OpenAI's research team recently announced a technological innovation aimed at enhancing the self-explanatory capabilities of large language models (LLMs) such as GPT-4. This technological innovation draws inspiration from the concept of the "prover-verifier game" and utilizes an interactive competition mechanism between two AI models to improve the verifiability and transparency of model answers. In this game, one model plays the role of the "prover" and is responsible for providing answers, attempting to convince a weaker model acting as the "verifier" of the correctness of its answers. The verifier model focuses on verifying the accuracy of the answers, thereby driving the prover model to provide clearer and more accurate explanations. Through multiple rounds of gameplay and iterative training, the prover model demonstrates significant improvements in generating answers that are easily understandable to humans, while the verifier model also enhances its ability to identify incorrect answers. According to OpenAI's research report, employing this game-based training method results in better performance of the models in providing both correct and comprehensible answers. This research emphasizes the importance of improving the transparency and verifiability of AI system outputs, particularly in critical fields such as healthcare, law, and energy where this characteristic is crucial. Furthermore, this technology showcases its potential for potentially being used in the future to fine-tune AI models that surpass human intelligence, thereby enhancing their trustworthiness and security in practical applications. This discovery provides new perspectives and possibilities for the further development of AI technology.