ChatGPT-5.2 versus Gemini 3 Pro in neck pain and interventional pain management: a multidimensional comparison of performance, safety, readability, and short-term temporal consistency

Abstract Background Large language models (LLMs) are increasingly used by patients and clinicians to obtain health-related information, yet their reliability in neck pain and interventional pain management remains uncertain. This study compared ChatGPT-5.2 and Gemini 3 Pro using a multidimensional, short-term repeated-assessment framework evaluating clinical performance, safety, readability, expert assessment, and temporal consistency. Methods A standardized, investigator-developed set of 50 neck pain-related questions included patient-oriented questions, interventional pain management, safety and red-flag scenarios, and clinical vignettes. The same questions were submitted to both models at Week 0 and Week 2, with one response obtained for each model–question–time-point combination. Responses were anonymized, and evaluators were blinded to model identity. Accuracy, completeness, usefulness, and guideline adherence were assessed using predefined 5-point ordinal Likert-type scales, while red-flag omission, hallucinated facts, and overconfident inappropriate recommendations were evaluated as binary outcomes. Readability was assessed for patient-oriented questions, and reviewer comment scores were independently assigned by two evaluators. Analyses were exploratory, with no formal adjustment for multiple comparisons; p-values < 0.05 were considered nominally statistically significant. Results Two hundred responses were evaluated. At Week 0, Gemini had higher completeness and usefulness scores, whereas no nominally significant between-model differences were observed for accuracy or guideline adherence. At Week 2, ChatGPT had higher completeness scores, whereas Gemini had higher guideline adherence and usefulness scores; no nominally significant difference was observed for accuracy. Reviewer comment scores were lower for Gemini at both time points. Several within-model between-time-point differences were nominally significant but were interpreted as short-term variation rather than evidence of true longitudinal change. No nominally significant between-model differences were observed in the interventional pain management subgroup. Red-flag omissions were more frequent with ChatGPT than Gemini at Week 0 (24% vs. 8%; p = 0.008), whereas no nominally significant between-model difference was observed at Week 2. No nominally significant between-model differences were observed for hallucinated facts or overconfident inappropriate recommendations. ChatGPT had higher Flesch–Kincaid Grade Level scores at both time points, indicating greater textual complexity. Inter-rater reliability for reviewer comment scores ranged from good to excellent. Conclusions ChatGPT-5.2 and Gemini 3 Pro generally received high scores within the predefined evaluation framework, but between-model differences varied by outcome and assessment time point and did not consistently favor either model. Differences between Week 0 and Week 2 should be interpreted as short-term variation in observed model output rather than evidence of true improvement or deterioration. These findings support multidimensional, time-specific evaluation of LLMs. LLMs should be regarded as informational or decision-support tools rather than substitutes for professional clinical reasoning or individualized clinical decision-making, particularly in safety-critical or interventional contexts.