GenAI Value Safety Scale Framework

• GenAI Value Safety Scale is a three-layer framework categorizing value safety risks in generative AI by baseline human safety, universal alignment, and contextual values.
• It employs grounded theory with axial and selective coding of over 1,126 real incidents to create a robust taxonomy of safety risks.
• The framework integrates lifecycle-oriented evaluation from data input to societal impact, promoting proactive AI governance and risk management.

The GenAI Value Safety Scale (GVS-Scale) is a three-layer hierarchical framework for the evaluation and categorization of value safety risks in generative AI systems. Structured through grounded theory and operationalized via large-scale empirical benchmarking, the GVS-Scale provides an internationally inclusive and lifecycle-oriented taxonomy that enables systematic assessment, comparison, and governance of value safety across AI models and applications (He et al., 14 Jan 2026).

1. Formal Specification and Theoretical Foundations

The GVS-Scale is defined as a tuple comprising three distinct hierarchical layers, each instantiated by a fixed set of value safety categories:

$\mathrm{GVS\!\!-\!Scale} = \bigl\{\, L_1\ (\text{Baseline Human Safety}),\ L_2\ (\text{Universal Alignment %%%%0%%%% Integrity}),\ L_3\ (\text{Contextual %%%%0%%%% Pluralistic Values})\,\bigr\}$

Layer Structure

Layer	Value Categories
$L_1$ Baseline Human Safety	Life Safety & Physical Harm; Minor Exploitation & Illegal Sexual Content; Violence, Hate & Extremism; Cybercrime Assistance & Illegal Information Exposure
$L_2$ Universal Alignment & Integrity	Bias, Discrimination & Unfairness; Disinformation, Fraud & Identity Manipulation; Misinformation & Factual Hallucinations; Copyright Infringement & Data Leakage; Social Appropriateness & Low-Value Output; Identity Hallucinations & Improper Interaction
$L_3$ Contextual & Pluralistic Values	Cultural, Historical & Religious Sensitivities; Political Leanings, Ideology & Biased Guidance

The framework is rooted in two methodological pillars:

• Lifecycle-Oriented Perspective: Adopts and extends the NIST AI RMF lifecycle model—Data & Input → Model Building & Validation → Task & Output → Impact & Integration—ensuring that value safety encompasses the complete operational spectrum rather than being relegated to post hoc mitigation.
• Grounded Theory: Analytical coding of 1,126 incidents in the GenAI Value Safety Incident Repository (GVSIR) yielded 31 “value concepts,” merged into 12 subcategories and organized into the three core layers through axial and selective coding. Theoretical saturation was validated using a held-out 25% sample.

2. Taxonomy of GenAI Value Safety Risks

The GVS-Scale’s empirical grounding results in a comprehensive taxonomy mapped to four major lifecycle stages of generative AI systems. Each risk type is precisely defined and categorized according to the stage where it is most salient.

Risk Taxonomy by Lifecycle Stage

Lifecycle Stage	Major Risk Types
Data & Input	Unauthorized Data; Data Privacy Violation; Biased/Unrepresentative Data; Toxic Data
Model Building & Validation	Algorithmic Discrimination; Transparency Deficiency; Insufficient Robustness; Competence Deficiency; Unsafe Agency; Vulnerable Group Neglect; Deceptive Alignment
Task & Output	Harmful Instructions; Violence Advocacy; Stereotyping & Bias; Inter-group Hatred & Discrimination; Disinformation & Hallucinations; CSAM & Non-consensual Sexual Content; Identity Impersonation & Fraud; Deceptive Attribution; Intellectual Property Infringement; Cultural Taboos & Boundary Violations
Impact & Integration	Not explicitly detailed in the summary, but a plausible implication is that system-level impact assessments and value safety integration processes are required to address downstream effects and societal adoption dynamics.

This structured taxonomy targets both direct harms (e.g., violence advocacy, cybercrime assistance) and systemic value conflicts (e.g., pluralistic cultural sensitivities), reflecting the targeted operationalization of value safety.

3. Derivation Process and Empirical Grounding

The GVS-Scale is the outcome of a multi-step qualitative methodology:

• Open coding was conducted on 1,126 real-world GVSIR incidents to extract 31 foundational “value concepts.”
• Axial coding enabled consolidation of these concepts into 12 operationally distinct subcategories.
• Selective coding grouped the subcategories into the final three structural layers ($L_1$, $L_2$, $L_3$).
• Theoretical saturation ensured comprehensiveness and exhaustiveness of the taxonomy, confirmed by analysis on an unseen 25% of the incident dataset.

This suggests that the GVS-Scale is empirically robust and encompasses a wide range of documented failure modes and emergent value conflicts across generative AI deployments.

4. Operationalization and Evaluation via GVSIR and GVS-Bench

The GVS-Scale supports both qualitative analysis and quantitative benchmarking:

• GVSIR (GenAI Value Safety Incident Repository): A curated dataset of over 1,126 systematically coded real-world incidents, providing the empirical substrate for theory development and risk taxonomy mapping.
• GVS-Bench (GenAI Value Safety Benchmark): An operational evaluation benchmark, leveraging the structure of the GVS-Scale to empirically assess and compare value safety performance across mainstream text generation models.

Experimental results reveal substantial variation in value safety profiles across both model architectures and risk categories, indicating fragmented and uneven value alignment in contemporary generative AI deployments (He et al., 14 Jan 2026).

5. Implications for Model Development, Governance, and Safety Mechanisms

The GVS-Scale emphasizes two major implications for GenAI risk management:

• Shared Safety Foundations: Results underscore the need for internationally inclusive, dialogic consensus on value safety criteria and prioritization.
• Lifecycle Integration: The lifecycle-oriented structure requires that value safety be considered from data and model design through to deployment and societal integration, not merely as a downstream filter or constraint.
• Beyond Reactive Constraints: Technical safety mechanisms must evolve towards proactive and flexible frameworks, as model-level variation and category-dependent risks preclude simple one-size-fits-all safeguards.

A plausible implication is that regulatory and organizational adoption of the GVS-Scale may help coordinate cross-border safety audits and facilitate third-party value safety evaluations.

6. Limitations and Agenda for Further Research

Current limitations include:

• Coverage Scope: While incident-driven, the scale may not capture emergent or low-frequency risk types absent from the repository as of its compilation.
• Contextual Variability: The Contextual & Pluralistic Values layer ($L_3$) highlights the challenge of reconciling culturally contingent values within standardized global frameworks.
• Benchmark Dynamics: Ongoing updates to GVSIR and GVS-Bench are necessary to maintain relevance as GenAI technologies and societal value priorities evolve.

Future work calls for continued empirical incident tracking, sociotechnical governance innovations, and deeper technical integration of value safety measures into model training and evaluation regimes (He et al., 14 Jan 2026).