Prompt Framing Modulates Safety in Shoulder and Elbow Red-Flag Vignettes: A Large Language Model Study

Background: Large language models (LLMs) are increasingly used for musculoskeletal health information, yet their safety in time-sensitive shoulder and elbow presentations with red-flag features remains insufficiently defined. We evaluated safety behavior using standardized vignettes, focusing on safety-critical under-triage and prompt-dependent performance differences. Methods: Eighty fictional vignettes (40 shoulder, 40 elbow) were created and classified a priori as red-flag (n = 24) or non-urgent (n = 56). Each vignette was queried in a single-turn format using three fixed prompt types (patient-, general physician-, and specialist-oriented), yielding 240 responses. Two blinded orthopedic surgeons rated outputs using a prespecified 0-8 rubric across four domains. Safety-critical under-triage was defined as failure to recommend timely urgent evaluation in red-flag presentations. Decision stability was assessed using 20 paired vignette sets differing by one predefined clinical variable. Results: The overall mean score was 6.42 ± 1.12 and was lower for red-flag than for non-urgent responses (5.28 ± 1.21 vs. 6.93 ± 0.81). Across the 72 prompt-specific responses generated for the 24 red-flag vignettes, urgency was correctly recognized in 53 responses (73.6%). Safety-critical under-triage occurred in 19 of 72 red-flag responses (26.4%) and was most frequent with patient-oriented prompts (10/24, 41.7%), followed by general physician-oriented prompts (6/24, 25.0%) and specialist-oriented prompts (3/24, 12.5%). Decision instability, defined as an inconsistent directional change after modification of a single risk-related variable, occurred in 6 of 20 paired vignette sets (30.0%). Conclusions: The evaluated LLM performed consistently well in non-urgent scenarios but showed prompt-dependent safety vulnerabilities in red-flag conditions, driven primarily by under-recognition of urgency. These findings support caution for unsupervised patient-facing use, highlight the need for explicit safeguards in high-risk presentations, and underscore the value of safety-focused evaluation frameworks in musculoskeletal care.