Papers by Year
Venue Type Distribution
Analytical Dimensions
Each card previews one analytical axis of the corpus. Click to explore the full tab.
Top 10 Categories
Data Availability
| Year | Paper | Venue | Ethics | Moral Psych | Religion | Model | Flags |
|---|
Where research is published shapes how it is received: peer-reviewed conference and journal papers undergo expert scrutiny that preprints do not. This section maps the vault's 2,043 papers across their publication venues — from flagship NLP conferences (ACL, EMNLP, NeurIPS) to domain-specific workshops and open-access preprint servers. The venue distribution reveals both the field's institutional centre of gravity and its long tail: a handful of top-tier AI venues account for a disproportionate share of values-alignment work, while hundreds of smaller venues contribute one or two papers each.
Venue Type Distribution
Top 30 Venues
Conference & Venue Papers
0Individual venue names fragment what are really conference families — ACL, NAACL, EACL, and AACL-IJCNLP all belong to the ACL ecosystem; NeurIPS, ICML, and ICLR form the ML core. This sub-tab groups venues into families and ecosystems, revealing which research communities drive the ethics conversation. The treemap shows relative mass, the ecosystem chart maps disciplinary homes, and the year heatmap tracks how each family's engagement has evolved.
Conference Family Treemap
Ecosystem Distribution
Conference Family x Year Heatmap
Conference Family Papers
0Journal publications complement the conference-driven pace of AI research with longer review cycles and broader disciplinary reach. Many ethics and social-science perspectives enter the AI values conversation through journals rather than conference proceedings. The chart below ranks the 20 most-represented journals in the vault, spanning AI-specific outlets, interdisciplinary ethics journals, and domain venues in law, philosophy, and public policy.
Top 20 Journals
Journal Papers
0Publication patterns shift over time as new venues emerge and the balance between peer-reviewed and preprint outlets evolves. The timeline below tracks how venue types — conferences, journals, workshops, and preprints — have contributed to the vault year by year. The peer-review ratio chart reveals whether the field's rapid growth has maintained its proportion of expert-scrutinised work or increasingly relied on unreviewed preprints.
Venue Type Timeline
Peer-Review Ratio Over Time
Publication Papers
0Every paper in the vault carries a multi-label classification spanning ethics focus, research methodology, data nature, and topical domain — assigned through a seven-stage keyword-and-heuristic pipeline. This overview visualises how those categories distribute over time and interact with each other. The year heatmap reveals when particular sub-fields gained traction, the co-occurrence matrix shows which topics are studied together (and which remain siloed), and the model-family chart highlights which AI systems attract the most taxonomic scrutiny.
Category x Year Heatmap
Category Co-occurrence
Top Model Families
The vault contains — benchmark candidates identified through deep-read survey. This sub-tab analyzes benchmark coverage across ethical traditions, cultures, and languages — revealing where the evaluation ecosystem is robust and where critical gaps persist.
Benchmark x Tradition
Benchmark Language Coverage
Benchmark Timeline
Benchmark Venue Types
Benchmark Candidates
0Cross-dimensional analysis reveals the intersections that isolated tabs cannot show. The most important questions in values alignment are inherently cross-dimensional: Which models have been tested on non-WEIRD populations? Do papers using MFT also engage with religious traditions? Which conferences publish cross-cultural work? These heatmaps expose exactly where the research space is dense and where it is empty.
Model x Ethical Framework
Model x Culture/Region
Venue x Dimension
Framework x Religion
This section maps the cultural landscape of LLM ethics research across the vault's 2,043 papers. Cultural values alignment is critical because ethical norms vary profoundly across societies — what is considered fair, respectful, or virtuous in one cultural context may be perceived very differently in another. Most current LLM benchmarks reflect WEIRD (Western, Educated, Industrialized, Rich, Democratic) value systems, creating systematic blind spots for the majority of the world's population.
Data is drawn from two sources: paper-level YAML classifications (national/ethnic/cultural tags on all 2,043 papers) and the 989-candidate benchmark survey (language, tradition, and framework extraction via NLP).
Region Distribution
Region Timeline
Country and nationality tags reveal which societies' values the AI ethics literature actually examines. The bar chart below ranks the 30 most-studied cultures and nationalities. While the Regions sub-tab aggregates by continent, this view preserves the granularity needed to spot imbalances — for example, whether "East Asian" coverage is driven primarily by Chinese or Japanese studies, or whether African representation is concentrated in a single country.
Culture/Nationality Papers
0Top 30 Cultures/Nationalities
Culture Focus Shift (2016–2026)
Culture Co-Occurrence Matrix
Ethical Framework × Culture
Benchmark Gap by Culture
Venue Type × Culture
Language is both a proxy for cultural reach and a practical barrier to evaluation. A benchmark available only in English cannot test whether an LLM's moral reasoning holds in Mandarin, Arabic, or Yoruba. This sub-tab charts the languages explicitly studied or used in the vault's papers — revealing how far the field has moved beyond English-only evaluation and where monolingual blind spots persist.
Language Papers
0Top 20 Languages
Language Family Treemap
Language Coverage Timeline (2016–2026)
Language Benchmark Gap
Language–Culture Alignment by Region
Venue Type × Language
The WEIRD framework (Western, Educated, Industrialized, Rich, Democratic) identifies a sampling bias that pervades psychology and, by extension, AI alignment research. Papers focused on populations outside WEIRD societies often uncover value structures that mainstream benchmarks miss entirely. This chart classifies every culturally-tagged paper in the vault as WEIRD or non-WEIRD, quantifying the extent of that imbalance.
WEIRD vs Non-WEIRD
WEIRD Analysis
WEIRD = Western, Educated, Industrialized, Rich, Democratic societies. Most AI ethics research focuses on WEIRD populations.
The treemap provides a hierarchical view of the vault's cultural coverage, nesting individual countries and nationalities within their parent regions. Rectangle size encodes paper count, making it immediately visible where research mass concentrates and which regions contain only a thin scattering of studies. Click or hover to drill into any region's constituent cultures.
Region → Culture Treemap
Gap analysis reveals where the LLM ethics evaluation ecosystem falls short. Despite growing interest in cross-cultural AI assessment, significant blind spots persist. The following frameworks and best practices can guide future benchmark development.
Best Practices for Cultural Values Alignment
Beyond WEIRD-Centric Evaluation
Current benchmarks over-represent Western moral intuitions. Evaluation suites must actively include non-Western ethical frameworks as first-class perspectives, not afterthoughts.
Culturally-Situated Evaluation
Translation alone does not equal cultural adaptation. Benchmarks must be designed with local cultural experts who understand the nuances of moral reasoning in their communities.
Multilingual ≠ Multicultural
A benchmark translated into 20 languages but designed with Western moral assumptions still measures Western values in other languages. True multicultural evaluation requires culturally-native scenario design.
Engage Local Communities
Meaningful cultural values alignment requires participatory design with affected communities, not top-down application of pre-existing frameworks.
What's Missing: Underrepresented Traditions
Ubuntu Ethics
Ubuntu philosophy emphasizes communal harmony and interconnectedness. Despite rich philosophical tradition, virtually no standardized LLM benchmarks exist.
Buddhist Ethics
Buddhist moral reasoning (ahimsa, karma, middle path) influences billions but has near-zero representation in LLM evaluation literature.
Hindu Ethics
Dharmic traditions cover complex moral philosophy (dharma, karma, ahimsa) for 1.2B+ people — essentially absent from benchmarks.
Indigenous Ethics
First Nations, Aboriginal, and Indigenous ethical frameworks emphasize land, kinship, and reciprocity — completely absent from LLM evaluation.
Latin American Ethics
Liberation theology, buen vivir, and Latin American moral philosophy remain underexplored in AI ethics benchmarking.
Southeast Asian Ethics
Diverse moral traditions across ASEAN nations (Filipino bayanihan, Thai Buddhist ethics, Indonesian pancasila) lack dedicated benchmarks.
Normative ethics prescribes what ought to be — unlike descriptive approaches such as Moral Foundations Theory that catalogue how people actually reason. For LLM assessment, normative benchmarks test whether models can reason consistently within established ethical frameworks: applying virtues, following duties, or maximizing welfare. This section maps the vault's coverage across the three major normative traditions.
Classification Distribution
Framework Distribution
Framework Timeline
Framework Coverage Gap Matrix
Tier 3: Advanced Analytics
Meta-Ethics Concepts
Philosopher Citations
Author × Framework
Full-Text Keyword Density
Virtue Ethics Papers
0Virtue Ethics (Aristotle, Confucius) evaluates moral character rather than actions or outcomes. It asks whether an agent — including an LLM — exhibits or simulates virtues such as honesty, prudence, justice, and temperance. As LLMs increasingly serve as conversational agents, tutors, and advisors, VE benchmarks test whether their responses embody the phronesis (practical wisdom) and eudaimonia (human flourishing) that character-based ethics demands.
Key benchmarks include JETHICS (78K Japanese scenarios with a dedicated VE category), BehaviorBench (testing alignment via the ETHICS dataset's VE subset), and CharacterEval / CharMoral for measuring character-level moral reasoning. Agarwal et al. (2024) uniquely operationalise VE across six languages, while Ohlhorst (2025) proposes a virtue framework specifically for AI.
Gaps: VE benchmarks remain the least developed of the three traditions. Most VE papers are theoretical rather than empirical. Confucian VE and eudaimonic well-being evaluations remain nascent paradigms with virtually no standardised test sets.
Year Distribution
Top Models
Research Type
Venue Types
Geographic Focus
Virtues Studied
Engagement Depth
Keyword Co-occurrence
Sub-Traditions
Concept Evolution
Virtue Ethics Papers
0Deontological Ethics Papers
0Deontological Ethics (Kant) judges actions by whether they conform to rules, duties, and the categorical imperative — 'act only according to that maxim you could will as a universal law.' For LLMs, deontological benchmarks test whether models can identify duties, respect rights, reason about obligations, and distinguish permissible from impermissible actions independent of consequences.
Key benchmarks: Samway et al. (2025) present 600+ trolley-problem variants and find that chain-of-thought prompting skews models toward deontological answers; Lu et al. (2025) introduce a Kantian κ parameter for calibrating LLM moral reasoning; the ETHICS dataset (Hendrycks 2021) includes a dedicated deontology subset; and JETHICS covers Japanese deontological norms.
Gaps: Current benchmarks focus on simple rule-following. Tests for perfect vs. imperfect duties, Ross's prima facie duties, and multi-stakeholder rights conflicts remain largely absent from the benchmark literature.
Year Distribution
Top Models
Research Type
Venue Types
Geographic Focus
Deontological Concepts
Engagement Depth
Keyword Co-occurrence
Sub-Traditions
Concept Evolution
Deontological Ethics Papers
0Consequentialism/Utilitarianism Papers
0Consequentialism / Utilitarianism (Bentham, Mill) judges actions solely by their outcomes — 'the greatest good for the greatest number.' For LLMs, consequentialist benchmarks test trade-off reasoning, welfare aggregation, and whether models can compute or approximate utility functions under uncertainty.
Key benchmarks: the Greatest Good Benchmark (Marraffini 2024) directly tests utilitarian reasoning; Moore et al. (2024) compare Utilitarian Sum vs. Nash Product welfare functions; Keshmirian et al. (2025) show multi-agent deliberation boosts utilitarian alignment. Moral Machine derivatives test utilitarian trade-offs in autonomous driving scenarios.
Gaps: Most tests probe simple utility calculus (trolley-like dilemmas). Rule utilitarianism, preference utilitarianism, and satisficing consequentialism remain essentially untested in the LLM evaluation literature.
Year Distribution
Top Models
Research Type
Venue Types
Geographic Focus
CU Concepts Studied
Engagement Depth
Keyword Co-occurrence
Sub-Traditions
Concept Evolution
Consequentialism/Utilitarianism Papers
0Care Ethics Papers
0Care Ethics (Gilligan, Noddings, Held) centres morality on relationships, empathy, and responsiveness to others' needs — in contrast to impartial, rule-based frameworks. For LLMs, care-ethics benchmarks test whether models can recognise relational context, respond with appropriate empathy, and navigate care-giving dilemmas where abstract principles fall short.
Gaps: Care ethics remains under-represented in LLM evaluation. Most work focuses on empathy detection rather than testing care-based moral reasoning as a distinct normative framework.
Year Distribution
Top Models
Research Type
Venue Types
Geographic Focus
CE Concepts Studied
Engagement Depth
Keyword Co-occurrence
Care Ethics Papers
0Contractualism Papers
0Contractualism / Rawlsian Ethics (Rawls, Scanlon) grounds morality in principles that rational agents would agree to under fair conditions. The 'veil of ignorance' thought experiment and Scanlon's 'what we owe to each other' test whether decisions can be justified to all affected parties. For LLMs, contractualist benchmarks probe fairness reasoning, distributive justice, and whether models can simulate impartial deliberation.
Gaps: Most contractualist LLM work focuses on fairness metrics rather than testing Rawlsian reasoning directly. Scanlon-style justifiability tests and difference-principle evaluations are nascent.
Year Distribution
Top Models
Research Type
Venue Types
Geographic Focus
CO Concepts Studied
Engagement Depth
Keyword Co-occurrence
Contractualism Papers
0Principlism Papers
0Principlism (Beauchamp & Childress) organises bioethics around four principles — autonomy, beneficence, non-maleficence, and justice. Originally developed for medical ethics, it increasingly appears in AI ethics discussions around healthcare AI, clinical decision support, and patient-facing LLMs.
Year Distribution
Principlism Concepts
Principlism Papers
0Most real-world ethical dilemmas do not fit neatly into a single tradition. Papers tagged with more than one normative framework — virtue ethics, deontological, or consequentialist — reveal where the literature engages in genuinely pluralistic moral reasoning. This sub-tab maps those overlaps: which framework combinations appear most often, how multi-framework scholarship has grown over time, whether certain models or geographic regions attract more cross-tradition analysis, and how benchmark density compares across single- versus multi-framework papers.
Framework Overlap
Benchmark Density by Combination
Comparative Radar
Overlap Typology
Overlap Growth Over Time
Framework Exclusivity
Top Models in Overlapping Papers
Framework Combination × Region
Overlapping Papers
Normative × Descriptive Bridge: This intersection reveals which papers engage with both prescriptive ethical frameworks (VE/DE/CU) and descriptive psychological theories (MFT, Schwartz, etc.). Papers at this intersection are uniquely positioned to test whether LLMs can distinguish between how people do reason morally and how they ought to.
Normative × Moral Psych Heatmap
Bridge Papers Over Time
Bridge Papers
0Moral psychology studies how people actually reason about morality — empirical and descriptive, not prescriptive. This is fundamentally different from normative ethics (deontology, utilitarianism, virtue ethics), which asks what ought to be. For LLM alignment, moral psychology frameworks reveal whether models replicate human moral intuitions, developmental stages, and value priorities — or diverge in systematic ways.
Key frameworks include Moral Foundations Theory (Haidt/Graham: 6 innate moral foundations), Schwartz Basic Human Values (10 universal values in 4 higher-order groups), Kohlberg's Cognitive Developmental stages, the Trolley/Moral Machine paradigm, and Dual Process Theory (Greene: automatic vs. controlled moral reasoning). This section maps which theories the vault's literature actually engages with.
Theory Distribution
Theory Timeline
Theory Dominance Shift (%)
WEIRD vs Non-WEIRD by Theory
Theory Co-occurrence
Moral Foundations Theory
0Moral Foundations Theory (Haidt & Graham, 2007) represented a paradigm shift in moral psychology, challenging decades of rationalist models by arguing that moral judgment is primarily driven by rapid intuitive responses rather than deliberative reasoning. MFT identifies six innate psychological foundations: Care/Harm, Fairness/Cheating, Loyalty/Betrayal, Authority/Subversion, Sanctity/Degradation, and Liberty/Oppression. Crucially, these foundations are not equally weighted across cultures. Research consistently shows that WEIRD (Western, Educated, Industrialized, Rich, Democratic) populations disproportionately emphasize Care and Fairness, while non-WEIRD populations and politically conservative groups draw more evenly on all six foundations. This cross-cultural variability is precisely what makes MFT indispensable for alignment research.
For values alignment, MFT offers something rare: a psychometrically validated instrument — the Moral Foundations Questionnaire (MFQ) — with large-scale cross-cultural norming data. When researchers administer the MFQ to an LLM, they can directly compare the model's moral foundation profile against human population baselines. The central questions are whether LLMs exhibit stable and coherent foundation profiles, whether those profiles shift under different prompting conditions or persona instructions, and whether models trained predominantly on English-language text reproduce the Care/Fairness skew characteristic of WEIRD populations. These are not merely academic concerns: an AI system deployed globally that systematically underweights Loyalty, Authority, or Sanctity may generate advice or judgments that feel alien or offensive to billions of users whose moral intuitions rest on those foundations.
The analyses below map the current state of MFT-LLM research. The Year Distribution chart tracks the rapid growth of this subfield. The Foundation Radar reveals which of the six foundations receive the most empirical attention and which remain understudied. Top Models shows concentration of testing across specific LLM families. Research Approach and Research Type distinguish benchmark-driven evaluation from theoretical work, while Geographic Focus and Venue Types expose whether the research community itself suffers from the same WEIRD sampling bias that MFT was designed to diagnose.
Year Distribution
Foundation Radar
Top Models
Research Approach
Venue Types
Geographic Focus
Research Type
Moral Foundations Theory
0Schwartz Value Theory
0Schwartz's Theory of Basic Human Values (1992) is arguably the most extensively validated cross-cultural framework in all of moral and social psychology. It identifies 10 universal value types — Self-Direction, Stimulation, Hedonism, Achievement, Power, Security, Conformity, Tradition, Benevolence, Universalism — organized in a circular motivational continuum where adjacent values are psychologically compatible and opposing values conflict. These map onto four higher-order groups: Openness to Change vs. Conservation, and Self-Enhancement vs. Self-Transcendence. Validated across 80+ countries using the Schwartz Value Survey (SVS) and the Portrait Values Questionnaire (PVQ), this circular structure has proven remarkably stable across cultures, making it a powerful diagnostic lens for examining what values AI systems implicitly prioritize.
For values alignment, Schwartz theory offers a uniquely structural advantage: because the 10 values form a circumplex with predictable trade-off relationships, researchers can test not just which values an LLM endorses, but whether the model's value profile exhibits the coherent motivational structure observed in human populations. The World Values Survey (WVS), the largest cross-cultural values dataset in the social sciences, provides population-level baselines against which LLM value profiles can be compared. Key alignment questions include whether models trained on multilingual corpora exhibit different value hierarchies depending on the language of interaction, whether fine-tuning for helpfulness systematically shifts models toward Self-Transcendence values at the expense of Conservation values, and whether value profiles remain consistent across semantically equivalent prompts — a basic test of whether the model has internalized coherent values rather than producing contextually triggered surface responses.
The analyses below unpack these questions empirically. The Value Dimensions Radar shows which of the 10 values receive concentrated research attention and which remain in the periphery. Higher-Order Quadrants reveals the balance across the four poles of the circumplex. Year Distribution and Top Models track the growth trajectory and model coverage of Schwartz-based evaluation. Geographic Focus is particularly informative here: Schwartz's own work demonstrated systematic cross-national value differences, so the geographic distribution of AI-values research reveals whether the field is leveraging or ignoring that foundational insight.
Year Distribution
Value Dimensions Radar
Top Models
Higher-Order Quadrants
Venue Types
Geographic Focus
Research Type
Schwartz Value Theory
0Moral Reasoning (Kohlberg)
0Moral reasoning research in the Kohlbergian tradition examines how moral judgment develops through qualitatively distinct stages of cognitive sophistication. Lawrence Kohlberg's (1958, 1984) cognitive-developmental theory proposes three levels — preconventional (self-interest), conventional (social norms and law), and postconventional (universal ethical principles) — each comprising two stages, for six stages total. The Defining Issues Test (DIT/DIT-2), developed by James Rest, operationalizes this framework as a standardized psychometric instrument and remains the most widely used measure of moral reasoning maturity. While Kohlberg's stage model has faced important critiques — notably Carol Gilligan's argument that it undervalues care-oriented reasoning and cross-cultural challenges to its claim of universality — it continues to provide the dominant developmental framework for studying moral cognition.
For values alignment, Kohlbergian stage theory raises a provocative question: at what developmental level does an LLM reason? Early findings suggest that large language models tend to cluster at the conventional level (Stages 3–4), producing responses oriented toward social conformity, rule-following, and maintaining social order. This is significant because RLHF and safety fine-tuning may systematically reinforce conventional-level reasoning — training models to defer to established norms rather than engage in principled, postconventional moral reflection. Whether an AI system should reason at Stage 5 (social contract) or Stage 6 (universal principles) is itself an open alignment question, but understanding where models currently fall on the developmental spectrum is essential for making that determination deliberately rather than by accident.
The charts below profile this subfield. Year Distribution tracks research activity over time — notably, Kohlberg-based LLM evaluation is a younger and smaller subfield than MFT or Schwartz-based work. Top Models reveals which LLM families have been tested for moral reasoning maturity. Research Type distinguishes empirical evaluation from theoretical discussion, while Venue Types and Geographic Focus illuminate the disciplinary and cultural origins of this research stream.
Year Distribution
Top Models
Research Type
Venue Types
Geographic Focus
Moral Reasoning Papers
0Moral Dilemmas (Trolley Problems)
0Moral dilemma research has a distinguished lineage in both philosophy and experimental psychology. Philippa Foot (1967) and Judith Jarvis Thomson (1985) formalized the trolley problem as a tool for probing moral intuitions about harm, intention, and the act/omission distinction. Joshua Greene's neuroimaging work (2001–2009) transformed the field by demonstrating that personal moral dilemmas (pushing someone off a bridge) and impersonal ones (pulling a lever) activate fundamentally different neural systems — an insight formalized as dual-process theory, which distinguishes automatic emotional responses from controlled utilitarian reasoning. More recently, the Moral Machine experiment (Awad et al., 2018) scaled dilemma research to millions of participants across 233 countries, revealing systematic cross-cultural variation in moral preferences for autonomous vehicle decision-making.
For values alignment, moral dilemmas are uniquely revealing because they force systems to make explicit trade-offs that expose implicit value hierarchies. When an LLM resolves a trolley-type scenario, its response reveals whether it defaults to consequentialist aggregation, deontological constraints, or some hybrid strategy — and whether that default is stable across framing variations. Dual-process theory raises a further question: do LLMs exhibit anything analogous to the emotional-automatic versus controlled-deliberative distinction observed in humans, or do they produce a single-process output that merely pattern-matches to one reasoning style? Beyond the laboratory, moral dilemma research has direct deployment implications for autonomous vehicles, medical triage systems, resource allocation algorithms, and any AI application where conflicting stakeholder interests demand principled resolution. The pronounced cross-cultural variation documented by the Moral Machine project underscores that no single resolution pattern will satisfy global moral intuitions.
The visualizations below map this research landscape. Year Distribution reveals the growth trajectory of dilemma-based LLM evaluation. The Dilemma Paradigms chart — unique to this section — shows which experimental designs dominate (classic trolley variants, Moral Machine adaptations, medical scenarios, autonomous driving, or novel constructions) and whether the field is diversifying beyond its philosophical origins. Top Models, Research Type, Venue Types, and Geographic Focus round out the picture, revealing the models, methods, publication venues, and cultural contexts that shape this critical area of alignment research.
Year Distribution
Dilemma Paradigms
Top Models
Research Type
Venue Types
Geographic Focus
Moral Dilemma Papers
0Cross-Framework Analysis
How do moral psychology theories interact with AI models, normative ethics frameworks, and research methodologies? This section provides cross-cutting analytical views.
Model × Theory Heatmap
Normative Ethics Bridge
Alignment Outcomes
Research Methodology
Instruments Used
Theoretical Sophistication
Research Blind Spots
Dual Process Theory
Zero CoverageGreene's automatic/controlled moral reasoning framework — foundational to moral cognition research — has no dedicated LLM benchmark despite explaining why humans give different answers to trolley variants.
Social Intuitionist Model
Zero CoverageHaidt's SIM (moral judgments are driven by intuition, not reasoning) challenges the assumption that LLM 'reasoning' reflects human moral cognition. No benchmarks test this.
Dyadic Morality
Zero CoverageGray & Schein's theory that all moral judgment reduces to perceived harm between an agent and patient — zero LLM evaluations despite growing empirical support.
Moral Identity
Minimal CoverageAquino & Reed's moral identity theory — how self-concept shapes moral behavior — is virtually untested in LLMs despite relevance to persona-based alignment.
Best Practices
Use multiple complementary frameworks, not just one. MFT captures moral intuitions but misses deliberative reasoning — complement with DIT. Schwartz values measure cross-cultural value priorities — ideal for alignment testing. Trolley paradigms are overused and ecologically invalid — complement with everyday moral scenarios.
Moral development stages (Kohlberg) can assess LLM reasoning sophistication — does the model reason at pre-conventional, conventional, or post-conventional levels?
Descriptive alignment ≠ normative correctness: An LLM that perfectly matches human moral psychology may still be 'wrong' by normative standards. Matching human biases and heuristics is not the same as ethical reasoning.
Religious ethics shape moral reasoning for over 5 billion people worldwide — 1.2B Muslims, 2.4B Christians, 1.2B Hindus, 500M Buddhists, 400M+ practitioners of East Asian traditions, and hundreds of millions following Indigenous and African moral philosophies. LLMs deployed globally must navigate these diverse moral frameworks, yet the current benchmark landscape is radically imbalanced: Islamic QA systems dominate with 11+ benchmarks, while most other traditions have zero or one dedicated evaluation.
This section maps which religious ethical traditions the vault's literature engages with, identifies the most critical coverage gaps, and outlines what rigorous evaluation would require for each tradition. The goal is not to rank or compare religions, but to ensure LLM evaluation reflects the full spectrum of human moral reasoning.
Tradition Distribution
Tradition Timeline
Tradition Co-occurrence
Islamic Ethics
0Islamic ethics derives from Usul al-Fiqh (principles of jurisprudence), with the Maqasid al-Shariah (5 higher objectives: protection of religion, life, intellect, progeny, and wealth) as the overarching framework. Moral actions are classified on a 5-point scale: fard/wajib (obligatory), mandub/mustahabb (recommended), mubah (permissible), makruh (discouraged), and haram (forbidden).
The fatwa reasoning chain — question → Quran → Hadith → Ijma (consensus) → Qiyas (analogy) — represents a structured, source-hierarchy-driven form of moral reasoning. LLM evaluation in this space is the most developed among religious traditions, with benchmarks for Quranic QA, Hadith verification, Fatwa generation, and Sharia compliance checking.
Key challenge: Diversity within Islam (Sunni/Shia, 4 madhahib schools of jurisprudence), context-sensitivity of fatwas, and the critical safety concern of models 'hallucinating' religious rulings for 1.2B potential users.
Year Distribution
Top Models
Islamic Knowledge Domains
Geographic & Linguistic Origins
Resource Building vs. Bias Detection
Islamic Textual Source Coverage
Cross-Tradition Connections
Islamic Ethics
0Christian Ethics
0Christian ethics spans multiple traditions: the Natural Law tradition (Aquinas) holds that morality is discoverable by reason and grounded in human nature. Divine Command Theory grounds morality in God's will. Thomistic virtue ethics adds theological virtues (faith, hope, charity) to the cardinal virtues. Protestant traditions emphasize sola scriptura (scripture alone), while Liberation Theology prioritizes the 'preferential option for the poor.'
LLM application: Bible QA systems exist, but few test moral reasoning. Challenges include denominational diversity (Catholic vs. Protestant vs. Orthodox), faith-reason tensions, and translation issues across Bible versions. The Imago Dei principle — that humans are made in God's image — raises critical questions about whether AI systems can possess moral status within Christian anthropology.
Year Distribution
Top Models
Denominational Landscape
Theological Concepts
Sacred Text Coverage
Research Focus
Normative Ethics Overlap
Cross-Tradition Connections
Christian Ethics
0Jewish Ethics
0Jewish ethics centers on Halakhic reasoning — Jewish law derived from Torah, Talmud, and rabbinic commentary. Key principles include Tikkun Olam (repairing the world), Pikuach Nefesh (sanctity of life overrides most commandments), Tzedakah (righteous giving as obligation), and Mussar (character development). B'tselem Elohim (in the image of God) anchors human dignity.
LLM application: Halakhic reasoning — structured, precedent-based, and analogical — is arguably well-suited for LLM evaluation but has essentially zero dedicated benchmarks. The 2,000+ year tradition of structured ethical argumentation (Talmudic dialectic) represents an untapped resource for AI ethics evaluation.
Note: With only ~5 papers, some charts below may be sparse. This itself highlights the critical coverage gap.
Year Distribution
Top Models
Core Concepts
Textual Sources
Research Focus
Normative Ethics Overlap
Cross-Tradition Connections
Jewish Ethics
0Buddhist Ethics
0Buddhist ethics derives from the Four Noble Truths, the Eightfold Path, and the Three Marks of Existence (impermanence, suffering, non-self). Core moral principles include ahimsa (non-harm), karuna (compassion), metta (loving-kindness), and upaya (skillful means). The Middle Way serves as both metaphysical and ethical principle. Dependent Origination (pratityasamutpada) and Sunyata (emptiness) provide the metaphysical ground for ethical reasoning.
LLM application: Ahimsa aligns naturally with LLM safety/harmlessness objectives. The Humanistic Buddhism Corpus (HBC) is the most relevant benchmark. Buddhist ethics spans diverse schools — Theravada, Mahayana, Vajrayana, Zen, and Secular Buddhism — each with distinct emphases that challenge monolithic evaluation.
Year Distribution
Top Models
School Detection
Core Concepts
Textual Sources
Research Focus
Cultural Contexts
Language Coverage
Cross-Tradition Connections
Buddhist Ethics
0Hindu Ethics
0Hindu ethics centers on dharma as both cosmic order and individual moral duty. Karma (moral causation across lifetimes), the four purusharthas (goals: dharma, artha, kama, moksha), and varna-ashrama-dharma (context-dependent duty) create a deeply contextual ethical framework. Svadharma — one's personal duty determined by role and circumstance — challenges LLMs' need for universal rules.
LLM application: Ahimsa (non-harm) is shared across Indic traditions. The textual tradition spans the Vedas, Upanishads, Bhagavad Gita, and Dharmashastra. BRAND (2025) covers Hindu religious bias but near-zero benchmarks test dharma-based moral reasoning despite 1.2B Hindu adherents. Six philosophical schools (Vedanta, Samkhya, Yoga, Nyaya, Vaisheshika, Mimamsa) offer distinct epistemological and ethical frameworks.
Year Distribution
Top Models
Textual Tradition
Core Concepts
Philosophical Schools
Research Focus
Cultural Contexts
Language Coverage
Cross-Tradition Connections
Hindu Ethics
0Confucian Ethics
0Confucian ethics centers on five core virtues: Ren (仁, benevolence/humaneness), Li (禮, ritual propriety), Yi (義, righteousness), Zhi (智, wisdom), and Xin (信, fidelity/trustworthiness). The Five Relationships — ruler-subject, parent-child, husband-wife, elder-younger sibling, friend-friend — each carry reciprocal duties that are hierarchical but mutually obligated.
The Junzi (君子, exemplary person) is the moral ideal, achieved through study, reflection, and practice. Mencius' four innate moral sprouts — compassion → ren, shame → yi, deference → li, right/wrong → zhi — provide a developmental moral psychology. Social harmony (和, he) serves as the overriding moral goal.
LLM application: Confucian ethics is relational and contextual — who is speaking to whom fundamentally matters. This differs from rule-based Western ethics that most LLMs are trained on. STORAL-ZH and CMoralEval benchmark Chinese moral reasoning with Confucian grounding. Chinese LLMs (Qwen, ChatGLM, Baichuan) may exhibit Confucian reasoning patterns not seen in Western-trained models.
Year Distribution
Top Models
Confucian-Sphere Cultural Footprint
Western Framework Overlap
Chinese vs. Western LLMs
CJK Language Coverage
Empirical vs. Theoretical Trajectory
Confucian Concept Frequency
Confucian Ethics
0Shinto Ethics
0Shinto, Japan's indigenous spiritual tradition, centers on kami (spirits/gods) inhabiting natural phenomena and the concept of wa (harmony). Its emphasis on ritual purity (harae), coexistence with nature, and animistic worldview has generated a unique strain of 'techno-animism' literature exploring whether robots and AI can possess kami-like qualities — a question with no parallel in Abrahamic or Dharmic traditions.
LLM application: Shinto's animistic framework raises unique questions about AI moral status, robot personhood, and the ethics of anthropomorphism in Japanese cultural context. The concept of musubi (creative interconnection) offers a relational framework for understanding human-AI interaction.
Year Distribution
Top Models
Core Concepts
Research Themes
Cultural Context
Cross-Tradition Connections
Gap vs. Vault Average
Shinto Ethics
0Ubuntu Ethics
0Ubuntu ethics is grounded in the principle umuntu ngumuntu ngabantu — "a person is a person through other persons." Personhood is not a given but is achieved through relationships, communal participation, and moral conduct. Core values include communal harmony, consensus-based decision-making (indaba), restorative justice, and the inseparability of individual well-being from collective flourishing. Ubuntu rejects atomistic individualism in favor of relational ontology.
LLM application: Ubuntu's relational ontology challenges the atomistic, individual-agent framing of most LLM evaluations. There are currently zero dedicated benchmarks for Ubuntu moral reasoning — a critical gap given that Ubuntu offers fundamentally different conceptions of personhood, agency, and moral responsibility than Western individualist ethics.
Year Distribution
Top Models
Core Concepts
Research Themes
Pan-African Geographic Spread
Gap vs. Vault Average
Cross-Tradition Connections
Ubuntu Ethics
0Indigenous Ethics
0Indigenous ethical frameworks span thousands of distinct traditions — First Nations, Aboriginal Australian, Native American, Maori, and many others — yet share common threads: kinship obligations extending to land and non-human beings, reciprocity with nature, intergenerational responsibility (Seven Generations principle), and oral tradition as the primary medium for moral knowledge. Ethics is embedded in place, story, and ceremony rather than codified in texts.
LLM challenge: 5,000+ distinct Indigenous peoples cannot be reduced to a single benchmark. Data sovereignty frameworks (OCAP, CARE principles) complicate whether such benchmarks should be created without community governance. These traditions are oral rather than textual (limited training data), community-embedded rather than individual (moral authority is collective), and land-based rather than abstract.
Year Distribution
Top Models
Specific Traditions
Core Themes
Geographic Origins
Gap vs. Vault Average
Cross-Tradition Connections
Indigenous Ethics
0Other Traditions
0Additional religious traditions — including Jainism, Sikhism, Taoism, Baha'i, and Zoroastrianism — each contain rich ethical frameworks with distinct implications for AI evaluation. Jain ethics emphasizes extreme non-violence (ahimsa) and epistemological humility (anekantavada — many-sidedness). Sikh ethics centers on selfless service (seva) and radical equality. Taoist ethics teaches wu wei (effortless action) and harmony with the Dao.
Note: These traditions have minimal representation in the vault (0-2 papers each), reflecting a critical gap in AI ethics research coverage.
Year Distribution
Top Models
Tradition Summary
Key Concepts per Tradition
Coverage Gap
Other Traditions
0Religious Ethics: Landscape & Gaps
The landscape is radically imbalanced: Islam has 11+ benchmarks, most other traditions have 0–1. Religious QA (factual recall about scriptures) is not the same as religious ethics benchmarking (moral reasoning within a tradition's framework).
Critical needs: Benchmarks testing moral reasoning within each tradition, not just factual recall. Cross-tradition consistency tests — does an LLM reason differently about the same dilemma when framed in Islamic vs. Christian vs. Buddhist terms? And religious sensitivity — avoiding benchmarks that reduce traditions to stereotypes.
Missing entirely: Sikhism (30M+ adherents, seva and equality principles), Taoism/Daoism (wu wei, balance of opposites), Shinto (purity, harmony with nature), Jainism (most rigorous ahimsa tradition, anekantavada/non-absolutism), and most Indigenous ethical systems are completely absent from LLM evaluation.
Which LLMs get evaluated — and which do not — defines the boundaries of what we know about AI alignment. This section maps the landscape of model evaluation across the vault's literature, drawn from both LLM assessment metadata and benchmark survey results. The distribution reveals a pronounced concentration: a small number of proprietary model families (GPT, Claude, LLaMA) dominate the evaluation literature, while most open-source and non-English-centric models remain under-tested. The timeline tracks the rapid post-2022 expansion of evaluation work, and the repository chart shows where researchers share their evaluation artifacts.
LLM Model Families
Model Coverage Timeline
Data Repository Distribution
LLM-Assessed Papers
0Model depth analysis reveals which LLMs have been evaluated broadly across ethical frameworks and cultural contexts, versus those tested narrowly. A model tested only on WEIRD populations with Western ethical frameworks has unknown alignment properties for the majority of the world's population.
Model x Ethical Framework
Model Cultural Exposure
Open vs. Closed Source Models
Model Coverage Breadth
Model Depth Papers
0Who publishes on AI ethics — and how they collaborate — shapes the field's direction. This tab maps the 9,300+ researchers behind the vault's 2,043 papers, revealing patterns of productivity, team formation, and collaboration networks. The median team size of 4 co-authors reflects AI ethics' interdisciplinary character, while the growth of new entrants each year shows a field still rapidly expanding.
Team Size Distribution
Author Productivity (Lotka's Law)
New vs Returning Authors per Year
Median Team Size Over Time
A small number of researchers shape the field disproportionately. The top 30 authors are shown with their position breakdown — first author (principal contributor), last author (typically lab director), or middle position. The radar chart reveals how the most prolific voices distribute across the vault's six analytical dimensions, while the timeline heatmap shows whether their output is sustained or concentrated in bursts.
Top 30 Authors by Paper Count
Author Papers
0Top 10 Author Subfield Radar
Top 15 Author Timeline Heatmap
Collaboration networks reveal the social fabric of research. The force-directed graph maps co-authorship ties among the 80 most prolific authors — node size reflects paper count, edge thickness reflects shared publications. The chord diagram shows the strongest bilateral collaboration channels among the top 30 researchers. Together, they expose both tight research clusters and isolated silos.
Collaboration Network (Top 80 Authors)
Top 20 Co-Author Pairs
Team Composition
Co-Authorship Chord Diagram (Top 30)
Who gets to define AI ethics? The concentration of authorship matters: a Lorenz curve reveals how evenly publication output is distributed. AI Lab affiliations show the role of major industry players, while the equality line in the Lorenz chart provides a benchmark for comparison.
AI Lab Contributions
Lorenz Curve — Publication Equality
Author metadata quality directly affects the reliability of collaboration analysis. This section provides transparency on data completeness — how many papers have full author lists, how many names are abbreviated, and what fraction of authors have enriched profiles.