"Just a strange pic": Evaluating 'safety' in GenAI Image safety annotation tasks from diverse annotators' perspectives

TL;DR

This study explores how diverse annotators assess safety in AI-generated images, revealing that moral, emotional, and contextual factors heavily influence judgments beyond predefined safety categories.

Contribution

It uncovers the complex reasoning behind safety judgments, emphasizing the influence of guidelines and subjective perceptions, and suggests improvements for evaluation frameworks.

Findings

Annotators invoke moral, emotional, and contextual reasoning.

Annotation guidelines shape harm interpretation and judgment.

Factors like image quality and prompt mismatch influence safety assessments.

Abstract

Understanding what constitutes safety in AI-generated content is complex. While developers often rely on predefined taxonomies, real-world safety judgments also involve personal, social, and cultural perceptions of harm. This paper examines how annotators evaluate the safety of AI-generated images, focusing on the qualitative reasoning behind their judgments. Analyzing 5,372 open-ended comments, we find that annotators consistently invoke moral, emotional, and contextual reasoning that extends beyond structured safety categories. Many reflect on potential harm to others more than to themselves, grounding their judgments in lived experience, collective risk, and sociocultural awareness. Beyond individual perceptions, we also find that the structure of the task itself -- including annotation guidelines -- shapes how annotators interpret and express harm. Guidelines influence not only which images are flagged, but also the moral judgment behind the justifications. Annotators frequently cite factors such as image quality, visual distortion, and mismatches between prompt and output as contributing to perceived harm dimensions, which are often overlooked in standard evaluation frameworks. Our findings reveal that existing safety pipelines miss critical forms of reasoning that annotators bring to the task. We argue for evaluation designs that scaffold moral reflection, differentiate types of harm, and make space for subjective, context-sensitive interpretations of AI-generated content.