Ethics Testing: A New Method to Identify Generative AI Harms
A new arXiv preprint (2604.22089) introduces 'ethics testing,' a systematic approach to identify software harms in content generated by Generative Artificial Intelligence (GAI) systems like ChatGPT, which rely on Large Language Models (LLMs). The paper argues that while tools such as ChatGPT have gained popularity for automatically generating code and images, misuse of this content can lead to serious consequences due to potential harms. Despite the importance of quality assurance, no existing methodology systematically generates tests for such harms. Ethics testing differs from fairness testing, which focuses on discrimination, by aiming to detect a broader range of software harms. The approach is proposed as a proactive measure to identify issues before deployment.
Key facts
- arXiv preprint 2604.22089 introduces ethics testing for GAI systems.
- Ethics testing aims to systematically generate tests for identifying software harms.
- Current methodologies like fairness testing focus on discrimination, not general harms.
- GAI systems like ChatGPT use LLMs to generate content such as code and images.
- Misuse of automatically generated content can have serious consequences.
- No existing approach systematically tests for harms in GAI-generated content.
- The paper proposes proactive identification of harms before deployment.
- Ethics testing is distinct from existing testing methodologies.
Entities
Institutions
- arXiv