Invited Talk: Detecting AI-Generated Academic Content

With the rapid advancement of large language models, generative AI has penetrated academic research extensively, bringing efficiency gains as well as severe challenges to academic authenticity. Reliable detection of AI-generated academic text has become an urgent research issue. Existing methods mainly target general texts and perform poorly on professional standardized academic writing, and their robustness declines rapidly with evolving LLMs. This study proposes an AI academic text detection method from the fundamental perspective of text complexity. We construct a multi-dimensional complexity indicator system and model joint feature distributions via Gaussian Mixture Model. Evaluated on datasets covering GPT-3.5-turbo, GPT-4o and GPT-5, our method achieves leading performance compared with 12 baselines, shows strong cross-model adaptability and remains stable under adversarial paraphrasing attacks.