{"path":"research/normative-reasoning.md","content":"# Computational Handling of Normative Claims: The \"Should\" Problem\n\n*Research compiled March 27, 2026. Deep dive into how deliberation platforms can handle normative/ethical claims alongside factual ones.*\n\n---\n\n## Table of Contents\n\n1. [The Fact/Value Boundary](#1-the-factvalue-boundary)\n2. [Moral Foundations Theory](#2-moral-foundations-theory)\n3. [Structured Values Elicitation](#3-structured-values-elicitation)\n4. [Computational Ethics](#4-computational-ethics)\n5. [LLMs and Moral Reasoning](#5-llms-and-moral-reasoning)\n6. [Normative Argument Schemes](#6-normative-argument-schemes)\n7. [Deliberative Polling](#7-deliberative-polling)\n8. [The \"Ethical Filter\" Concept](#8-the-ethical-filter-concept)\n9. [Value-Sensitive Design](#9-value-sensitive-design)\n10. [Practical Implications for Deliberus](#10-practical-implications-for-deliberus)\n\n---\n\n## 1. The Fact/Value Boundary\n\n### Hume's Guillotine (The Is-Ought Problem)\n\nDavid Hume observed that writers frequently shift from \"is\" statements (descriptive) to \"ought\" statements (prescriptive) without justification. His thesis: **an ethical or normative conclusion cannot be inferred from purely descriptive factual statements.** This is sometimes called \"Hume's Guillotine\" — it severs the logical bridge from facts to values.\n\n**Key implications for Deliberus:**\n\n- **Factual claims** (\"Sweden's CO2 emissions decreased 5% in 2025\") can be verified empirically\n- **Normative claims** (\"Sweden should increase carbon taxes\") require *both* factual premises *and* value premises\n- **The gap is not a ban on reasoning** — it insists that normative conclusions must make their value premises *explicit*. Factual discoveries must be supplemented by explicit value premises to yield normative conclusions\n\nThis is actually an **opportunity** for a deliberation platform: by forcing users to decompose normative arguments into their factual and value components, the platform can make the hidden value premises visible — which is exactly what Hume was asking for.\n\n### How Existing Systems Handle the Boundary\n\nMost argumentation platforms (Kialo, Argüman, DebateGraph) **do not** distinguish between factual and normative claims at the data model level. They treat all claims as undifferentiated nodes in an argument tree. This is a significant gap.\n\n**Approaches that do engage with the boundary:**\n\n- **ASPIC+ framework**: Distinguishes between \"brute facts,\" \"institutional facts,\" and \"deontic facts\" (obligations, permissions). Normative arguments are represented as sequences of (brute, institutional, deontic) triples, with constitutive norms deriving institutional facts from brute facts, and regulative norms deriving deontic facts from institutional facts\n- **AIF (Argument Interchange Format)**: The upper ontology distinguishes information nodes (I-nodes) from scheme application nodes (S-nodes), and separates normative structures of inference from descriptive argument material. However, it does not have a first-class \"normative claim\" type\n- **Walton's Argumentation Schemes** (2008): Explicitly includes normative schemes (argument from values, argument from consequences) with critical questions that probe the value assumptions. The most relevant existing framework for Deliberus\n\n### Can LLMs Classify Claims as Descriptive vs. Normative?\n\nRecent research (2024-2025) shows progress:\n\n- **BENR dataset** (Bioethical Explanations and Normative Reasoning): Designed to expose the internal structure of normative explanations in NLI tasks. Logic-internal approaches consistently outperform logic-external methods in normative reasoning\n- **Normative reasoning in LLMs** ([ACL 2025](https://aclanthology.org/2025.blackboxnlp-1.17.pdf)): Provides comparative analysis of how models handle deontic reasoning patterns\n- **Claimify** (Microsoft Research): Already in Deliberus's technical roadmap — decomposes text into atomic claims. Could be extended with a normative/descriptive classifier as a second pass\n- **Practical feasibility**: Current LLMs can reliably distinguish \"X is the case\" from \"X should be the case\" at the sentence level. The harder problem is **implicit normativity** — claims that appear factual but smuggle in value assumptions (e.g., \"Education spending is insufficient\")\n\n**Sources:**\n- [Is-Ought Problem — Wikipedia](https://en.wikipedia.org/wiki/Is%E2%80%93ought_problem)\n- [AI, Hume and a Guillotine](https://euideas.eui.eu/2021/04/21/ai-hume-and-a-guillotine-the-dangers-of-machine-learning-loops/)\n- [ASPIC-based Legal Argumentation for Deontic Reasoning](https://ebooks.iospress.nl/publication/37811)\n- [AIF Specification](http://www.arg-tech.org/wp-content/uploads/2011/09/aif-spec.pdf)\n- [Normative Reasoning in LLMs](https://aclanthology.org/2025.blackboxnlp-1.17.pdf)\n\n---\n\n## 2. Moral Foundations Theory\n\n### The Six Foundations (Haidt, Graham, et al.)\n\nMoral Foundations Theory (MFT) proposes that morality is \"more than one thing\" — humans have innate moral \"taste buds\" that shape moral judgment:\n\n| Foundation | Virtue / Vice | Evolutionary Origin |\n|-----------|--------------|-------------------|\n| **Care / Harm** | Compassion / Cruelty | Attachment, protecting offspring |\n| **Fairness / Cheating** | Justice / Injustice | Reciprocal altruism |\n| **Loyalty / Betrayal** | Patriotism / Treachery | Tribal coalitions |\n| **Authority / Subversion** | Respect / Rebellion | Hierarchical social interactions |\n| **Sanctity / Degradation** | Purity / Disgust | Pathogen avoidance |\n| **Liberty / Oppression** | Freedom / Tyranny | Resistance to domination |\n\nThe MFQ-2 (2023) split Fairness into **Equality** and **Proportionality**, based on a decade of empirical work by Mohammad Atari et al.\n\n### Political Mapping\n\nHaidt's core finding: Liberals rely primarily on Care + Fairness + Liberty. Conservatives use all six foundations roughly equally, including Loyalty, Authority, and Sanctity. This explains why political groups talk past each other — they are literally operating from different moral axioms.\n\n### Computational Detection of Moral Foundations\n\nA mature NLP subfield exists for detecting moral foundations in text:\n\n- **MoralStrength**: Semi-automated lexicon (~1,000 lemmas) quantifying moral relevance/strength per foundation via crowdsourced \"Moral Valence\" scores\n- **Extended Moral Foundations Dictionary (eMFD)**: Crowdsourced continuous scoring approach (vs. binary vice/virtue). Outperforms older MFD in multiple domains\n- **MoralBERT**: Fine-tuned language models for moral sentiment detection, trained on heterogeneous MFT-annotated datasets from Twitter, Reddit, Facebook\n- **ME2-BERT**: Integrates events and emotions for domain-invariant morality-relevant text representations\n- **Moral Foundations Twitter Corpus (MFTC)**: 35,108 tweets annotated for 10 categories of moral sentiment — benchmark dataset\n- **MFT-NLP benchmark**: [mft-nlp.com](https://www.mft-nlp.com/) tracks progress in NLP applications to MFT\n\n### Mapping to Argument Analysis for Deliberus\n\nMFT provides a powerful lens for understanding *why* people disagree on normative claims:\n\n- **Diagnostic use**: When two users disagree on a normative claim, the platform could identify which moral foundations each is invoking. \"You're arguing from Care; they're arguing from Sanctity — you're not actually disagreeing about facts\"\n- **Bridging use**: Inspired by Polis/Community Notes bridging algorithms — find normative positions that satisfy multiple moral foundations simultaneously\n- **Profiling use**: Users could take an MFQ-style assessment, producing a moral foundations profile. Arguments could then be tagged with which foundations they appeal to\n\n**Sources:**\n- [Moral Foundations Theory — Wikipedia](https://en.wikipedia.org/wiki/Moral_foundations_theory)\n- [moralfoundations.org](https://moralfoundations.org/)\n- [MFQ-2 Validation](https://www.sciencedirect.com/science/article/pii/S2451958825000247)\n- [eMFD Development](https://link.springer.com/article/10.3758/s13428-020-01433-0)\n- [MoralBERT](https://arxiv.org/html/2403.07678v2)\n- [Survey on MFT and Pre-trained Language Models (2025)](https://link.springer.com/article/10.1007/s00146-025-02225-w)\n- [MFT-NLP Benchmark](https://www.mft-nlp.com/)\n\n---\n\n## 3. Structured Values Elicitation\n\n### The Problem: Implicit Values\n\nMost normative disagreements are not about stated values but about **unstated** ones. Two people may agree on a policy conclusion but for completely different value reasons, or disagree on a policy while sharing the same values but weighting them differently. Making these implicit values explicit is essential for productive deliberation.\n\n### Schwartz's Theory of Basic Values\n\nShalom Schwartz identified **ten basic personal values** recognized across cultures, organized in a circular motivational structure:\n\n- **Self-Direction** (creativity, freedom, independence)\n- **Stimulation** (excitement, novelty, challenge)\n- **Hedonism** (pleasure, enjoyment)\n- **Achievement** (success, competence, ambition)\n- **Power** (authority, wealth, social recognition)\n- **Security** (safety, harmony, stability)\n- **Conformity** (obedience, self-discipline)\n- **Tradition** (respect, commitment, humility)\n- **Benevolence** (helpfulness, honesty, loyalty)\n- **Universalism** (equality, justice, environmental protection)\n\nThe circular structure captures **conflicts and compatibility**: adjacent values are compatible (e.g., Benevolence and Universalism); opposing values conflict (e.g., Power vs. Universalism). This structure appears to be culturally universal.\n\n### World Values Survey and the Inglehart-Welzel Map\n\nThe World Values Survey (WVS) maps societies on two dimensions:\n\n1. **Traditional vs. Secular-Rational** values (religion, authority, family vs. secularism, autonomy)\n2. **Survival vs. Self-Expression** values (economic/physical security vs. environmental protection, tolerance, participation)\n\nThese two dimensions explain >70% of cross-national variance. The WVS questionnaire (~250 questions, 400-800 measurable variables) provides a comprehensive framework for understanding value landscapes. Sweden consistently scores high on both Secular-Rational and Self-Expression dimensions — relevant context for a platform originating there.\n\n### Elicitation Methods for a Deliberation Platform\n\nResearch on online group deliberation for values elicitation (Afshar & Stranieri) shows:\n\n- **Structured elicitation is more productive online than face-to-face** — people find it difficult to discuss values in person but engage more freely in anonymous online deliberation\n- **Values arrived at through deliberation enjoy broader acceptance** than those imposed or surveyed\n- **Definitional method**: Participants read value definitions, then elicit subordinate constructs by answering \"what does someone who values X do? How do they behave?\" — operationalizes abstract values into concrete behaviors\n\n**Practical approach for Deliberus**: When a user makes a normative claim, the system could prompt: \"This claim seems to rely on [Value X]. Do you agree? What other values inform your position?\" Using Schwartz's circular model, the system could map value tensions and compatibilities visually.\n\n**Sources:**\n- [Schwartz Theory of Basic Values — Overview](https://scholarworks.gvsu.edu/cgi/viewcontent.cgi?article=1116&context=orpc)\n- [Online Group Deliberation for Values Elicitation](https://link.springer.com/chapter/10.1007/978-3-642-04941-5_17)\n- [Inglehart-Welzel Cultural Map](https://en.wikipedia.org/wiki/Inglehart%E2%80%93Welzel_cultural_map_of_the_world)\n- [World Values Survey](https://www.worldvaluessurvey.org/WVSContents.jsp?CMSID=Findings)\n- [Exploiting Value System Structure for Value-Aligned Decision-Making](https://link.springer.com/chapter/10.1007/978-3-031-58202-8_11)\n\n---\n\n## 4. Computational Ethics\n\n### Can You Compute Ethics?\n\nThree major ethical frameworks have been formalized computationally, each with different tractability:\n\n#### Deontic Logic (Deontological Ethics)\n\nDeontic logic is a branch of modal logic dealing with obligation (O), permission (P), and prohibition (F). Recent work (2024-2025):\n\n- **Deontic Temporal Logic** ([Priya T.V. et al., 2025](https://arxiv.org/html/2501.05765)): Uses first-order logic to define predicates and axioms capturing ethical requirements for AI systems. Expresses obligations, prohibitions, and permissions formally\n- **ASPIC+ deontic frameworks**: Non-monotonic variants of modal logics using argumentation theory to handle conflicting obligations\n- **s(CASP) deontic reasoning**: Goal-directed predicate answer set programming for modeling deontic logic\n- **Practical limitation**: Deontic logic handles clear rule conflicts well but struggles with the messiness of real moral reasoning where rules are vague and context-dependent\n\n#### Utilitarian Calculus (Consequentialist Ethics)\n\nBentham's felicific calculus evaluated actions across seven factors: **intensity, duration, certainty, propinquity, fecundity, purity, and extent** of pleasure/pain.\n\nComputational formulations:\n- **Combinatorial optimization**: Formulate as maximizing aggregate wellbeing subject to constraints, using integer programming\n- **Multi-criteria decision analysis**: Weigh different wellbeing factors, apply computational techniques\n- **Fundamental limitation**: \"Maximum happiness\" and \"minimum pain\" are subjective — the greatest happiness for the greatest number is notoriously difficult to compute because utilities are incommensurable across persons\n\n#### Computational Complexity of Ethics\n\nA 2024 analysis ([Springer](https://link.springer.com/article/10.1007/s10462-024-10732-3)) found that **most problems posed by normative ethical frameworks lead to tractability issues** — moral reasoning is computationally hard in formal systems. This suggests that practical platforms should use **heuristic and deliberative approaches** rather than attempting to compute \"correct\" moral answers.\n\n### The Delphi Experiment (Allen Institute)\n\nThe most ambitious attempt to build a computational moral judgment system: trained on 1.7 million ethical judgments (Commonsense Norm Bank), Delphi agreed with human ethical norms 92.1% in lab conditions, ~80% in the wild. **However, it inherited severe biases** — giving racially discriminatory responses — and was taken down. Key lesson: crowdsourced morality reflects existing biases, not ideal moral reasoning.\n\n**Implication for Deliberus**: Do not attempt to compute moral \"answers.\" Instead, structure the *process* of moral reasoning so that value assumptions are explicit, arguments are traceable, and diverse perspectives are represented.\n\n**Sources:**\n- [Deontic Temporal Logic for AI Ethics](https://arxiv.org/html/2501.05765)\n- [Computational Complexity of Ethics](https://link.springer.com/article/10.1007/s10462-024-10732-3)\n- [Computational Ethics — ScienceDirect (Trends in Cognitive Sciences, 2022)](https://www.sciencedirect.com/science/article/pii/S1364661322000456)\n- [Delphi: Can Machines Learn Morality?](https://arxiv.org/abs/2110.07574)\n- [Delphi Experiment — Nature Machine Intelligence (2024)](https://www.nature.com/articles/s42256-024-00969-6)\n\n---\n\n## 5. LLMs and Moral Reasoning\n\n### Can LLMs Decompose Normative Arguments?\n\nThis is the critical question for Deliberus's AI-assisted claim decomposition pipeline.\n\n#### Current Capabilities (2024-2025)\n\n**Structured moral reasoning frameworks** ([EMNLP 2025](https://arxiv.org/html/2506.14948v1)):\n- Structured prompts using First-Principles Reasoning and Schwartz + Care Ethics produce norm-aligned decisions\n- Probing classifiers can detect whether moral dimensions (care, fairness, dignity) are linearly separable in embedding spaces\n- Singular value decomposition identifies \"principal moral axes\" within high-dimensional embeddings\n- Models like GPT-4 show internal consistency but **contradictory behavior between abstract values and concrete moral evaluations**\n\n**LLM argument mining** ([Survey, 2025](https://arxiv.org/html/2506.16383v4)):\n- Chain-of-thought prompts that model the \"logical leap between premise and conclusion\" account for most performance gains\n- Symbolic scaffolding (Toulmin-style or Walton-style templates) before verdict emission improves structure\n- LLMs can detect argumentative components (claims, premises, evidence) and relationships (attack, support)\n- **Critical limitation**: \"LLMs lack an explicit representation of argumentative structure\" — they struggle when relations span paragraphs, rely on implicit premises, or cross language boundaries\n\n**Three-dimensional ethics benchmark** ([Scientific Reports, 2025](https://www.nature.com/articles/s41598-025-18489-7)):\n- Evaluates LLMs across foundational moral principles, reasoning robustness, and value consistency\n- Models show progress but still exhibit gaps between principled reasoning and applied judgment\n\n#### Multilingual Moral Reasoning\n\nResearch finds GPT-4 is the most consistent ethical reasoner across languages, while other models show significant moral value bias in non-English languages. Relevant for a platform targeting Swedish-language deliberation.\n\n#### What LLMs Can Do for Deliberus\n\n1. **Claim classification**: \"Is this claim factual or normative?\" — reliably achievable\n2. **Value premise extraction**: \"What values does this normative argument assume?\" — achievable with structured prompting (Walton templates)\n3. **Moral foundation tagging**: \"Which moral foundations does this argument invoke?\" — achievable using MoralBERT-style models or prompted LLMs\n4. **Argument scheme detection**: \"Is this an argument from consequences? From values? From authority?\" — achievable with Walton scheme templates\n5. **Consistency checking**: \"Does this person's stated values match their argument?\" — partially achievable, but fragile\n\n#### What LLMs Cannot Do (Yet)\n\n1. **Determine whether a normative claim is \"correct\"** — this is category-inappropriate\n2. **Reliably detect implicit normativity** — claims that look factual but encode values\n3. **Handle incommensurable values** — comparing Care vs. Liberty when they conflict\n4. **Maintain consistency** between abstract principles and concrete cases\n\n**Sources:**\n- [Structured Moral Reasoning in LLMs (EMNLP 2025)](https://arxiv.org/html/2506.14948v1)\n- [LLMs in Argument Mining — Survey](https://arxiv.org/html/2506.16383v4)\n- [LLM Ethics Benchmark — Scientific Reports](https://www.nature.com/articles/s41598-025-18489-7)\n- [LLMs and Moral Value Assessment — ScienceDirect (2025)](https://www.sciencedirect.com/science/article/pii/S2451958825000247)\n- [Ethical Reasoning Depends on Prompt Language](https://aclanthology.org/2024.lrec-main.560/)\n- [LLM-based Argument Mining meets Argumentation Logics](https://arxiv.org/html/2603.02858v1)\n\n---\n\n## 6. Normative Argument Schemes\n\n### Walton's Normative Schemes\n\nWalton, Reed & Macagno (2008) catalogued 96 argumentation schemes. Several are specifically normative:\n\n#### Argument from Values\n\n> **Premise 1**: Value V is positive as judged by agent A.\n> **Premise 2**: If V is positive, it is a reason for A to commit to goal G.\n> **Conclusion**: V is a reason for A to commit to goal G.\n\n**Critical Questions:**\n1. Is V really a positive value?\n2. Is the claim that V is promoted by G true?\n3. Are there alternative ways to promote V that are better?\n4. Is the value V the most relevant in this situation?\n5. Are there other values that should override V here?\n\n#### Argument from Consequences (Positive)\n\n> **Premise**: If action A is brought about, good consequences will plausibly occur.\n> **Conclusion**: Therefore A should be brought about.\n\n**Critical Questions:**\n1. How likely are these consequences really?\n2. Are there negative side effects?\n3. Are there alternative actions with better consequences?\n\n#### Argument from Consequences (Negative)\n\n> **Premise**: If A is brought about, bad consequences will occur.\n> **Conclusion**: Therefore A should not be brought about.\n\n#### Practical Reasoning (Value-Based Variant)\n\n> **Major Premise**: I have a goal G.\n> **Minor Premise**: Carrying out action A is a means to realize G.\n> **Conclusion**: Therefore, I ought (practically speaking) to carry out A.\n\n**Critical Questions include the \"side effects question\"**: What consequences of bringing about A should also be taken into account?\n\n### Relevance for Deliberus\n\nThe argument scheme framework is **highly actionable** for Deliberus:\n\n1. **Template-based decomposition**: When a user makes a normative claim, prompt them (or use LLM-assist) to fill in the scheme template. \"What value does this serve? What consequences do you predict?\"\n2. **Critical questions as structured challenges**: Instead of freeform rebuttals, the platform could offer scheme-specific critical questions as guided response options\n3. **Scheme detection via LLM**: Recent work shows LLMs can classify arguments by scheme type using symbolic scaffolding — \"Is this an argument from values, consequences, authority, or analogy?\"\n4. **Argument quality scoring**: Arguments that explicitly state their value premises and address critical questions could score higher on argument quality metrics\n\n**Sources:**\n- [Argumentation Schemes (Walton, Reed, Macagno, 2008)](https://www.cambridge.org/core/books/argumentation-schemes/9AE7E4E6ABDE690565442B2BD516A8B6)\n- [Argumentation Schemes and Argument Mining](https://windsor.scholarsportal.info/omp/index.php/wsia/catalog/download/106/106/1086?inline=1)\n- [Classification System for Argumentation Schemes (2015)](https://journals.sagepub.com/doi/10.1080/19462166.2015.1123772)\n- [Argumentation Scheme — Wikipedia](https://en.wikipedia.org/wiki/Argumentation_scheme)\n\n---\n\n## 7. Deliberative Polling\n\n### James Fishkin's Method\n\nFishkin (Stanford Deliberative Democracy Lab) developed Deliberative Polling® in 1988. The process:\n\n1. **Random representative sample** of citizens\n2. **Balanced briefing materials** presenting multiple perspectives\n3. **Small group discussions** (12-15 people) with trained moderators\n4. **Expert Q&A panels** where participants interrogate experts\n5. **Pre/post polling** measuring opinion change\n\n### Evidence on Value Change\n\nKey findings from 150+ experiments in 50+ countries:\n\n- **~70% of participants change their minds** on at least some issues\n- **Respondents tend to move toward center** — effect lasts at least one year\n- **Shifts toward cosmopolitan, egalitarian, and collectivist** value orientations are common\n- **The strongest value-laden shifts occur on empirical statements** — when people learn new facts, their normative positions shift accordingly\n- **Information provision and value reflection both contribute** — the mechanisms are not purely cognitive but also involve perspective-taking and empathy\n\n### What Causes Genuine Value Change vs. Opinion Shift?\n\nThe research distinguishes:\n\n- **Information effects**: People update factual beliefs, which changes the empirical premises of their normative arguments. This is the dominant mechanism — most \"value change\" is actually \"updating beliefs about consequences\"\n- **Perspective-taking effects**: Hearing diverse viewpoints in small groups triggers empathy and consideration of moral foundations one doesn't typically use\n- **Deliberation quality effects**: The four parameters that matter: equal participation, respect for others' opinions, willingness to adopt a societal perspective, and reasoned justification of positions\n\n### Implications for Deliberus\n\nFishkin's work provides **the strongest empirical evidence** that structured deliberation genuinely changes minds, and not just through social pressure. Key design takeaways:\n\n1. **Balanced information is prerequisite** — the platform must surface counter-evidence and competing expert views, not just pro/con arguments\n2. **Small group deliberation is more effective than mass debate** — consider cohort-based deliberation features\n3. **The biggest shifts come from updating empirical beliefs** — reinforces the importance of separating factual and normative claims, so people can update facts without feeling their values are under attack\n4. **One year persistence** — these aren't transient effects\n\n**Sources:**\n- [What is Deliberative Polling® — Stanford](https://deliberation.stanford.edu/what-deliberative-pollingr)\n- [Deliberative Opinion Poll — Wikipedia](https://en.wikipedia.org/wiki/Deliberative_opinion_poll)\n- [James S. Fishkin — Stanford](https://deliberation.stanford.edu/people/james-fishkin)\n- [Four Parameters for Measuring Democratic Deliberation](https://www.cogitatiopress.com/politicsandgovernance/article/viewFile/1199/710)\n- [Measuring Deliberation — Harvard](https://ash.harvard.edu/wp-content/uploads/2024/02/baechtiger_0.pdf)\n\n---\n\n## 8. The \"Ethical Filter\" Concept\n\n### Edvin's 2011 Idea\n\nThe original concept: users answer ethical dilemmas, the system computes a moral similarity profile, and argument summaries are viewed through that \"lens\" — showing how arguments align with or challenge the user's own value framework.\n\n### Existing Implementations and Research\n\n#### YourMorals.org\n\nThe closest existing implementation: users take the Moral Foundations Questionnaire and receive their moral foundations profile. Over 517,590 participants. However, it is **purely diagnostic** — it tells you your profile but does not use it to filter or personalize content.\n\n#### Polis (pol.is) — Bridging Algorithm\n\nPolis is the most sophisticated implementation of a related concept:\n\n- Participants submit short opinion statements and vote Agree/Disagree/Pass on others' statements\n- **PCA + K-means clustering** produces a 2D opinion space, identifying opinion groups\n- The key innovation: **\"bridging statements\"** — ideas that gain agreement across all clusters\n- Used in governance decisions (Taiwan's Uber regulation, 2015) and by civic organizations worldwide\n- Open source: [compdemocracy.org](https://compdemocracy.org/polis/)\n\nPolis inverts the typical social media dynamic: instead of amplifying divisive content, it surfaces consensus. This is a \"value filter\" operating at the group level.\n\n#### Community Notes (Twitter/X) — Bridging-Based Scoring\n\nX's Community Notes uses a bridging-based matrix factorization algorithm:\n\n- Contributors are placed on an opinion spectrum based on rating history\n- A note is deemed helpful **only if** it receives positive ratings from contributors with opposing viewpoints\n- Notes that are rated helpful only by one side are suppressed\n- Reduced likes/reposts of misinformation by 25-34%\n\nThis is a **normative content filter** that privileges cross-partisan agreement over popularity.\n\n#### Ethical Profiling in AI (Emerging Research)\n\nRecent work on automated ethical profiling envisions interactive loops: (i) system proposes an ethical explanation, (ii) user confirms/modifies/rejects, (iii) profile updates. This enables both user agency and model refinement, reducing misaligned ethical personalization.\n\n### AllSides — Media Bias as Value Proxy\n\nAllSides rates media sources on a Left-to-Right spectrum and lets users rate their own bias. While not explicitly about ethical values, it demonstrates that users are willing to engage with self-assessment of their own perspective. Their methodology uses multi-partisan editorial reviews and blind bias surveys.\n\n### Synthesis: What an Ethical Filter Could Look Like in Deliberus\n\nCombining these approaches:\n\n1. **Onboarding**: MFQ-style assessment producing a moral foundations profile (Care, Fairness, Loyalty, Authority, Sanctity, Liberty scores)\n2. **Argument tagging**: LLM-assisted tagging of arguments with moral foundations and value premises they invoke\n3. **Personalized views**: \"Arguments that challenge your dominant foundation\" / \"Arguments that align with your values\" / \"Bridging arguments that satisfy multiple foundations\"\n4. **Group dynamics**: Polis-style clustering to find bridging positions across value groups\n5. **Transparency**: Users always see *that* filtering is happening and *why* — the filter is a tool, not a gate\n\n**Risk**: Filter bubbles. The ethical filter must be designed to **expand** perspectives, not reinforce them. Show the user where they are on the value landscape and what they're not seeing.\n\n**Sources:**\n- [Polis — Wikipedia](https://en.wikipedia.org/wiki/Pol.is)\n- [Polis Platform](https://compdemocracy.org/polis/)\n- [LLMs for Scalable Deliberation with Polis](https://arxiv.org/html/2306.11932)\n- [Community Notes Ranking Algorithm](https://communitynotes.x.com/guide/en/under-the-hood/ranking-notes)\n- [Consensus Stability of Community Notes](https://arxiv.org/html/2601.14002)\n- [Advancing Automated Ethical Profiling](https://arxiv.org/html/2510.00881v1)\n- [AllSides Bias Checker](https://www.allsides.com/bias-checker)\n- [YourMorals.org](https://yourmorals.org/)\n\n---\n\n## 9. Value-Sensitive Design\n\n### The VSD Framework (Friedman, Hendry)\n\nValue-Sensitive Design is a theoretically grounded approach to technology design that accounts for human values in a principled and comprehensive manner. Developed by Batya Friedman (University of Washington), VSD uses a tripartite methodology:\n\n1. **Conceptual investigations**: Identify stakeholders, articulate values, examine value tensions\n2. **Empirical investigations**: Study how users actually experience values in practice\n3. **Technical investigations**: Design systems that support identified values\n\nVSD emphasizes **values with ethical import** — the assumption is that early awareness of values can help shape technology development in ways conducive to their realization.\n\n### Application to a Deliberation Platform\n\nVSD is directly applicable to Deliberus. Key values to design for:\n\n| Value | Design Implication |\n|-------|-------------------|\n| **Fairness** | Equal voice for all participants; no algorithmic amplification of popular views |\n| **Accountability** | Traceable argument chains; visible reasoning |\n| **Autonomy** | Users control their own value profiles and filter settings |\n| **Transparency** | Platform's reasoning about moral foundations, argument quality, and content surfacing must be inspectable |\n| **Informed Consent** | Users understand how their moral profile is used |\n| **Democratic Process** | The platform serves deliberation, not persuasion |\n| **Value Pluralism** | The system must accommodate genuinely incompatible value systems without privileging any |\n\n### The Value Pluralism Challenge\n\nThe Stanford Encyclopedia of Philosophy defines value pluralism as the thesis that **conflict between ultimate values is a permanent feature of the human condition**, not a sign of flawed reasoning. This is the deepest design challenge for Deliberus:\n\n- **Monist assumption** (incorrect for Deliberus): There is one right answer, the platform helps find it\n- **Pluralist reality**: Multiple valid moral frameworks exist that genuinely conflict. The platform must hold them in tension, not resolve them\n\nResearch suggests that acknowledging value pluralism requires:\n- Moving from **optimization frameworks** to **negotiation and justice frameworks**\n- Designing processes where outcomes are **procedurally legitimate** even when substantive consensus is impossible\n- Making **intrapersonal** value conflicts visible too — people often disagree *within themselves*\n\n**Sources:**\n- [Value Sensitive Design — Wikipedia](https://en.wikipedia.org/wiki/Value_sensitive_design)\n- [VSD: Theory and Methods (Friedman, 2003)](https://research.cs.vt.edu/ns/cs5724papers/6.theoriesofuse.cwaandvsd.friedman.vsd.pdf)\n- [VSD MIT Press Book](https://mitpress.mit.edu/9780262039536/value-sensitive-design/)\n- [Value Pluralism — Stanford Encyclopedia of Philosophy](https://plato.stanford.edu/entries/value-pluralism/)\n- [Value Pluralism in Environmental Policy](https://besjournals.onlinelibrary.wiley.com/doi/full/10.1002/pan3.10324)\n\n---\n\n## 10. Practical Implications for Deliberus\n\n### The Core Design Decision: Claim Ontology\n\nBased on this research, normative claims need **fundamentally different treatment** from factual claims. The minimum viable distinction:\n\n```\nClaim\n├── FactualClaim        → Can be verified empirically\n│   ├── evidence_status: verified | disputed | unverified\n│   └── confidence: 0.0-1.0\n│\n├── NormativeClaim      → Expresses a value judgment or prescription\n│   ├── value_premises: [Value]          → What values does this assume?\n│   ├── empirical_sub_claims: [FactualClaim]  → What facts does it depend on?\n│   ├── moral_foundations: [Foundation]   → Which MFT foundations it invokes\n│   ├── argument_scheme: Scheme          → Walton classification\n│   └── agreement_axes:\n│       ├── agreement: -1.0 to 1.0      → Do you agree with this prescription?\n│       └── importance: 0.0 to 1.0       → How important is this issue?\n│\n└── HybridClaim         → Contains both factual and normative elements\n    ├── factual_component: FactualClaim\n    └── normative_component: NormativeClaim\n```\n\n### Agreement + Importance (Not True/False)\n\nFor normative claims, **truth value is category-inappropriate**. Instead, Deliberus should score on two axes:\n\n1. **Agreement** (-1 to +1): \"How much do you agree with this prescription?\" Captures the normative stance directly\n2. **Importance** (0 to 1): \"How important is this issue to you?\" Prevents vocal minorities from dominating and surfaces what people care about most\n\nThis is more informative than simple upvote/downvote because it distinguishes \"I disagree but recognize this matters\" from \"I don't care about this.\"\n\n### The Argument Decomposition Pipeline\n\nWhen a user submits a normative claim, Deliberus should:\n\n1. **Classify**: Is this factual, normative, or hybrid? (LLM-assisted, ~90%+ accuracy achievable)\n2. **Decompose**: Extract value premises and empirical sub-claims (Walton template + LLM)\n3. **Tag**: Identify moral foundations invoked (MoralBERT or prompted LLM)\n4. **Scheme-match**: Which argumentation scheme is being used? (Walton's 96 schemes, reduced to ~10-15 common normative ones)\n5. **Surface critical questions**: Present scheme-appropriate critical questions as structured challenge prompts\n\n### The Mercier-Sperber Insight\n\nMercier and Sperber's argumentative theory of reasoning provides the theoretical foundation for *why* deliberation works:\n\n- **Individual reasoning is biased** (confirmation bias is a *feature*, not a bug — it helps us produce persuasive arguments)\n- **Group reasoning corrects individual bias** through **epistemic vigilance** — others challenge flawed reasoning\n- **But this only works with diverse groups** — like-minded groups amplify bias (group polarization)\n\n**Design implication**: Deliberus must ensure normative deliberation involves **diverse** participants, not self-selected echo chambers. The Polis bridging approach and Community Notes cross-partisan requirement are proven mechanisms for this.\n\n### Ethical Filter Architecture (Candidate Design)\n\n```\nUser Onboarding\n├── MFQ-2 Assessment → Moral Foundations Profile\n├── Schwartz Values Survey → Value Priority Profile\n└── Optional: WVS-style Cultural Position\n\nArgument Display\n├── Default View: All arguments, weighted by quality + engagement\n├── Challenge View: Arguments invoking foundations you score LOW on\n├── Alignment View: Arguments resonating with YOUR value profile\n├── Bridging View: Arguments with cross-cluster agreement (Polis-style)\n└── Foundation View: Arguments grouped by moral foundation\n\nUser Controls\n├── Toggle filter on/off (always available)\n├── See filter explanation (\"shown because it challenges your Care foundation\")\n├── Adjust foundation weights manually\n└── Compare your profile to the community's profile distribution\n```\n\n### Open Design Questions\n\nThese questions arise from the research but cannot be answered by research alone — they require the kind of structured deliberation that Deliberus itself is meant to enable:\n\n1. **How formal should value elicitation be?** MFQ-2 is validated but 36 items may feel heavy. Could implicit profiling from argument behavior work? (Risk: opacity)\n2. **Should the platform take a meta-ethical stance?** Value pluralism says \"no right answer.\" But the platform must still decide which arguments get surfaced and how — those are normative choices. VSD says: make those choices transparent\n3. **How to handle bad-faith normative arguments?** Trolley-problem-style trolling, arguments made in bad faith, strategic gaming of the moral foundations tagger\n4. **Cultural calibration**: MFT was developed primarily from WEIRD (Western, Educated, Industrialized, Rich, Democratic) samples. Schwartz's universality claim is stronger but still contested. How does Deliberus handle genuinely different moral ontologies?\n5. **The internalization question**: Fredrik's vision includes \"internalization\" — individuals personally grappling with arguments, not just voting. How does the ethical filter support this without becoming a crutch that lets people avoid engaging with uncomfortable moral positions?\n6. **Feed algorithm implications**: Infinite scroll vs. organic redraw vs. cohort-based deliberation — each has different implications for how normative claims are encountered and processed\n\n### Recommended Reading Order for Deep Design Work\n\n1. **Walton, Reed & Macagno (2008)** — *Argumentation Schemes*. The practical framework for normative argument structure\n2. **Fishkin (2018)** — *Democracy When the People Are Thinking*. Empirical evidence for deliberation changing minds\n3. **Haidt (2012)** — *The Righteous Mind*. Why people disagree morally and how to bridge\n4. **Friedman & Hendry (2019)** — *Value Sensitive Design*. How to build values into technology\n5. **Mercier & Sperber (2017)** — *The Enigma of Reason*. Why group reasoning works and individual reasoning doesn't\n6. **Berlin (1958/2002)** — *Liberty* / *The Proper Study of Mankind*. The philosophical foundation for value pluralism\n\n### Connection to Existing Deliberus Docs\n\n- **Object model** (`docs/object-model.md`): Needs expansion for NormativeClaim entity type with value premises, moral foundation tags, and dual-axis scoring\n- **Technical direction** (`docs/technical-direction.md`): LLM pipeline for claim decomposition should include normative/factual classification as first pass\n- **Academic foundations** (`docs/academic-foundations.md`): Walton schemes and MFT literature complement the existing ASPIC+ and DeliData references\n- **Competitive landscape** (`docs/competitive-landscape.md`): No existing platform handles the fact/value boundary explicitly — this is a genuine differentiator for Deliberus\n- **Vision** (`docs/vision.md`): The \"rationality standards built into the software\" formulation aligns with making value premises explicit, not with computing moral answers\n"}