring how moral factors including age, social status, species, and legality influence choices. Rather than aggregating preferences into single correct answers, the framework quantifies the relative strength of different moral considerations, yielding preference distributions that capture value pluralism. This design enables measurement of alignment with human preference patterns through post-hoc statistical analysis, assessing moral intuitions less likely to have been explicitly targeted during model training or alignment procedures. Our previous work [ 33 ] used this framework to provide initial evidence for scaling in moral judgment, observing improved alignment with human preferences in larger models. However, that analysis was limited to a small number of available models. The rapid proliferation of LLMs spanning orders of magnitude in size now enables systematic investigation of whether moral alignment follows predictable scaling laws. Several recent benchmarks evaluate ethical reasoning in LLMs using crowdsourced labels to establish correct answers. ETHICS [ 12 ] assesses knowledge across multiple ethical frameworks including utilitarianism and deontology. Moral Stories [ 8 ] evaluates moral reasoning through narrative understanding. Social Chemistry [ 9 ] measures understanding of social norms and rules of thumb. These benchmarks provide valuable tools for assessing moral knowledge and reasoning about established norms. However, for real-world dilemmas faced by autonomous systems—such as unavoidable collision scenarios or medical tradeoffs—where no objective ground truth exists and multiple reasonable perspectives coexist, frameworks that preserve preference distributions offer complementary advantages. The Moral Machine’s focus on preference alignment rather than classification accuracy makes it well-suited for studying how moral judgment capabilities scale with model size in contexts where human values are pluralistic. We systematically investigate scaling relationships in moral judgment across 75 LLM configurations spanning 0.27B to 1000B parameters, representing a substantial expansion from prior work. Our analysis reveals a consistent power-law relationship where larger models systematically demonstrate improved alignment with human moral preferences. Mixed-effects models confirm this relationship persists after controlling for model family and reasoning capabilities. These findings demonstrate that moral judgment follows predictable scaling patterns, extending scaling law research to value-based judgments and providing quantitative foundations for AI governance. 2 Methods 2.1 Large Language Models We evaluated 75 model configurations spanning 0.27B to 1000B parameters using the evaluation framework established in our previous work [ 27 , 33 ] . Complete methodological details including prompt formats, response coding procedures, and validation against the original Moral Machine human data are provided in those papers. The 75 configurations include different model sizes within families (e.g., Llama 3 8B, 70B), versions (e.g., DeepSeek V3, V3.1, V3.2), and reasoning modes (e.g., standard vs. extended reasoning models including reasoning-focused and thinking-mode models). For each model family, we prioritized instruction-tuned or chat-optimized variants over base models, as base models often produce unreliable responses to conversational prompts. This ensures that observed scaling relationships reflect genuine moral judgment capabilities rather than artifacts of poor instruction-following. The model set encompasses diverse architectures and training approaches, from both open-weight sources (Llama, Gemma, Qwen, Mistral, DeepSeek) and proprietary APIs (GPT and Claude). For open-weight models, we used officially reported parameter counts. For proprietary models, we used parameter estimates from published literature [ 1 ] , as official counts are not publicly disclosed. Complete model specifications, release dates, and evaluation results are provided in Table S1 . All models received identical prompts with default sampling parameters. Responses were coded following the procedures detailed in [ 27 ] . 2.2 Moral Machine Framework We employed the Moral Machine framework [ 3 ] , a validated instrument for measuring moral preferences in autonomous vehicle dilemmas. Following the methodology in [ 27 ] , we generated 10,000 distinct scenarios by systematically varying nine moral factors: age (young vs. elderly), gender (male vs. female), social status (high vs. low), physical fitness (fit vs. large), species (human vs. pet), legality (legal vs. illegal crossing), number of characters (more vs. fewer), intervention type (swerve vs. stay), and whether characters are passengers or pedestrians. Standard models and thinking-mode models with thinking disabled evaluated all 10,000 scenarios. Reasoning-focused models and thinking-mode models with thinking enabled evaluated the first 5,000 scenarios due to longer inference times. Complete prompt templates, response coding procedures, and scenario generation details are provided in [ 27 ] . 2.3 Evaluation Metrics Following [ 27 ] , we quantified alignment between model judgments and human preferences using the Average Marginal Component Effect (AMCE) [ 10 ] . AMCE measures the causal influence of each moral factor on decision-making by estimating the average change in choice probability when a factor level changes (e.g., from young to elderly) while holding other factors constant. For each model, we computed AMCE values for all nine moral factors, yielding a 9-dimensional preference vector (Table S2 ). We quantified overall alignment as the Euclidean distance between each model’s AMCE vector and the aggregate human AMCE vector derived from the original Moral Machine experiment [ 3 ] . Lower distances indicate greater correspondence to human moral preferences. Complete details on AMCE computation and human preference vectors are provided in [ 27 ] . 2.4 Statistical Analysis We analyzed the relationship between model size (parameter count) and alignment metrics using Spearman rank correlation to account for potential non-linear relationships and outliers. To test for power-law scaling, we fit log-linear models ( \log_{10}D\sim\log_{10}S ) and compared goodness of fit ( R^{2} ) against alternative functional forms including linear ( D\sim S ), logarithmic ( D\sim\log_{10}S ), and exponential models ( \log_{10}D\sim S ). The power-law exponent \alpha and its standard error were estimated from the slope of the log-linear regression. To assess whether the observed scaling relationship reflects effects of model size rather than correlated confounding factors, we employed linear mixed-effects models with model family (DeepSeek, Llama, Gemma, Qwen, Other) as random effects, including family-specific intercepts and slopes for model size. We included two additional predictors: (1) release date, converted to a continuous variable (years) to capture temporal trends and control for potential scale-independent improvements in model capabilities, such as advances in training methodology, architecture design, or alignment procedures; (2) reasoning capability, coded as a binary variable indicating whether models employ extended reasoning processes (reasoning-focused models and thinking-mode variants = 1, standard models = 0). We compared nested models using likelihood ratio tests and information criteria (AIC, BIC): (A) model size only, (B) model size and release date, (C) model size, release date, and reasoning capability, and (D) model size, release date, reasoning capability, and size \times reasoning interaction. All statistical analyses were performed using R software (version 4.4.2) with the lme4 [ 4 ] and lmerTest [ 18 ] packages. Mixed-effects models were fitted using restricted maximum likelihood (REML) for parameter estimation and maximum likelihood (ML) for model comparison. 3 Results 3.1 Overall Scaling Relationship We first examined the relationship between model size and moral judgment alignment across all 75 model configurations (Figure 1 ). We observed a strong negative correlation (Spearman \rho=-0.73 , p=1.8\times 10^{-13} ), indicating that larger models systematically align better with human moral preferences. The relationship follows a power-law pattern D\propto S^{-\alpha} , where D represents Euclidean distance from human preferences, S denotes model size in parameters, and \alpha=0.10\pm 0.01 . This power-law model achieved superior fit ( R^{2}=0.50 ) compared to alternative functional forms: linear ( R^{2}=0.16 ), logarithmic ( R^{2}=0.47 ), and exponential models ( R^{2}=0.19 ). The power-law formulation outperforms the next-best model (logarithmic) by \Delta R^{2}=0.03 , and substantially exceeds linear and exponential forms. The consistency of this relationship across 75 diverse models spanning three orders of magnitude in size (0.27B–1000B parameters) suggests a fundamental scaling law for moral judgment capabilities. Figure 1: Scaling relationship between model size ( S ) and moral judgment alignment with human preferences (distance from human, D ). Each point represents one LLM, colored by model family. The dashed line shows the fitted power-law relationship ( D\propto S^{-0.10\pm 0.01} ) Beyond improved mean alignment, visual inspection of Figure 1 reveals that variance in alignment decreases with model size. Smaller models exhibit substantial variability in performance, while larger models cluster more tightly around the power-law trend, indicating more reliable and consistent moral judgment capabilities at scale. 3.2 Robustness to Confounding Factors To assess whether the observed scaling relationship reflects genuine effects of model size rather than confounding factors, we fitted linear mixed-effects models with model family (DeepSeek, Llama, Gemma, Qwen, Other) as random effects. This approach accounts for family-specific differences in architecture and training approaches while estimating the effect of size. Model size correlates with several factors that could contribute to improved alignment, including computational budget, architectural sophistication, and advances in training methodology. While the first two can be partially controlled through family-specific random effects (as models within families share similar architectures and resource access), temporal improvements in model development are not directly observable. We used release date, converted to a continuous variable, as a proxy for potential scale-independent improvements over time, which could include advances in training practices, data curation, or alignment techniques. We compared nested models using likelihood ratio tests: (A) model size only, (B) model size and release date, (C) model size, release date, and reasoning capability, and (D) model size, release date, reasoning capability, and size \times reasoning interaction (Table 1 ). Table 1: Linear mixed-effects model comparison for moral alignment. All models include random intercepts and slopes for model family. Model comparison uses maximum likelihood estimation. Model Fixed Effects AIC BIC \chi^{2} p -value A Size -126.6 -112.7 – – B Size + Release date -124.7 -108.5 0.10 0.76 C Size + Release date + Reasoning -130.0 -111.4 7.24 0.007 D Size + Release date + Reasoning + Size \times Reasoning -133.0 -112.1 5.05 0.025 \chi^{2} and p -values from likelihood ratio tests against previous model Model comparison revealed that adding release date did not improve fit (Model B vs. A: \chi^{2}=0.10 , p=0.76 ), suggesting that temporal factors in model development, which could include advances in training methodology, architectural innovations, or alignment procedures, do not substantially contribute to moral alignment beyond the effects captured by model size and family. However, adding reasoning capability significantly improved the model (Model C vs. B: \chi^{2}=7.24 , p=0.007 ). The full model including the size \times reasoning interaction provided further significant improvement (Model D vs. C: \chi^{2}=5.05 , p=0.025 ). The final model (Model D; Table S3 ) confirmed the power-law scaling relationship while controlling for family differences and temporal trends. Model size showed a significant negative effect ( \beta=-0.12 , SE =0.016 , p<0.001 ), corresponding to a power-law exponent \alpha=0.12 consistent with the overall analysis. Extended reasoning models showed significantly better alignment ( \beta=-0.16 , SE =0.044 , p=0.001 ). Family-specific patterns (Figure S1 ) show predominantly negative relationships between size and alignment, though statistical power for detecting these relationships within individual families varies due to differences in sample size (8–19 models per family) and parameter range coverage. Random effects captured modest family-level variation (SD {}_{\text{intercept}}=0.019 , SD {}_{\text{slope}}=0.017 ; Table S3 ), with the mixed-effects approach providing robust estimates by pooling information across families while accounting for family-specific heterogeneity. Critically, the size \times reasoning interaction was positive and significant ( \beta=0.057 , SE =0.025 , p=0.024 ), indicating that the advantage of extended reasoning capabilities is more pronounced in smaller models (Figure S2 ). This suggests that architectural innovations in reasoning provide complementary pathways to improved alignment, with particularly strong benefits when computational scale is limited, while very large models may already capture much of this reasoning capability through scale alone. Comparison of coefficient estimates across all nested models is provided in Table S4 . Analysis of residuals after controlling for size showed no correlation with release date (Spearman \rho=0.047 , p=0.69 ; Figure S3 ; Table S4 ) further supporting that scale-independent temporal improvements in model development do not substantially contribute to alignment beyond scale and reasoning capabilities. This null result should be interpreted cautiously given the limited temporal range of our dataset (95% of models released in 2024–2025), which may constrain our ability to detect gradual improvements in training data quality over time. 4 Discussion Our findings establish that computational scale systematically predicts moral judgment capabilities in LLMs. Across 75 model configurations spanning three orders of magnitude (0.27B–1000B parameters), we observed a consistent power-law relationship ( D\propto S^{-0.10\pm 0.01} , R^{2}=0.50 , p<0.001 ), where distance from human moral preferences decreases predictably with model size. This relationship proved robust to potential confounding factors: mixed-effects models controlling for model family confirmed that the scaling pattern persists across diverse architectures and training approaches. Family-specific analyses revealed predominantly negative relationships between size and alignment across all major model families (DeepSeek, Llama, Gemma, Qwen), demonstrating that the power-law pattern is not driven by any single architectural approach or training methodology. While computational scale represents the primary determinant of moral alignment, additional factors contribute independently. Extended reasoning models demonstrated significantly better alignment ( \beta=-0.16 , p=0.001 ). Critically, this effect varies with model size: the size \times reasoning interaction ( \beta=0.057 , p=0.024 ) indicates that extended reasoning provides greater relative benefits in smaller models, while very large models already capture much of this capability through scale alone. This indicates that architectural or training innovations enabling extended reasoning, such as chain-of-thought processing [ 31 , 29 ] and iterative refinement [ 19 ] , enhance moral judgment, with this enhancement being particularly pronounced when computational scale is limited. In contrast, temporal factors in model development, proxied by release date, did not contribute substantially to alignment after controlling for size and reasoning capability ( p=0.76 ). This null result suggests that while model size and reasoning architecture represent primary pathways to improved moral alignment, scale-independent improvements over time—which could include advances in training practices, data curation, or alignment techniques—do not provide comparable benefits, at least within the temporal range of our dataset (2023–2025). The modest power-law exponent ( \alpha=0.10 ) reveals an important characteristic of moral judgment capabilities: they scale slowly with model size. A tenfold increase in parameters reduces distance from human preferences by only approximately 21%, indicating that moral judgment represents a particularly challenging emergent capability requiring substantial computational scale to approach human-level performance. This gradual scaling contrasts with steeper improvements observed in some other domains and, while subject to the interpretation of our distance-based metric, suggests fundamental computational constraints on acquiring human-like moral intuitions. For practical deployment, this slow scaling implies that achieving substantial improvements in moral alignment requires order-of-magnitude increases in model size or complementary approaches such as extended reasoning architectures. Extended reasoning provides particularly strong benefits in resource-constrained settings where deploying very large models is impractical, offering an alternative pathway to improved alignment without proportional increases in parameter count. Beyond improved mean alignment, larger models also exhibit reduced variance in performance, clustering more tightly around the power-law trend and indicating systematic emergence of more reliable and consistent moral judgment capabilities with scale. This dual benefit of improved alignment and increased reliability suggests that computational scale enhances both the quality and predictability of moral reasoning in safety-critical applications. These findings extend scaling law research from tasks with objective ground truth to value-based judgments where correctness is defined by alignment with human preferences rather than factual accuracy. Previous scaling studies [ 16 , 13 ] have focused predominantly on capabilities with verifiable solutions and factual question answering, where performance can be evaluated against definitive correct answers. Our demonstration that moral judgment follows analogous power-law patterns despite lacking objective ground truth suggests that scaling laws may represent a more general principle of capability emergence [ 30 ] in large language models. This extension has important theoretical implications: it indicates that models can acquire implicit knowledge of human value systems through exposure to training data reflecting diverse ethical perspectives, and that this acquisition process follows predictable quantitative relationships with computational scale. The consistency of scaling patterns across model families confirms that computational scale represents the primary determinant of moral alignment. While architectural innovations such as extended reasoning provide complementary improvements, the fundamental relationship between model size and moral judgment quality follows a universal power-law that transcends specific implementation choices. These findings provide quantitative foundations for risk-based deployment of LLMs in contexts requiring moral judgment. The power-law relationship enables prediction of alignment quality from model size, allowing system designers to make informed tradeoffs between computational costs and ethical reliability. For safety-critical applications where ethical failures carry substantial consequences, our results indicate that larger models or those enhanced with extended reasoning capabilities offer improved reliability. Extended reasoning represents a complementary pathway to enhanced alignment that does not require proportional increases in parameter count, with particularly strong benefits when computational scale is limited. Beyond mean alignment, the observed reduction in variance at larger scales implies more predictable performance, reducing the risk of outlier judgments that could undermine trust or violate ethical standards in high-stakes contexts [ 26 , 25 , 14 ] . Several limitations warrant consideration. Our evaluation uses Euclidean distance between model and human AMCE vectors as the primary alignment metric. While this metric provides mathematical rigor and enables quantitative comparison across models, the relationship between distance reductions and practical improvements in real-world moral judgment quality remains to be empirically validated. Future work should complement distance-based metrics with human evaluations to establish the practical significance of observed improvements. Model size correlates with multiple factors including computational budget, architectural sophistication, and advances in training methodology. While our mixed-effects analyses control for model family and temporal trends in model development, fully isolating the causal effect of scale from all correlated factors requires controlled experiments that are impractical at the scale of contemporary LLMs. Training data for most models is not publicly disclosed, making it impossible to assess whether Moral Machine scenarios appeared in pretraining corpora, though the consistency of scaling patterns across models from different organizations suggests systematic relationships rather than memorization artifacts. Our evaluation focuses on life-death dilemmas from the Moral Machine framework, representing one aspect of moral reasoning. We selected this framework because its preference-based evaluation approach, rather than correct-answer classification, reduces the risk of data contamination from training corpora while enabling measurement of spontaneous moral intuitions. Whether similar scaling relationships exist for other ethical dimensions such as fairness, autonomy, and virtue ethics remains an important future direction. The Moral Machine dataset predominantly reflects preferences from participants in Western countries, and cross-cultural validation [ 28 , 15 ] is essential to assess the generalizability of our findings across diverse value systems. Finally, different prompting strategies might yield quantitatively different relationships [ 23 , 17 ] , though we expect directional trends to remain consistent. Despite the limitations discussed above, our results indicate that computational scale represents an important factor in moral alignment. The observed power-law relationship provides empirical evidence that model size systematically predicts alignment with human moral preferences, suggesting that human-aligned moral capabilities emerge with computational scale. These findings offer initial quantitative insights for AI governance [ 7 , 5 ] and inform risk assessment in deploying AI systems [ 22 , 21 ] requiring value-based judgments. Supplementary Figures Figure S1: Family-specific scaling relationships. Log-log plot showing the relationship between model size and distance from human preferences for each model family. Points represent individual models, lines show linear regression fits with 95% confidence intervals. All families exhibit negative scaling relationships, demonstrating that the power-law pattern is not driven by any single architectural approach. Model families: DeepSeek (red circles), Gemma (blue triangles), Llama (green squares), Qwen (purple diamonds), Other (brown crosses). Figure S2: Effect of extended reasoning capabilities on moral alignment. Comparison of scaling relationships between standard models (gray circles) and extended reasoning models (red triangles). Extended reasoning models show systematically better alignment (lower distance) at comparable sizes, with the advantage being more pronounced in smaller models. The significant main effect of reasoning ( \beta=-0.16 , p=0.001 ) and the size \times reasoning interaction ( \beta=0.057 , p=0.024 ) indicate that architectural innovations in reasoning provide particularly strong benefits when computational scale is limited. Lines show linear regression fits with 95% confidence intervals. Figure S3: Temporal trends in moral alignment. Residual distance from human preferences (after controlling for model size) plotted against release date. The absence of a significant trend (Spearman \rho=0.047 , p=0.69 ) suggests that scale-independent temporal improvements in model development do not contribute substantially to alignment beyond the effects of scale and reasoning capabilities. Points represent individual models, line shows linear regression fit with 95% confidence interval. Supplementary Tables Table S1: Model metadata and alignment results. This table provides key information for all 75 model configurations evaluated in this study. OpenWeight : whether the model weights are publicly available (yes) or proprietary (no). Reasoning : whether the model employs extended reasoning capabilities (yes = reasoning-focused models or thinking-mode variants; no = standard models). API ID : official API identifier for proprietary models accessed via vendor APIs; blank for open-weight models evaluated locally. Size : model size in billions of parameters; for proprietary models, values represent estimates from published literature [ 1 ] as official counts are not publicly disclosed. Release Date : official release date in YYYY/MM/DD format. Distance : Euclidean distance between the model’s AMCE vector and the human AMCE vector (Table S2 ), where lower values indicate closer alignment with human moral preferences. Valid Response Rate : proportion of scenarios where the model provided a valid binary choice response that could be coded for analysis. This is publicly available from our GitHub repository ( https://github.com/kztakemoto/mmllm/ ). Table S2: Average Marginal Component Effect (AMCE) values for all models and human preferences. Each row represents one model configuration’s (or human) 9-dimensional preference vector, quantifying the causal influence of each moral factor on decision-making. AMCE values indicate the average change in choice probability when a factor level changes (e.g., from young to elderly) while holding other factors constant. Positive values indicate preference for the first level of each factor: young (vs. elderly), fit (vs. large), male (vs. female), swerve (vs. stay), legal crossing (vs. illegal), more characters (vs. fewer), passengers (vs. pedestrians), high status (vs. low), and human (vs. pet). Negative values indicate preference for the second level. Model size is reported in billions of parameters. The “Human” row represents aggregate preferences from the original Moral Machine experiment [ 3 ] with 40 million decisions across 233 countries. This is publicly available from our GitHub repository ( https://github.com/kztakemoto/mmllm/ ). Table S3: Mixed-effects model (Model D) parameter estimates. The model includes model size (log 10 parameters), release date, reasoning capability, and the size \times reasoning interaction as fixed effects, with random intercepts and slopes for model family. The model was fitted using restricted maximum likelihood (REML). Parameter Estimate SE df t -value p -value Fixed Effects Intercept 0.033 0.107 56.6 0.312 0.756 Model Size (log 10 ) -0.120 0.016 5.9 -7.397 < 0.001 Release Date 0.033 0.023 58.4 1.458 0.150 Reasoning -0.156 0.044 40.3 -3.544 0.001 Size \times Reasoning 0.057 0.025 56.3 2.313 0.024 Random Effects (Standard Deviations) Family: Intercept = 0.019 Family: Slope (Model Size) = 0.017 Correlation (Intercept, Slope) = -1.00 Residual = 0.091 Observations: 75; Groups (Family): 5 AIC = -119.7 (REML) Table S4: Comparison of fixed effect estimates across nested mixed-effects models. All models include random intercepts and slopes for model family. Standard errors in parentheses. Models were fitted using REML for parameter estimation. Parameter Model A Model B Model C Model D Fixed Effects Intercept 0.142* 0.093 0.006 0.033 (0.027) (0.110) (0.106) (0.107) Model Size (log 10 ) -0.096 *** -0.096 *** -0.101 * -0.120 *** (0.014) (0.014) (0.012) (0.016) Release Date – 0.010 0.034 0.033 (0.023) (0.023) (0.023) Reasoning – – -0.074 ** -0.156 ** (0.027) (0.044) Size \times Reasoning – – – 0.057* (0.025) Random Effects (SD) Family: Intercept 0.042 0.040 0.005 0.019 Family: Slope (Size) 0.014 0.009 0.005 0.017 Residual 0.095 0.095 0.094 0.091 Model Fit AIC (REML) -125.1 -119.5 -120.2 -119.7 * p<0.05 ; ** p<0.01 ; *** p<0.001