Background: As parents increasingly seek information about child health on digital platforms, the risk of misinformation has also risen. Aims: This study aimed to evaluate the performance of large language models (LLMs) in identifying, correcting and generating guideline-adherent, evidence-based responses to paediatric misinformation. Method: Twenty common paediatric misinformation statements across nine thematic areas were identified based on literature review and expert opinion. These statements were presented to four different LLMs (ChatGPT, Gemini Advanced, Claude, Microsoft Copilot) with three repetitions each. The generated responses were evaluated by one physician and two nurses in terms of accuracy, guideline adherence, explanation quality and risk. Results: ChatGPT and Gemini demonstrated the highest and most consistent performance in accuracy and guideline adherence. Claude showed some deficiencies in certain explanations, while Copilot exhibited lower performance in guideline adherence and explanation depth compared to the other models. Risk scores were low across all models, and no hazardous content was observed. Furthermore, the models' abilities to correct misinformation varied in terms of guideline-compliant explanations and risk communication. Conclusion: Within the scope of the evaluated paediatric misinformation statements and current model versions, LLMs may provide conditionally safe and guideline-aligned information for parents. However, professional oversight and adherence to evidence-based paediatric guidelines remain essential. This study systematically highlighted the potential applications and limitations of LLMs in digital health safety and paediatric nursing practice.