Adapting Vision-Language Models for Human-Readable XAI in Industrial Object Detection

AI explanations need to be designed for actual users, not just experts: The paper shows that standard AI explanation tools often fail to communicate clearly to blue-collar workers and other non-technical users in manufacturing settings, highlighting that responsible AI deployment requires tailoring explanations to the people who actually use the systems.

Off-the-shelf AI models struggle with real-world context: The researchers found that popular vision-language models like GPT-4o-mini "often fall short in delivering clear, context-relevant explanations" for industrial quality control tasks, demonstrating that general-purpose AI tools need customization to work responsibly in specific industries.

Fine-tuning improves trustworthiness and usability: By adapting a vision-language model specifically for manufacturing contexts, the team achieved measurable improvements in "explanation clarity, instruction adherence, image groundedness, and contextual awareness," suggesting that investment in customization can make AI systems more trustworthy and accessible to diverse workers.

Accessibility is a key component of responsible AI: The paper emphasizes that explainability alone isn't enough—explanations must be "intuitive and applicable" to the people relying on AI systems for critical decisions like quality control, making user accessibility a core responsibility rather than an afterthought.

Manufacturing workers deserve transparent AI tools: By focusing on making AI decisions understandable to frontline workers, this work recognizes that responsible AI integration means respecting workers' need to understand and potentially question the systems affecting their work.