Societal AI Taxonomy: Frameworks & Governance

• Societal AI Taxonomy is a framework categorizing AI risks, opportunities, and normative impacts across social, political, and technical dimensions.
• It employs multiple axes—such as social norms, systemic harms, and regulatory structures—to align technical properties with sociocultural contexts.
• The taxonomy underpins audit, governance, and regulation by providing actionable insights for comparative benchmarking and risk assessment.

A Societal AI Taxonomy is a structured framework for categorizing the risks, opportunities, and normative dimensions by which artificial intelligence systems interact with, reshape, or threaten society. Such taxonomies provide foundations for evaluation, governance, regulation, auditing, and technical control of AI’s societal-scale effects. Recent research efforts have produced multiple, largely complementary taxonomies, which can be decomposed along axes of social norms, systemic harms, political risks, regulatory structure, organizational integration of AI, and multi-scale ethical theory. These taxonomies extend standard risk categories by aligning technical properties of models and deployments with sociocultural context, power structures, and institutional oversight.

1. Social Norms: Contextual Taxonomies for Human-AI Interaction

The “Cultural Compass” taxonomy formalizes the evaluation of AI models' adherence to sociocultural norms across three pillars: Contextual Frame, Norm Specification, and Mechanism of Enforcement (Cheng et al., 12 Jan 2026). The Contextual Frame encodes (a) the cultural context (e.g., country), (b) the situational context (general vs. situation-specific norm), and (c) the interactional context distinguishing norms AI should recognize between humans (H-H) from those it must itself uphold in human–AI (H-AI) dialogues. Norm Specification annotates the domain of conduct (behavior, belief, language) and the mode of articulation (descriptive or prescriptive). Mechanism of Enforcement distinguishes formal (legal) from informal (social sanction) adherence.

Operationalization proceeds via a four-step evaluation pipeline.

1. Diverse prompt generation (user intent × scenario × culture mention)
2. LLM model response collection
3. LLM-judged surfacing of relevant norms
4. Automated violation detection

Empirically, violation rates vary by model, interactional context, and explicit cultural framing; e.g., Gemini-2 exhibits the highest violation rates across all contexts, whereas Claude Sonnet shows the lowest. Notably, explicit country mention lowers violation rates for H-H norms but is much less effective in reducing violations of H-AI norms. Functional content generation tasks elicit disproportionately high violation rates, sharply so in H-AI, situation-specific contexts.

This taxonomy enables nuanced, metadata-driven evaluation and supports the differentiation of knowledge-oriented vs. behavioral compliance for AI systems (Cheng et al., 12 Jan 2026).

2. Systemic and Societal Harms Taxonomies

Large-scale sociotechnical harms are categorized into granular themes and subthemes to provide actionable targets for risk reduction and governance. One representative taxonomy includes:

• Representational Harms: Outputs reflecting or amplifying unjust social hierarchies—stereotyping, demeaning, erasing, alienating, forced categorization, essentialism.
• Allocative Harms: Denial or distortion of access to resources and opportunities, economic deprivation.
• Quality-of-Service Harms: Systemic disparaties in performance and service for particular groups; increased labor for marginalized users; alienation from digital services.
• Interpersonal Harms: Distortion or harm to human relationships, loss of agency, technology-facilitated violence.
• Social System/Societal Harms: Disruption or erosion of social institutions, trust, or large-scale public goods (Shelby et al., 2022).

Operationalizing these themes supports longitudinal policy design, participatory harm elicitation, and the development of intersectional benchmarking protocols.

The HARM66+ taxonomy formalizes 66+ harm types across exo-human (environment, digital, infrastructure) and endo-human (physical, cognitive, social, political, financial) domains, with each harm annotated by victim entity taxonomy, reversibility, and durance, and mapped to 11 ethical theories. Severity functions combine these attributes and ethical context to enable structured, theory-aware risk assessments (Khan et al., 23 Jan 2026).

3. Political, Regulatory, and Systemic Risk Axes

Societal AI risk taxonomies now extend to political risk, regulatory clustering, and systemic risk propagation. Examples include:

• Political Risk Taxonomy: Groups risks into Geopolitical Pressures (AI arms race, AWS, technopolarity), Malicious Usage (misinformation, defamation, cyber attacks), Environmental/Social/Ethical (carbon, economic, bias), and Privacy/Trust Violations (surveillance, censorship, trust erosion) (Arda, 2024).
• Six-Dimensional Regulatory Schema: Regulation Layer (technology/application/hybrid), Coverage Scope, Timing of Intervention, Legal Maturity, Enforcement Mechanisms, and Stakeholder Participation, with empirically mapped coordinates for major regulatory regimes (EU, US, China, etc.). This multi-axis schema reveals distinct global clusters: “comprehensive,” “transitional,” and “minimalist” regulation (Alanoca et al., 19 May 2025).
• Systemic Risk Taxonomy: Thirteen categories reflect societal-scale threats: Control, Democracy, Discrimination, Economy, Environment, Fundamental Rights, Governance, Harms to Non-Humans, Information, Irreversible Change, Power, Security, and Warfare. These are driven by three meta-sources—knowledge gaps, harm-recognition challenges, and unpredictable AI trajectories—amplifying risk propagation across the value chain (Uuk et al., 2024).

4. Entity, Integration, and Governance Typologies

Taxonomies for classifying AI, robots, and agents for governance adopt integration-centered frameworks. The CPST (Cyber, Physical, Social, Thinking) integration space theory encodes an agent’s profile as a 4-vector across computational autonomy, physical/embodiment, relational (social), and cognitive (goal complexity, self-modification) axes (Ning et al., 7 Apr 2026). A three-tier governance scale:

• Tier 1: Confined Actors (low integration, no social embedding)—governed via enhanced product liability.
• Tier 2: Socially-Aware Interactors (multi-axis, especially social, integration)—subject to relational governance and operational rights.
• Tier 3: CPST-Integrated Agents (full integration across all axes)—with qualified legal personhood, rights, and bespoke accountability.

Transitions and boundary adjudications are handled through periodic re-assessments and explicit notification/appeal processes. This taxonomy supports proportional, emergent policy design, tailored to an agent’s demonstrable integration and societal embedment (Ning et al., 7 Apr 2026).

5. Causal/Domain Hybrid Risk Repositories and Modularity

Meta-reviews extracting risk statements from hundreds of taxonomies have produced hybrid frameworks that combine causal (who/why/when) lenses with mid-level domain/hazard categories (Slattery et al., 2024):

• Causal taxonomy:
  • Entity: {Human, AI, Other}
  • Intentionality: {Intentional, Unintentional, Other}
  • Timing: {Pre-, Post-deployment, Other}
• Domain taxonomy:
  • Discrimination & Toxicity
  • Privacy & Security
  • Misinformation
  • Malicious Actors & Misuse
  • Human–Computer Interaction
  • Socioeconomic & Environmental
  • AI System Safety, Failures, & Limitations (23 subdomains)

Prevalence metrics, cross-tabulation, and modular risk registration enable dynamic extension as risks and architectures evolve. These repositories (e.g., airisk.mit.edu) are adaptable for audit, regulation, and technical policy prioritization (Slattery et al., 2024).

6. Normative, Ethical, and Governance-First Foundations

Several taxonomies operate at the level of macroethics: responsible AI must be evaluated using a suite of normative principles—deontology, egalitarianism, utility, virtue, care, rights, contract, justice, natural law, environmental, pragmatist, existential. Each societal or technical harm can then be scored by alignment or violation under each theory and mapped to practical controls (Woodgate et al., 2022, Khan et al., 23 Jan 2026). Democracy-oriented dual taxonomies explicitly link societal harms (autonomy erosion, unfairness, power asymmetry, authoritarianism, discourse breakdown, principle undermining, trust loss) to technical and procedural mitigations (agency, robustness/safety, privacy, diversity/fairness, transparency, societal well-being, accountability) (Mentxaka et al., 19 May 2025).

Alternatively, “governance-first” approaches emphasize that technical taxonomies risk reducing incommensurable, context-dependent harms to quantifiable metrics, and should instead serve as input for inclusive, adaptive, participatory governance structures (Berman et al., 2024).

7. Synthesis: Integration, Comparative Utility, and Application

Societal AI taxonomies are not mutually exclusive; they function as layered, interoperable scaffolds supporting technical audit, regulatory calibration, policy coordination, and participatory governance. Key differences among frameworks reflect:

• Classification granularity (from high-level systemic risk to fine-grained harm type)
• Target application (e.g., conversational norms, system-level governance, legislative alignment, human–agent interaction)
• Ethical and political commitments (procedural, consequentialist, participatory)

Practical use cases include regulatory gap analysis (e.g., EU AI Act compliance gaps for autonomous weapon systems and systemic risk cutoffs (Arda, 2024)), governance calibration (regulatory sandboxes, international standards (Ning et al., 7 Apr 2026)), and the design/implementation of AI risk repositories for socio-technical audit and red-teaming (Slattery et al., 2024).

The systematic categorization of societal AI risks, harms, and opportunities provides researchers, technologists, regulators, and civil society with a common, operational vocabulary—enabling comparative benchmarking, targeted intervention, and the iterative refinement of both AI systems and their governance regimes.