一句话：将系统安全工程的论证模型、模板与操作引入 AI 安全，并把模型卡重构为可操作的安全工件。
AI/ML 系统虽未必属于安全关键系统，仍可能造成严重伤害；现有 AI Safety 更接近模型验证，而非安全保证。本文主张将系统安全工程的论证模型、模板与操作引入 AI Safety，以更广泛考虑潜在伤害并构建显式、结构化的安全论证。作者将事实标准 Model Cards 重构为可操作安全工件，使安全声明与支撑证据显式化。通过分析 GPT-OSS、Claude 3、Gemma 3N 等先进大语言模型，验证了可从现有工件识别安全模板、用 justification diagrams 表达为论证模型，并将其操作化，在证据不再支持安全声明时向 AI 开发者提供即时反馈。
Artificial intelligence (AI) and Machine Learning (ML) models are increasingly deployed in systems that, while not safety-critical per se, may still cause harm (e.g., a recent tragedy involving the death of a sixteen-year-old teenager and ChatGPT). Recent incidents have highlighted how unsafe such systems can be in practice. While safety engineering is a mature discipline for safety-critical systems, AI Safety remains closer to model validation than to safety assurance. In this paper, we propose integrating engineering practices from system safety, including argumentation models, templates, and operations, into AI Safety. This alignment enables a broader consideration of potential harm and supports the construction of explicit, structured safety arguments. We validate our approach by reframing Model Cards, a de facto industry standard, as actionable safety artifacts that make safety claims and supporting evidence explicit. We analyze state-of-the-art large language models (GPT–OSS, Claude 3, Gemma 3N) to show that it is possible to (i) identify safety templates in existing artifacts, (ii) express them as argumentation models using justification diagrams, and (iii) operationalize these models to provide immediate feedback to AI developers when evidence no longer supports safety claims.
@inproceedings{paperbot12043,
title = {Model Cards for Responsible AI: Stop Carding, Start Modelling},
author = {Kalvin Thuan-Phong Khuu and Nicolas Lacroix and Baptiste Lacroix and Richard Paige and Mireille Blay-Fornarino and Sébastien Mosser},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3786582.3786804}
}