Large language models for automated consent form generation: A design and feasibility pilot.

602 Background: Informed consent forms (ICF) are essential decision-making aids for clinical trial participants. ICFs must be written in plain language while preserving legal and scientific accuracy yet can take weeks to develop, even with templates. Study teams face significant challenges including the lack of simple alternatives for complex terminology, tendency to overestimate participants' health literacy, and inherent difficulty achieving 6th grade reading levels. These obstacles often result in ineffective communication with excessive jargon and complexity. When properly engineered, large language models (LLMs) excel at transforming complex information into accessible explanations while maintaining accuracy. This feasibility pilot aimed to develop and evaluate an LLM-based system for generating first-draft ICFs from research protocols to reduce development time while maintaining appropriate reading levels and regulatory compliance. Methods: We utilized Amazon Bedrock with Anthropic's Claude 3.7 for systematic prompt engineering. From the research protocols we extracted key clinical trial information step-by-step and structured the information into simplified sentences formatted as JavaScript Object Notation (JSON). We inserted the outputs into a predetermined template using placeholders. The system incorporated conditional logic for study-type-specific content requirements. Principal investigators and IRB leaders reviewed drafts and provided feedback. Results: Through iterative prompt optimization and one-to-two shot examples we achieved 8 th grade reading level outputs from the LLM. Reviewer feedback was consistently positive regarding content accuracy and readability. However, human oversight remained essential for content validation and formatting. A significant challenge emerged: mandatory verbatim language sections comprised approximately half of the consent forms and substantially elevated reading levels, with legal sections showing highest complexity. We provided simplified, concise alternatives to IRB and legal teams for review. Conclusions: This LLM implementation represents a transformative innovation for study teams and IRB through dramatic improvement in efficiency by reducing ICF development time and providing high-quality initial drafts for further human refinement. This approach addresses a critical pain point by ensuring consistency between research protocols and ICFs, eliminating common discrepancies that frequently delay IRB approval. However, achieving true reading level targets requires institutional commitment to revising mandatory language sections. This technology enables teams to focus on higher-level content review, potentially transforming clinical trial recruitment and participant trust.