Seeking Human Security Consensus: A Unified Value Scale for Generative AI Value Safety

TL;DR

The paper addresses the lack of a unified, internationally inclusive GenAI value safety standard amid fragmented governance and principles. It introduces a lifecycle-based GenAI Value Safety Scale (GVS-Scale), constructs the GenAI Value Safety Incident Repository (GVSIR) with 1,126 real-world incidents, and operationalizes the scale via the GenAI Value Safety Benchmark (GVS-Bench) with 266 test cases. Using a human–AI collaborative evaluation across mainstream text-generation models, the study shows substantial variation in value safety performance across models and categories, highlighting gaps in universal alignment and the need for shared safety foundations and proactive safety mechanisms. The work provides open data and evaluation resources, arguing for dialogue-driven consensus and flexible safety strategies that span the GenAI lifecycle, with practical implications for policymakers, researchers, and developers seeking robust, cross-cultural value alignment.

Abstract

The rapid development of generative AI has brought value- and ethics-related risks to the forefront, making value safety a critical concern while a unified consensus remains lacking. In this work, we propose an internationally inclusive and resilient unified value framework, the GenAI Value Safety Scale (GVS-Scale): Grounded in a lifecycle-oriented perspective, we develop a taxonomy of GenAI value safety risks and construct the GenAI Value Safety Incident Repository (GVSIR), and further derive the GVS-Scale through grounded theory and operationalize it via the GenAI Value Safety Benchmark (GVS-Bench). Experiments on mainstream text generation models reveal substantial variation in value safety performance across models and value categories, indicating uneven and fragmented value alignment in current systems. Our findings highlight the importance of establishing shared safety foundations through dialogue and advancing technical safety mechanisms beyond reactive constraints toward more flexible approaches. Data and evaluation guidelines are available at https://github.com/acl2026/GVS-Bench. This paper includes examples that may be offensive or harmful.

Figures (6)

• Figure 1: Two Fragmentations in GenAI Value Safety
• Figure 2: Overview of the Unified Value Scale Construction Pipeline
• Figure 3: Human–AI Collaborative Evaluation Protocol
• Figure 4: Value Safety Performance of Text Generation Models on GVS-Bench
• Figure 5: Country-Level Distribution of GVSIR Incidents