3.1. Uncovering implicit values

LLMs are trained on extensive datasets that encapsulate the diversity of human language, culture, and knowledge. However, this diversity also embeds implicit biases, assumptions, and value systems that may not align with universally accepted ethical standards or principles of social good. The process of uncovering these implicit values is essential for identifying misalignments and serves as the foundation for subsequent steps in the proposed framework. This section elaborates on the methodology for identifying the implicit values underlying an LLM model, focusing on topic selection, reference value set selection, and prompt design.

To systematically uncover the implicit values embedded in LLMs, it is necessary to define the domains where ethical considerations are critical. Based on prior research and existing frameworks for AI ethics, this study selected the six core topics representing key pillars of AIfSG (Li et al., Reference Li, Lam and Cui2021, Reference Li, Lam, Han and Cheung2024), including (1) Reasoning and Interpretability, (2) Bias Removal, (3) Transparency and Accountability, (4) Security and Privacy, (5) Moral and Ethical Observations, and (6) Public Understanding.

A core component of uncovering implicit values is selecting an appropriate reference value set that represents a desirable set of moral and ethical standards. Using a fixed, absolute value set allows systematic comparison of LLM outputs against pre-defined benchmarks, offering a consistent and interpretable method for assessing value alignment. The proposed value-driven framework assumes the reference value set to be an input defined by relevant stakeholders, such as domain experts, policymakers, or affected communities. To demonstrate the framework’s versatility across different ethical paradigms, we selected two illustrative reference sets: the Ten Commandments, representing a concise and historically influential moral code often used in philosophical discussions of universal ethics (Irwin, Reference Irwin2011), and the GDPR, reflecting a contemporary, legally grounded framework that operationalizes core principles such as fairness, accountability, and respect for individual rights (Kasirzadeh and Clifford, Reference Kasirzadeh, Clifford, Fourcade, Kuipers, Lazar and Mulligan2021). Together, these two sets capture both moral-philosophical and policy-based dimensions of value alignment, showcasing the framework’s adaptability to diverse normative contexts.

The Ten Commandments, inherently rooted in religious and moral traditions, offer a concise and universally recognizable foundation widely accepted by a considerable portion of the global population (Green, Reference Green2000; Friedman, Reference Friedman2025). This study adopts selected Commandments as a reference value set, with each one mapped to the corresponding AIfSG topics (see Table 1). The correspondences between the Ten Commandments and contemporary ethical constructs (e.g., privacy, fairness, accountability) are illustrative and heuristic rather than literal. They are used to demonstrate how the proposed framework can flexibly translate abstract moral ideas into policy-relevant domains for the purpose of quantifying and comparing value differences before and after realignment. Specifically, Commandment 9 (“You shall not bear false witness against your neighbor”) underpins reasoning, interpretability, and bias removal by emphasizing truthfulness and fairness, key principles for reducing algorithmic bias and ensuring explainable AI outputs (Schlender and Spanakis, Reference Schlender, Spanakis, Baratchi, Cao, Kosters, Lijffijt, van Rijn and Tak2020; Voorhees, Reference Voorhees2021). Commandments 9 and 3 (“You shall not take the name of the Lord your God in vain”) relate to transparency and accountability by promoting honesty in AI disclosures and discouraging misuse of authority (Voorhees, Reference Voorhees2021; Umbrello, Reference Umbrello2025). Security and privacy concerns align with Commandments 8 (“You shall not steal”), 9, and 10 (“You shall not covet”), which condemn theft, deception, and covetous data practices, grounding AI governance in respect for data ownership and informed consent (Thacker, Reference Thacker2020; Sangwa and Mutabazi, Reference Sangwa and Mutabazi2025). Broader ethical observance draws from Commandments 4 (“Remember the Sabbath day, to keep it holy”), 5 (“Honor your father and your mother”), 8, 9, and 10 means designing AI systems that promote human flourishing through respect for rest, family, and community and that refrain from cheating, deceiving, or exploiting others (DeLashmutt, Reference DeLashmutt2025; Sangwa and Mutabazi, Reference Sangwa and Mutabazi2025). Finally, public understanding is supported by the principle in Commandment 8, ensuring an informed public guards against exploitation and affirms the dignity and rights of all, which is consistent with the spirit of the commandment against theft (Heritage History, 2025).

The GDPR establishes a legal-ethical framework for responsible data processing and governance. Building on Article 5 of the GDPR, this study adopts its seven core principles as a reference value set, including (1) lawfulness, fairness, and transparency; (2) purpose limitation; (3) data minimization; (4) accuracy; (5) storage limitation; (6) integrity and confidentiality; and (7) accountability. Each principle is mapped to the corresponding AIfSG topics (see Table 1). Specifically, lawfulness, fairness, and transparency underpin reasoning, interpretability, and public understanding through explainable and trustworthy AI design (Roundtree, Reference Roundtree, Degen and Ntoa2023; Cheong, Reference Cheong2024; Singhal et al., Reference Singhal, Neveditsin, Tanveer and Mago2024). Accuracy and fairness guide bias detection and correction (Alvarez et al., Reference Alvarez, Colmenarejo, Elobaid, Fabbrizzi, Fahimi, Ferrara, Ghodsi, Mougan, Papageorgiou, Reyero, Russo, Scott, State, Zhao and Ruggieri2024). Transparency and accountability ensure traceability and auditability in decision processes (Wieringa, Reference Wieringa, Celis, Ruggieri, Taylor and Zanfir-Fortuna2020). Integrity and confidentiality protect data security and privacy (Lemieux and Werner, Reference Lemieux and Werner2024). Purpose limitation, data minimization, and storage limitation govern moral and ethical data use and retention (Mühlhoff and Ruschemeier, Reference Mühlhoff and Ruschemeier2025).

While the Ten Commandments and the GDPR serve as a starting point, the proposed value-driven framework can be adapted to incorporate other moral and ethical standards, such as the Universal Declaration of Human Rights (UDHR) (United Nations, 1948), or professional and policy guidelines, such as the IEEE Ethically Aligned Design Principles (Shahriari and Shahriari, Reference Shahriari, Shahriari, Krishnan, Fernando and Istrate2017), the OECD AI Principles (OECD, 2024), and the EU AI Act (European Commission, 2024). This adaptability ensures that the proposed value-driven framework can accommodate diverse ethical, cultural, and regulatory contexts.

To extract implicit values and ensure consistent value realignment, a structured prompting strategy was developed to systematically engage the LLM across all six AIfSG topics. The prompting process comprised three stages: (1) Before Value Realignment: generating original principles to ensure a selected AIfSG topic, capturing the LLM’s implicit value assumptions before realignment, (2) Aligning with the Reference Value Set: generating equivalent principles aligned with the selected reference value set, and (3) After Value Realignment: generating equivalent principles that reflect ethical awareness without explicit reference to any reference value set. When aligning the original principles to the reference value set, the prompt design explicitly incorporated considerations of translation and interpretive nuances, particularly for historically derived frameworks such as the Ten Commandments, to mitigate semantic ambiguity and ensure contextually faithful alignment. Detailed prompt templates are provided in Supplementary Table S1 in the Supplementary Information.

3.2. Quantifying value difference

Once the implicit values embedded in LLMs have been uncovered, the next step involves quantifying the extent of misalignment between these values and the reference value set. This quantification provides a measurable basis for evaluating the LLM’s ethical consistency and identifying specific areas requiring further intervention. The proposed value-driven framework employs high-dimensional embeddings, similarity metrics, and visualization techniques to capture and analyze value differences systematically.

To enable quantitative analysis, both the LLM-generated outputs before realignment and the LLM-generated outputs aligned with the reference value set were transformed into high-dimensional vector representations (i.e., embeddings). Value embeddings provide a dense numerical representation of semantic meaning, enabling comparisons across diverse textual inputs. A pre-trained sentence embedding model was used to generate embeddings for each text sample. This embedding model encodes the linguistic and semantic content of a sentence into a fixed multi-dimensional vector space, facilitating quantitative comparisons. For each topic, the LLM-generated responses and their corresponding responses aligning with the reference value set were converted into value embeddings. The same embedding model was used across all topics to ensure consistency and comparability.

The value alignment between LLM outputs and the reference value set was quantified using cosine similarity ranging from −1 to 1, a metric widely used to compare the similarity of vectors in high-dimensional spaces. The value difference was calculated as cosine distance, that is, one minus the cosine similarity (see Equation (1)).

(1) $$ \mathrm{Value}\ \mathrm{Difference}=1- cosine\left(u,v\right) $$

where $ u $ represents the embedding of one LLM-generated response before realignment and $ v $ represents the embedding of the corresponding LLM-generated response aligned with the reference value set. As a result, the value difference ranges from 0 to 2. Alignment refers to cases where the value difference approaches zero, indicating strong correspondence with the reference value set, while misalignment denotes higher value differences that signify divergence from the reference value set. For example, a value difference of 0.1 indicates that the LLM output closely reflects the reference value. In contrast, a value difference of 0.9 highlights a substantial divergence, warranting realignment. For each topic, the average cosine distance across all sample pairs (LLM-generated outputs before realignment and LLM-generated outputs aligned with the reference value set) was calculated, providing an overall value misalignment score for each topic and enabling comparative analysis across different ethical dimensions.

Moreover, to facilitate the interpretation of high-dimensional value embeddings and their differences, dimensionality reduction analysis was performed to project value embeddings into a two-dimensional space. Specifically, the mean value embeddings were calculated for each topic. Multidimensional Scaling (MDS) (Hout et al., Reference Hout, Papesh and Goldinger2013) was employed to reduce the dimensionality of the mean value embeddings, preserving the pairwise distance relationships between them in a lower-dimensional space. This approach allows the visualization to better reflect the value differences in the value embedding space, providing an interpretable representation of the relationship between the LLM outputs and the reference values across topics.

3.3. Analyzing topic-specific value difference

Once value difference has been quantified and visualized, the proposed value-driven framework identifies and prioritizes topics based on the difference in values. Topics were ranked by the average value distance of their respective outputs. Topics with the smallest and largest deviations from the reference value set were identified, highlighting the topics with the least and most significant ethical gaps. Topics with larger average value differences are deemed to have more significant misalignments and are prioritized for further intervention. Conversely, topics with smaller average differences are noted for their closer alignment, requiring minimal adjustment and further intervention.

3.4. Topic-based iterative value realignment

The iterative realignment process serves as the central mechanism for addressing topic-specific value disparities identified during the previous step. This method relies on a systematic, iterative approach to refine the behavior of the LLM, ensuring that its outputs align closely with pre-defined reference values across diverse topics. By iteratively making targeted adjustment, the model can eventually achieve a balance between adhering to moral and ethical standards and preserving its wide-ranging utilities. The iterative framework is particularly effective in handling complex, multi-dimensional value difference, as it allows for incremental, data-driven improvements over successive iterations. The process of iterative realignment systematically reduces value differences through the following stages:

1. (1) Dataset Construction: For each AIfSG topic, pairs of the original principle (LLM-generated output before realignment) and the reference-aligned principle (LLM-generated output aligned with the reference value set) were first obtained from the value-uncovering stage (see Section 3.1). These pairs were combined to form a dataset in which each input (user message) was a prompt instructing the LLM to generate an equivalent principle so that it reflects ethical awareness without explicitly referencing any ethical framework, and each output (assistant message) was the corresponding principle aligned with the reference value set. A topic label was retained for each pair to enable topic-specific weighting during fine-tuning. The dataset was then divided into training, validation, and test subsets, stratified by topic, to ensure balanced representation across all topics. This dataset enables the model to learn how to transform a generic principle into one that aligns with a specific reference value set. By mapping each principle to its corresponding reference-aligned counterpart, the LLM acquires an implicit understanding of how its internal value representations should be adjusted to align with the reference value set.

2. (2) Topic Prioritization Based on Value Difference: To identify and prioritize topics requiring stronger ethical realignment, the mean value difference for each topic was aggregated and normalized to the range [0, 1], yielding a topic-specific loss weight. These weights were used to scale the training loss during fine-tuning so that the total loss for each batch was computed as the weighted mean of individual cross-entropy losses across examples. Topics with larger mean value differences, therefore, contributed more strongly to loss optimization, directing learning toward areas requiring greater value realignment while maintaining balanced representation across all topics.

3. (3) LLM Fine-tuning: Fine-tuning was performed using a topic-weighted supervised fine-tuning approach implemented with parameter-efficient adaptation (LoRA) (Hu et al., Reference Hu, Wallis, Allen-Zhu, Li, Wang, Wang, Chen, Liu, Finn, Choi and Deisenroth2022). Each input–output pair was encoded as a conversational message, where the user message contained the prompt to generate an equivalent principle without specifying the reference value set, and the assistant message contained the reference-aligned principle. Training optimized the token-level cross-entropy loss, the standard objective for causal language modeling, which measures how well the predicted token distribution matches the ground-truth aligned output. The loss for each example was multiplied by its topic-specific weight, ensuring that topics requiring greater value realignment exerted proportionally stronger influence on parameter updates. This formulation enables fine-grained, topic-based optimization of value realignment while seamlessly integrating with LoRA-based fine-tuning for efficient and scalable adaptation.

4. (4) Evaluation and Feedback Loop: After each fine-tuning iteration, the updated LLM generated realigned principles for the validation set. The value differences between these realigned principles and their corresponding reference-aligned principles were recalculated to assess post-training value differences. New topic-wise mean value differences were computed and normalized to update the topic-specific loss weights for the next iteration. This feedback mechanism created an adaptive loop in which topics requiring further value realignment were automatically prioritized in subsequent training cycles. The mean value difference across all topics served as the primary metric for monitoring convergence and evaluating iterative improvements in the value realignment process.

5. (5) Convergence and Termination Criteria: The iterative realignment process continued until one of two stopping conditions was met: (a) the number of iterations reached a predefined threshold or (b) the relative improvement in mean value difference between consecutive iterations fell below a predefined threshold, indicating diminishing returns. The checkpoint yielding the lowest mean value difference was selected as the best model and evaluated on the held-out test set to confirm the generalization of value realignment.

4.1. Experimental setup

In this study, the experiments were structured in two main phases: value uncovering and value realignment, corresponding to the diagnostic and corrective stages of the proposed value-driven framework.

In the value uncovering phase, the implicit values encoded within LLM outputs were identified and compared against reference value sets. Two open-source base models were used for text generation, including Meta AI’s Llama 3.2 [3B parameters] (Touvron et al., Reference Touvron, Lavril, Izacard, Martinet, Lachaux, Lacroix, Rozière, Goyal, Hambro and Azhar2023) and Google DeepMind’s Gemma 2 [2B parameters] (Mesnard et al., Reference Mesnard, Hardin, Dadashi, Bhupatiraju, Pathak, Sifre, Rivière, Kale, Love, Tafti, Hussenot and Sessa2024). Two reference value sets were examined, including the Ten Commandments and the GDPR. For each of the six AIfSG topics, 100 principles were generated to capture the model’s implicit value assumptions and their corresponding reference-aligned counterparts. Both the original and reference-aligned principles were embedded using a sentence transformer model (all-mpnet-base-v2) (Reimers and Gurevych, Reference Reimers, Gurevych, Inui, Jiang, Ng and Wan2019). Cosine distance between embeddings provided a quantitative measure of value difference, which served as the basis for topic prioritization in the subsequent fine-tuning phase.

In the value realignment phase, the base LLMs were iteratively fine-tuned to minimize topic-specific value differences. The value-realignment dataset was divided into 60% training, 20% validation, and 20% testing subsets, stratified by topic for balanced representation. Fine-tuning employed LoRA-based parameter-efficient adaptation with rank = 16, $ \alpha $ = 32, and dropout = 0.05, applied to the attention and projection layers of each model. Training used a batch size of 4, a learning rate of 1 × 10⁻⁵, the AdamW optimizer with weight decay = 0, $ {\beta}_1 $ = 0.9, and $ {\beta}_2 $ = 0.999, and cosine learning rate scheduling for 5 epochs per iteration.

Each value realignment experiment was limited to five fine-tuning iterations, with early stopping triggered when relative improvement in mean value difference on the validation set fell below 1%. After each iteration, topic-specific loss weights were recalculated from the latest validation value differences, adaptively emphasizing topics requiring greater value realignment. The best checkpoint, defined by the lowest mean validation value difference, was then evaluated on the held-out test set to verify the generalization of value realignment.

4.2. Results

The results have demonstrated the efficacy of the proposed value-driven framework for uncovering and realigning the implicit values of LLMs. This section focuses on the results of quantifying value differences, analyzing misaligned topics, and evaluating the impact of iterative realignment.

Figures 2–5 present the mean value differences across six AIfSG topics before and after realignment for two open-source LLMs: Llama 3.2 [3B] and Gemma 2 [2B], under two reference value sets: the Ten Commandments and the GDPR. These figures can serve as diagnostic tools for assessing ethical consistency, pinpointing areas where misalignment persists, and supporting the monitoring, auditing, and governance of AI systems. It should be noted that these analyses provide a structured view of how LLMs express normative principles but not operational performance, that is, how these principles manifest in concrete decisions in applied contexts. For example, Bias Removal does not measure whether the model statistically eliminates bias in its downstream outputs; rather, it evaluates how the model frames fairness and inclusion as ethical priorities within its responses. From a policy perspective, these topic-level insights reveal where an AI system’s ethical framing diverges from socially endorsed standards, helping regulators and auditors target human oversight and prioritize governance interventions in high-risk domains.

Figure 2.

Value difference across topics before and after realignment (LLM: Llama 3.2 [3B]; Reference Value Set: Ten Commandments).

Figure 3.

Value difference across topics before and after realignment (LLM: Gemma 2 [2B]; Reference Value Set: Ten Commandments).

Before value realignment, both LLMs exhibited notable value differences across topics, with the largest discrepancies appearing in Public Understanding (Llama 3.2; Ten Commandments), Reasoning and Interpretability (Gemma 2; Ten Commandments), Moral and Ethical Observations (Llama 3.2; GDPR), and Bias Removal (Gemma 2; GDPR). These variations indicate that the LLMs initially diverged most in domains emphasizing public understanding, reasoning, interpretability, fairness, and broader moral and ethical observations, reflecting topic-dependent sensitivities to ethical reasoning under different reference frameworks. In contrast, smaller value differences were observed in topics such as Security and Privacy and Transparency and Accountability, suggesting a relatively stronger baseline alignment with principles of transparency, accountability, and data protection. Overall, the pre-realignment results reveal that while both LLMs exhibited partial ethical consistency across topics, substantial misalignments persisted in several key domains, underscoring the necessity of targeted, topic-weighted fine-tuning to achieve balanced and comprehensive value realignment.

Supplementary Figures S1 and S2 in the Supplementary Information illustrate the effectiveness of the iterative fine-tuning process in progressively reducing the mean value difference between LLM outputs and the selected reference value sets. These figures visualize the dynamics of realignment, providing a quantitative basis for assessing when iterative interventions have achieved satisfactory convergence and when additional oversight, evaluation, or retraining may be warranted. As shown in Supplementary Figure S1, when the Ten Commandments were used as the reference value set, both Llama 3.2 and Gemma 2 models exhibited a sharp initial decline in value difference after the first iteration, followed by gradual convergence over subsequent iterations. The reduction was particularly pronounced within the first two iterations, suggesting that most realignment occurs early in the process. Likewise, Supplementary Figure S2 shows consistent patterns when the GDPR principles were used as the reference value set. Both models demonstrated substantial decreases in mean value difference, reaching stable convergence within three to five fine-tuning iterations. Collectively, these findings confirm that the proposed topic-weighted iterative fine-tuning approach can effectively minimize value differences between LLM outputs and the chosen reference value sets, demonstrating robustness and generalizability across different LLMs and value systems.

When the Ten Commandments were selected as the reference value set (Figures 2 and 3), both LLMs exhibited consistent convergence toward the reference values after realignment. For Llama 3.2, the largest improvements were observed in Security and Privacy (−70.5%) and Public Understanding (−61.6%), followed by Bias Removal (−57.5%) and Transparency and Accountability (−52.6%). Moderate yet meaningful reductions occurred in Reasoning and Interpretability (−38.0%) and Moral and Ethical Observations (−26.4%). The sharper declines among privacy- and transparency-related topics suggest that realignment most effectively corrected ethical gaps related to data protection and truthful representation. For Gemma 2, reductions were similarly broad, though slightly less pronounced in magnitude, ranging from −47.0% (Security and Privacy) to −28.3% (Moral and Ethical Observations). Topics emphasizing reasoning and fairness, Reasoning and Interpretability (−40.8%), Bias Removal (−45.6%), and Public Understanding (−40.0%) showed steady improvement, indicating generalizability of the realignment process to smaller-parameter models.

When the GDPR was selected as the reference value set (Figures 4 and 5), value differences again decreased markedly across all topics after realignment. Llama 3.2 achieved the largest relative improvement in Bias Removal (−68.5%) and Public Understanding (−61.3%), with moderate reductions in Security and Privacy (−50.1%), Transparency and Accountability (−41.1%), and Reasoning and Interpretability (−36.8%). Gemma 2 displayed comparable performance, with reductions exceeding −45% in Bias Removal and Security and Privacy, and substantial alignment gains in Transparency and Accountability (−65.3%) and Public Understanding (−69.3%).

Figure 4.

Value difference across topics before and after realignment (LLM: Llama 3.2 [3B]; Reference Value Set: GDPR).

Figure 5.

Value difference across topics before and after realignment (LLM: Gemma 2 [2B]; Reference Value Set: GDPR).

Taken together, these results demonstrate that (1) the proposed topic-weighted iterative fine-tuning process effectively minimizes value differences across diverse ethical frameworks and (2) the observed improvements are consistent across model scales and architectures. The relatively larger reductions in Security and Privacy, Transparency and Accountability, and Bias Removal indicate that topics directly tied to fairness, honesty, and responsible data handling benefit most from targeted value realignment. In contrast, smaller yet consistent reductions in Moral and Ethical Observations suggest that abstract moral reasoning may require additional iterations or richer contextual grounding for full value realignment. Overall, the convergence patterns observed across all figures confirm that the proposed value-driven framework offers a scalable and generalizable pathway for embedding moral and ethical values into LLMs.

Moreover, Figures 6–9 visualize the two-dimensional value embedding spaces before and after realignment for both Llama 3.2 and Gemma 2, using the Ten Commandments (Figures 6 and 7) and GDPR (Figures 8 and 9) as the reference value sets. MDS (Hout et al., Reference Hout, Papesh and Goldinger2013) was used to project the mean value embeddings of each AIfSG topic into a two-dimensional space. MDS was selected because it preserves the pairwise distances among embeddings more faithfully than linear reduction methods such as principal component analysis (PCA), enabling clearer visualization of how the relative proximities between model and reference values evolve through the realignment process. Before realignment, large separations between LLM outputs (circles) and reference values (squares) indicate substantial misalignments across topics. After realignment, the realigned outputs (diamonds) converge markedly toward their corresponding references, confirming reduced value differences across all topics. These visualizations provide an intuitive geometric validation of the quantitative results: the topic-specific embeddings move significantly closer to their reference values, revealing a clear reduction in pairwise distances and enhanced alignment across all six AIfSG topics.

Figure 6.

Value difference in the value-embedding space before and after realignment (LLM: Llama 3.2 [3B]; Reference Value Set: Ten Commandments).

Figure 7.

Value difference in the value-embedding space before and after realignment (LLM: Gemma 2 [2B]; Reference Value Set: Ten Commandments).

Figure 8.

Value difference in the value-embedding space before and after realignment (LLM: Llama 3.2 [3B]; Reference Value Set: GDPR).

Figure 9.

Value difference in the value-embedding space before and after realignment (LLM: Gemma 2 [2B]; Reference Value Set: GDPR).

Further, to complement the quantitative value-difference analysis, we performed a qualitative validation to examine whether the realigned LLM outputs more closely reflect the intended reference value sets. For each model and each reference framework, we identified the LLM-generated outputs that exhibited the largest reductions in value difference before and after realignment (see Supplementary Tables S2–S5 in the Supplementary Information). Manual inspection of these samples indicated that large changes in value difference were often associated with substantive shifts in ethical emphasis. For example, under the Ten Commandments, in Public Understanding (Supplementary Table S2), broad procedural statements such as “AI should be accountable and responsive to societal needs” were rephrased as explicit ethical directives like “You shall not manipulate public discourse for personal gain or to obscure the truth.” Similarly, in Reasoning and Interpretability (Supplementary Table S3), the outputs moved from technical, process-oriented descriptions toward statements underscoring honesty and integrity. Under the GDPR, the outputs showed closer value alignment with principles of fairness, accountability, and transparency, reframing general design recommendations into formulations that emphasize lawful and trustworthy AI behavior (Supplementary Tables S4 and S5). Taken together, these qualitative observations suggest that the largest quantitative improvements in cosine-based value difference are typically accompanied by meaningful shifts in ethical orientation. Nevertheless, expert evaluations will be necessary to systematically verify whether these observed changes align with human moral judgments across diverse contexts.