Conference Agenda

The proliferation of Text-to-Image Generative AI (GenAI) shifts design from manual representation to linguistic specification. However, mainstream models trained on low-context Western data conflict with high-context Eastern aesthetics, causing "Aesthetic Aphasia." This study proposes the "Double Semantic Gap" model to analyze this cultural friction. Integrating Hall’s cultural context and Kalyuga’s expertise reversal theories, we investigate cognitive costs in AI collaboration. A pilot study on mood board creation reveals an expertise reversal effect: experts suffer higher cognitive load due to schema conflicts, while novices benefit from AI as semantic scaffolding. This research identifies mechanisms of cultural friction and advocates for localized prompting guidelines to preserve cultural diversity.

Artificial intelligence is increasingly discussed as a transformative technology in product development, yet its industrial adoption remains uneven. This study examines how academic and industrial experts perceive its applications, limitations and needs for action. Based on a qualitative survey of 26 participants, responses were systematically coded and clustered. The results reveal a partial divergence in perceived application areas. While both groups highlight data analysis, decision support and design automation, academia places stronger emphasis on simulation and modelling, whereas industry prioritizes knowledge management, routine automation and quality assurance. Despite these differences, both groups converge on key barriers, particularly data availability and quality, user trust, hallucinations and integration into existing toolchains. Several industrial participants report no active usage of artificial intelligence, indicating an implementation gap between research and practice. Overall, artificial intelligence in product development is currently positioned primarily as an efficiency-enhancing support tool rather than an autonomous design agent.

The influence of ChatGPT support on sequential reasoning in synchronous collaborative CAD activity was examined. Twenty-two student pairs redesigned a crankshaft for forging with or without ChatGPT support. Verbal protocols were coded and analysed using first-order transition probabilities between the Problem and Solution space and Analysis, Synthesis and Evaluation (ASE) design operations. With ChatGPT support, transitions from Analysis to Synthesis and from Evaluation to Synthesis were increased, within Analysis and returns to Analysis were reduced, fewer returns to Problem were observed, and coordination around the ChatGPT was indicated.

Parametric CAD models encode engineering design knowledge through feature-based representations, yet generative AI approaches to design automation treat these models as geometric shapes, discarding functional semantics and manufacturing constraints. Text-to-CAD methods rely on vision-language models to generate descriptions from renders, producing annotations that capture visual appearance but lack engineering content, limiting their ability to generate functionally valid outputs. We present a methodology for capturing machining feature knowledge in parametric CAD representations. Our approach enriches models with structured metadata linking geometric features to functional purposes and manufacturing requirements while maintaining traceability to design operations. We validate this on 1,770 shaft components with dual annotations—VLM-only baseline versus feature-augmented. Training generative models on feature-augmented data yields substantial improvements in geometric accuracy and functional correctness, demonstrating that structured engineering metadata enables better capture of design intent. We release our code for reproducibility.