Recent developments in artificial intelligence have focused mainly on language models capable of interpreting and generating text. However, AI’s capabilities steadily expand beyond textual applications into sectors such as manufacturing and the service industry. These domains, which often involve manual labour and physical processes, may not initially appear to benefit significantly from AI integration. Yet, a new interdisciplinary study from the University of Notre Dame reveals that AI tools can enhance product quality and worker safety in these environments, opening the door to substantial improvements in industrial practices.
The study, published in the journal Information Fusion, centres on a class of AI systems known as multimodal large language models (MLLMs). Unlike traditional AI models that operate within a single input type—usually text—MLLMs can process and integrate information from multiple sources, such as images, text, and data tables. This capacity enables more nuanced reasoning and decision-making. While most academic investigations into AI’s impact on the workplace have focused on white-collar occupations, the Notre Dame research team aimed to evaluate AI’s potential within industrial and production-based settings, where its adoption has thus far been limited.
To conduct their research, the team collaborated with vocational institutions and trade professionals across Northern Indiana, a region noted for its high concentration of manufacturing jobs. Partner organisations included the Elkhart Area Career Centre, Plymouth High School, Career Academy South Bend, Plumbers & Pipefitters Local Union 172, and Ivy Tech Community College. These partnerships, facilitated through Notre Dame’s iNDustry Labs, allowed researchers to collect real-world weld imagery to assess AI model performance. iNDustry Labs has worked with over 80 regional businesses on more than 200 applied projects, making it a valuable conduit for integrating academic research with industrial needs.
The study specifically focused on the welding practices found in the recreational vehicle, marine, aeronautical, and agricultural sectors. Researchers tasked MLLMs with evaluating images of welds and determining their suitability for various manufacturing applications. While the AI models demonstrated a promising ability to assess weld quality using high-resolution, curated images sourced online, their performance declined when applied to images captured in real-world settings. This discrepancy highlights a significant challenge: models trained on idealised data do not generalise well to messy, unstructured environments that typify most manufacturing floors.
According to Nitesh Chawla, the Frank M. Freimann Professor of Computer Science and Engineering at Notre Dame and founding director of the Lucy Family Institute for Data and Society, this finding emphasises the need to retrain models using actual industrial data and more sophisticated techniques. “This discrepancy underscores the need to incorporate real-world welding data when training these AI models, and to use more advanced knowledge distillation strategies when interacting with AI,” he explained. Such approaches could allow AI systems to assess weld integrity more accurately, thus improving both product performance and occupational safety.
Interestingly, the study also found that the complexity or scale of an AI model did not necessarily correlate with better outcomes. In several cases, performance improved significantly when researchers used context-specific prompts rather than larger or more computationally intensive models. This insight challenges prevailing assumptions in AI development and suggests that customisation and prompt engineering may be more effective than merely scaling up existing systems. As such, the authors recommend that future research focus on equipping AI models with stronger domain-specific reasoning abilities and adaptability.
The implications of this study extend far beyond welding. As industries move towards greater AI adoption, the challenge will be to balance the power of general-purpose models with the practicality and precision of fine-tuned tools tailored to specific environments. Yong Suk Lee, associate professor at Notre Dame’s Keough School of Global Affairs and chair of technology ethics at the Institute for Ethics and the Common Good, stressed the importance of explainability in AI. “As AI adoption in industrial contexts grows, practitioners will need to balance the trade-offs between using complex, expensive general-purpose models and opting for fine-tuned models that better meet industry needs,” he said. Integrating transparent and accountable AI systems will build trust and ensure these technologies benefit the workforce and society.
More information: Grigorii Khvatskii et al, Do multimodal large language models understand welding? Information Fusion. DOI: 10.1016/j.inffus.2025.103121
Journal information: Information Fusion Provided by University of Notre Dame