港科大教授实测 AI 眼镜考试“作弊”：30 分钟交卷，碾压 95% 的学生

Core Insights - The experiment conducted at Hong Kong University of Science and Technology demonstrated that an AI-powered smart glasses, equipped with the ChatGPT-5.2 model, could successfully complete a final exam in Computer Network Principles, achieving a score of 92.5, placing it in the top 5 among over 100 human students [1][9][12]. Hardware Selection - The research team evaluated 12 mainstream commercial smart glasses, including products from Meta, Xiaomi, and Rokid, to find a suitable device for the experiment [3][4]. - The final choice was Rokid smart glasses due to their superior SDK and development flexibility, which were essential for the experiment's requirements [5][7]. AI Model Selection - The team selected OpenAI's latest model, ChatGPT-5.2, for its strong response speed and general knowledge capabilities, which were crucial for the exam performance [7]. Exam Process - The exam process involved the AI glasses capturing questions via a camera, transmitting them to the cloud for processing, and then displaying the answers back on the glasses for the student to copy [9][12]. - The AI achieved full marks in multiple-choice and single-page short answer questions, and performed well in multi-page short answer questions, showcasing strong reasoning capabilities [9][12]. Technical Limitations - The experiment revealed significant limitations in current commercial AI glasses, particularly regarding power consumption and camera clarity, which directly affected the AI's performance [12][13]. - The glasses' battery dropped from 100% to 58% within 30 minutes under exam conditions, indicating a need for better power management for prolonged use [12]. Educational Implications - The results raised questions about the validity of traditional educational assessment methods, which focus primarily on standardized answers, as AI can excel in these areas [13][14]. - The experiment highlighted the need for a shift in educational evaluation from merely assessing final answers to understanding the reasoning processes and learning paths of students [25][29].