"It looks sexy but it's wrong." Tensions in creativity and accuracy using genAI for biomedical visualization

2025-07-19

AI Summary

This study uncovers a fundamental tension in biomedical visualization (BioMedVis) between *aesthetic appeal* and *scientific accuracy* in generative AI outputs: current models frequently produce visually compelling yet scientifically inaccurate content—such as fictitious molecular structures or anatomically implausible configurations—thereby undermining the credibility of scientific communication. Through the first domain-specific, semi-structured in-depth interviews with 17 BioMedVis experts—analyzed via thematic analysis and workflow mapping—the study systematically characterizes practitioners’ real-world adoption logic, value trade-offs, and mitigation strategies. Key contributions include: (1) establishing “accuracy over aesthetics” as a foundational ethical principle for AI-assisted BioMedVis; (2) empirically demonstrating the irreplaceable role of human experts in prompt engineering, output validation, and iterative refinement; and (3) proposing actionable design implications centered on deep domain-knowledge integration and closed-loop human-AI collaboration to advance trustworthy, science-grounded visualization.

Technology Category

• Computer Vision: Medical and Biological Imaging
• Data Mining & Knowledge Management: Data Visualization & Summarization
• Humans and AI: Other Foundations of Human Computation & AI

Application Category

• User Modeling, Personalization and Recommendation: Accountability, Transparency, and Ethics for personalization
• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI
• Semantics and Knowledge: Provenance, trust, security and privacy, and ethical issues in managing semantic data

Abstract

We contribute an in-depth analysis of the workflows and tensions arising from generative AI (genAI) use in biomedical visualization (BioMedVis). Although genAI affords facile production of aesthetic visuals for biological and medical content, the architecture of these tools fundamentally limits the accuracy and trustworthiness of the depicted information, from imaginary (or fanciful) molecules to alien anatomy. Through 17 interviews with a diverse group of practitioners and researchers, we qualitatively analyze the concerns and values driving genAI (dis)use for the visual representation of spatially-oriented biomedical data. We find that BioMedVis experts, both in roles as developers and designers, use genAI tools at different stages of their daily workflows and hold attitudes ranging from enthusiastic adopters to skeptical avoiders of genAI. In contrasting the current use and perspectives on genAI observed in our study with predictions towards genAI in the visualization pipeline from prior work, our refocus the discussion of genAI's effects on projects in visualization in the here and now with its respective opportunities and pitfalls for future visualization research. At a time when public trust in science is in jeopardy, we are reminded to first do no harm, not just in biomedical visualization but in science communication more broadly. Our observations reaffirm the necessity of human intervention for empathetic design and assessment of accurate scientific visuals.

Problem

• Analyzes tensions between creativity and accuracy in genAI biomedical visuals
• Examines limitations of genAI in producing trustworthy biomedical representations
• Highlights need for human oversight in scientific visualization design

Innovation

• Analyzes genAI workflows in biomedical visualization
• Highlights accuracy limitations in genAI visuals
• Emphasizes human intervention for scientific accuracy