The paper introduces **ChatSpamDetector**, a system that leverages large language models (LLMs) to detect phishing emails. By converting email data into prompts suitable for LLM analysis, the system accurately determines whether an email is phishing and provides detailed reasoning for its decisions. The evaluation using a comprehensive phishing email dataset showed that the system, particularly when using GPT-4, achieved an accuracy of 99.70%, outperforming other models and baseline systems. The system's advanced contextual interpretation capabilities enable the detection of various phishing tactics and impersonations, making it a powerful tool in the fight against email-based phishing threats. The contributions of the paper include the proposal of **ChatSpamDetector**, its evaluation, and a detailed analysis of LLM responses, highlighting their sophisticated ability to extract and analyze key information from email headers and bodies. The system's architecture, dataset preparation, and experimental results are discussed, along with its limitations and potential future improvements.The paper introduces **ChatSpamDetector**, a system that leverages large language models (LLMs) to detect phishing emails. By converting email data into prompts suitable for LLM analysis, the system accurately determines whether an email is phishing and provides detailed reasoning for its decisions. The evaluation using a comprehensive phishing email dataset showed that the system, particularly when using GPT-4, achieved an accuracy of 99.70%, outperforming other models and baseline systems. The system's advanced contextual interpretation capabilities enable the detection of various phishing tactics and impersonations, making it a powerful tool in the fight against email-based phishing threats. The contributions of the paper include the proposal of **ChatSpamDetector**, its evaluation, and a detailed analysis of LLM responses, highlighting their sophisticated ability to extract and analyze key information from email headers and bodies. The system's architecture, dataset preparation, and experimental results are discussed, along with its limitations and potential future improvements.