This article shares a method for real-time monitoring of ChatGPT performance changes by adding specific prompts to custom instructions, making the model output its current identifier in every conversation, thereby quickly detecting whether ‘intelligence degradation’ or hallucination issues occur. Through testing, the author found that this method can effectively identify model switching or performance degradation, especially detecting anomalies in the first conversation. Although GPT-4o and GPT-4.1 sometimes don’t follow instructions, overall it provides a simple and practical AI monitoring solution that has high value for users and developers concerned about model stability.
