Burning Money on Impractical AI Glasses? How "Cheating Glasses" Found Their "Scene" in the Exam Room — Exclusive Interview with HKUST Professor Reveals the Birth Story

Invest in stocks with Golden Kylin Analyst Reports—authoritative, professional, timely, comprehensive—helping you uncover potential thematic opportunities!

Daily Economic News Reporter | Li Xukui Ding Zhouyang Daily Economic News Editor | Huang Bowen

At the start of the spring semester, universities worldwide are caught in a wave of anxiety and excitement over AI (Artificial Intelligence).

“We have cut 16 undergraduate majors and directions, including translation and photography,” said Liao Xiangzhong, Party Secretary of the Communication University of China. “The future will be an era of human-machine division of labor,” and classrooms must be thoroughly restructured.

In March, Guo Yi, Vice President of Hong Kong University of Science and Technology (HKUST), told the Daily Economic News that HKUST has launched a series of AI education innovation programs this year, including a mandatory 6-credit AI general education course for all students. HKUST will also promote the “AI for X” model across all courses, “making AI an essential basic skill for science and engineering students.” How teaching, exams, and question-setting should change in the AI era is one of HKUST’s key focuses this year.

Interestingly, Professor Zhang Jun and Assistant Professor Meng Zili from HKUST’s Department of Electronic and Computer Engineering have conducted an experiment to pre-test the application of AI wearable devices in exams.

They developed a pair of “AI cheating glasses,” based on domestic AR smart glasses Rokid, equipped with OpenAI’s GPT-5.2 model. In their experiment, the seemingly ordinary black-rimmed glasses, with faint green lights on the temples, could automatically recognize questions and display answers on the lenses. Wearing these glasses for 30 minutes allowed students to complete notoriously difficult “Computer Network Principles” exams with a score of 92.5, surpassing 95% of test-takers. A scene from the movie “Genius” was actually reenacted.

Experiment photos Source: Provided by interviewee

In fact, there have been real cases of students cheating with AI glasses at universities. On social media, some users reported that their schools issued notices mentioning students using smart glasses for cheating.

As tools evolve to easily help people score high, what should universities test? The Daily Economic News recently visited HKUST and interviewed Zhang Jun and Meng Zili. Here is a summary of the interview.

The idea of conducting an “AI glasses” cheating experiment originated from a misunderstanding by Meng Zili. During an invigilated exam, he saw a student wearing sunglasses and suspected cheating with AR glasses. Upon closer inspection, he realized it was just ordinary sunglasses. However, this “imaginary enemy” was highly instructive: if AI cheating glasses really appeared, how should exams respond? This explores smart hardware and directly addresses the deep water of university exam reform in the AI era.

NBD: Before the experiment, what preparations were made for developing these glasses? Why choose domestic AI glasses Rokid?

Meng Zili: Our project started in summer 2025. Initially, the goal was simple—have the AI glasses answer a full exam under real test conditions. But it’s not as simple as just buying a pair of glasses. We iterated for four months, optimizing algorithms to ensure the glasses could effectively capture exam paper information, transmit it to a large model, and quickly produce answers. The delay couldn’t be too long; even a few minutes would be unacceptable.

Zhang Jun: There are many issues involved. Since we both have backgrounds in networking and communications, we see many problems that require network communication to solve. For example, a test paper viewed through the glasses has a very small field of view, and capturing and reading information clearly within that small range is challenging. Current glasses have limited capabilities; they don’t have the stable camera systems of drones, and even high-definition cameras are limited in this form.

Meng Zili: We tested many glasses, including Meta and various domestic startups. We bought about ten models that developers could modify. Rokid was the only one capable of testing for an hour; others ran out of power in about ten minutes. Rokid’s battery life was the longest, and its camera was relatively clear. Some glasses are more advanced, like Meta’s, supporting full-color images and videos, but for exam scenarios, that’s unnecessary.

NBD: During the experiment, could you see answers directly after wearing the glasses? How do the glasses recognize the exam paper?

Meng Zili: Yes, the student we tested with was our teaching assistant. We told her to copy whatever the glasses displayed. The large model just needed to recognize the questions and automatically infer the user’s needs—no additional commands needed. For multiple-choice, true/false, Q&A, matching questions, the glasses would display answers. The only limitation was that images couldn’t be displayed, but it could tell you how to connect items or the specific method to answer.

NBD: Did anything in the experiment surprise you?

Meng Zili: Yes, one question was answered incorrectly because the AI used extraneous knowledge. I graded the AI’s paper together with all students’ papers. Since its answers differed from the standard, I marked it wrong. Later, I found out it was using out-of-syllabus knowledge, which shocked me.

Sample of incorrect answers Source: Provided by interviewee

The AI glasses track has shifted from a few startups’ experiments to a battleground among global tech giants. Companies like Meta, Google, Apple, Huawei, Xiaomi, Alibaba, and others are entering the field, increasing the track’s popularity. Meanwhile, companies specializing in AI glasses like Thunderbird and Yingmu have completed multiple funding rounds, expanding industry supply.

However, despite the booming supply of AI glasses, application scenarios on the demand side remain unclear.

NBD: In March, Alibaba’s Qianwen launched its first AI glasses, “Qianwen AI Glasses.” How do you view the “arms race” among major model manufacturers in AI glasses? What is the current capability of AI glasses?

Zhang Jun: First, battery life is an issue. For certain scenarios, short-term use is feasible. If only using voice interaction, power consumption is lower; if involving cameras and real-time video interaction, power demands increase. Currently, AI glasses can’t support real-time video interaction; Meta’s AI glasses only record photos and videos. Due to power constraints, the cameras are small and image quality isn’t comparable to smartphones.

Meng Zili: Different AI glasses manufacturers have different goals. For example, Qianwen focuses on intelligence—hardware mainly supports their models. Some prioritize lightweight design; for instance, Yingmu’s AI glasses are very light, similar to regular glasses. Others aim for long battery life, hoping users can wear them all day. We welcome diverse development in AI glasses, as it promotes market progress. As for specific scenarios where users need AI glasses, that’s still an unsolved issue. Although cheating is unethical, it’s a strong motivation for students to use AI glasses. Other functions are less in demand now. We’re exploring this and have developed some apps.

NBD: Are you also developing applications for AI glasses?

Meng Zili: We focus more on software to make AI glasses easier to operate. Recently, we made some interesting AI glasses apps, like a “Subtext Translation” app. For example, when chatting with a leader who says “Let’s go,” the app can suggest responses.

Zhang Jun: We don’t want AI glasses to have only one function. In some demo videos, AI glasses tag everything seen—this is a bicycle, that is a car. I don’t need all those labels; I don’t want to be overwhelmed with information outside. I want a “wearable AI assistant” that’s always with me, following the wearer around. To achieve this, it needs to record what you see and understand you—so it can give you hints and information in advance. There are many systemic and technical challenges to reach this level of personalized human-device interaction, and there’s still a long way to go.

The arrival of AI acts as an amplifier of capabilities, quickly widening the gap between students with ideas and those without. In Zhang Jun and Meng Zili’s teams, this change is especially evident: previously, good ideas took one or two months of refinement; now, students can turn ideas into reality in a day with AI.

NBD: Many educators are cautious about students completing assignments entirely with AI, and some students overly rely on AI for answers in class. As computer science teachers, what’s your view?

Meng Zili: I always encourage students to use AI. I think this trend is inevitable—students asking AI for help with homework or questions is normal. We teachers also generate exams and slides with AI daily. I don’t think AI will make students lazy; before AI, many students copied answers online anyway. Conversely, some students don’t even know how to use AI proactively. My first class teaches students how to use intelligent agents, and I plan to teach them how to use OpenClaw. We built an AI platform so students can ask questions about course content directly to AI. The goal is for AI to pinpoint the exact slide and page in the PPT where the knowledge is, and tell students directly.

Zhang Jun: Yes, we want students to communicate and ask questions actively with AI. We’re not worried about students just staring at screens; we care whether they’re proactive. Currently, students don’t know what to ask AI or how to generate questions after reviewing a lecture. If they passively receive knowledge and just hand their homework to AI, it doesn’t help their growth much.

NBD: How do you prefer to select students? How do the abilities you mentioned reflect in your admissions process?

Meng Zili: I think it’s obvious—some students lack initiative. It’s hard to describe this feeling precisely.

Zhang Jun: We mainly recruit PhD students. Besides academic records, we value the interview process. Even though some students use AI to assist during online interviews, their genuine reactions and on-the-spot performance still reflect their overall qualities. Many technical tasks in the future may be replaced by AI, but genuine human interaction will always be needed. No one wants to only interact with machines. For example, programmers doing only coding are more easily replaced by AI; but those in leadership or management roles who can direct AI and communicate effectively within teams will find soft skills increasingly important.

A 15-year-old admitted to Tsinghua’s Department of Electronic Engineering, with top grades in university, awarded Tsinghua’s top scholarship, and holding a PhD from Tsinghua, joined HKUST at 23.

For Meng Zili, “winning in exams” is routine. But when his self-developed AI tools start disrupting traditional exam rules, this former “problem solver” has a very different perspective on future assessment methods.

NBD: Does this experiment mean traditional closed-book exams are now obsolete? How should universities adjust their evaluation methods?

Zhang Jun: Doesn’t this prove exams are more meaningful? At least for testing understanding of knowledge, closed-book exams will become more important. Of course, other formats should also be included, like oral exams, which are currently rare. These can give a more direct view of students’ grasp of knowledge, similar to how many companies now ask candidates to write code on a whiteboard during interviews. Presenting a problem and asking for a real-time response doesn’t necessarily require a correct answer, but it reveals the thought process, which is very valuable.

Meng Zili: I have a more “radical” view. I believe current exams serve two purposes: first, to verify if students have learned the knowledge; second, to filter and stratify. The skills needed in real work often don’t fully align with what exams test. I think the core should be cultivating abilities—students should use various AI tools without worrying about whether exams are closed or open book, or about specific knowledge points. With AI, we can better evaluate a student’s ability to complete projects and solve engineering problems—core skills that are more practical for their future careers. Specific knowledge testing becomes less important.

Zhang Jun: From a skills development perspective, we hope students can better utilize AI tools, making AI their personal assistant. It can tailor learning methods based on each student’s habits and personality. During this process, AI can identify students’ strengths, helping them discover their interests and passions. Many people struggle to find their strengths, but through AI interaction, it can guide and suggest, helping students develop personalized paths. Traditional education struggles with personalization, but AI offers this opportunity.

(Intern Chang Songzhen also contributed to this article.)

Cover image source: Meijing Media Library