Comparative evaluation of child and adolescent psychiatric-mental health nursing care plans generated by ChatGPT, Gemini, and DeepSeek: A cross-sectional text analysis.

Large language models are increasingly used in nursing documentation and education, but their value for child and adolescent psychiatric-mental health nursing care planning remains uncertain. This cross-sectional comparative text-analysis study evaluated care-plan texts generated by ChatGPT, Gemini, and DeepSeek. Thirty child and adolescent psychiatric-mental health nursing diagnoses were entered using a standardized prompt, producing 90 care plans. Text characteristics, reference accessibility, seven readability formulas from two calculators, DISCERN scores, and reliability indicators were summarized. Because each diagnosis was evaluated across all models, between-model comparisons used Friedman tests and Bonferroni-adjusted Wilcoxon signed-rank tests. Gemini had the highest DISCERN total score (63.8 ± 4.3), followed by ChatGPT (61.2 ± 3.8) and DeepSeek (59.7 ± 4.5; Q = 12.07, P = 0.002). ChatGPT produced the easiest texts by averaged Flesch Reading Ease score (53.9 (43.4-66.0)), whereas DeepSeek produced the most complex texts (42.1 (26.7-55.6); Q = 30.87, P < 0.001). All models exceeded recommended sixth-grade readability levels across grade-level formulas. Accessible references were most consistently associated with higher DISCERN total scores. Findings support task-specific model selection, source checking, readability revision, and psychiatric-mental health nurse review before AI-generated care plans are used in clinical or educational settings.