AI Has Evolved into Smart Glasses, Yet Universities Still Say ‘Don’t Use It’: Question Design Holds the Key to AI-Era Education

Generative AI Becomes the Default for Learning
Universities’ Blanket Bans Collide With Entrenched Use
“Teach Questions, Not Answers”: Evidence-Based Inquiry Holds the Key

Controversy over AI-enabled academic misconduct has reignited across South Korean universities after more than half of the students enrolled in a Seoul National University computer science course were penalized for using generative artificial intelligence (AI) to complete an assignment. While universities seek to curb its use through notices and grade penalties, AI has already penetrated assignments, online examinations and even in-person testing venues through smart glasses. Students have embraced AI as a routine learning tool, yet critics say university assessment systems remain tethered to standardized answers and tests of memorization.

Assignments Breached by AI, Universities Resort to Grade Penalties

According to a Seoul National University official on August 5, 36 of the 62 students enrolled in a major course offered by the university’s Department of Computer Science and Engineering recently received grade penalties for using AI on an assignment. Students who voluntarily disclosed their use reportedly received a zero on the assignment, while those whose use was detected were given a D-minus in the course.

The course had explicitly announced from the outset that “generative AI must not be used in any class-related activity.” According to the syllabus posted before the semester began, the professor stated: “To cultivate students’ independent problem-solving abilities and critical-thinking skills, the course will proceed without the use of generative AI in any class activity.”

The controversy intensified after the Department of Computer Science and Engineering’s student council issued a statement questioning the consistency of the criteria used to identify AI use and the scope of enforcement. In an opinion submitted to the professor, the council argued that the measure was unfair because “students who answered honestly were penalized, while those who actually used generative AI but denied doing so or escaped detection may have faced no consequences.” The council thus challenged the penalties on the grounds of fairness in enforcement.

An official from the department’s student council explained, “The point was that when human and AI-generated answers are similar, students who did not use AI may have been mistakenly identified as having done so and penalized.” The professor responded: “I simply cannot accept the argument that punishing an offender is unfair because not every person who committed the offense was caught.”

AI Adoption Outpaces Detection

The problem extends well beyond Seoul National University. Last year, a professor teaching “Natural Language Processing and ChatGPT” at Yonsei University detected indications of widespread misconduct during an online midterm examination. After confirming that some students had used generative AI tools such as ChatGPT to answer questions, the professor announced: “Those who confess will receive a zero; those who conceal their actions will face suspension proceedings.” The incident triggered a fundamental debate on campus over whether using AI constitutes cheating. Misconduct involving AI in assignments and examinations has already been widely documented at leading universities overseas.

The United Kingdom has produced some of the most detailed figures on the scale of AI-enabled academic misconduct. According to the British daily The Guardian, more than 6,900 cases involving the use of AI were detected during the 2023–2024 academic year, from September 2023 through August 2024. The rate exceeded 5.1 cases per 1,000 students, more than triple the 1.6 recorded in the previous academic year. The figures reflect only data from universities that separately classified AI misuse, prompting faculty members to warn that the reported cases represent merely the tip of the iceberg.

Table 1. AI Use Among University Students

Clashing Perceptions Amid Ubiquitous AI Use

Underlying these conflicts is a deep divide in perceptions of AI use. A substantial share of students regard AI as a default tool, making blanket bans difficult for them to understand. A university professor said, “Students who use AI as a matter of course may regard work as their own as long as AI provided only some assistance and did not produce the entire output. The students’ criticism appears to reflect that perception.” The professor added, “The objection may seem unreasonable, but under these circumstances, universities need detailed guidelines defining precisely what forms and degrees of use are prohibited.”

AI has already become embedded in students’ routine learning practices. A 2025 survey by the Digital Education Council (DEC), a global consortium of universities and corporations, found that 86% of 3,839 university students across 16 countries used AI for academic work. Of the respondents, 24% said they used AI daily and 54% weekly. Each student used an average of 2.1 AI tools.

A 2025 survey of 1,041 undergraduate students by the UK Higher Education Policy Institute (HEPI) put the overall AI adoption rate at 92%. The share using generative AI for assignments and assessments reached 88%. A study by Cornell University researchers analyzing approximately 95,000 students at 20 public research universities in the United States found that 37% used AI every month and 9% had used it to engage in academic misconduct. Regular usage among computer science majors stood at 62%, while 26% of daily users reported experience using AI for misconduct.

Technology Races Ahead as Assessment Methods Stand Still

AI’s penetration of examination venues has expanded from laptops and smartphones to smart glasses. In May, three candidates were caught wearing AI glasses during national technical qualification examinations, including tests for electrical and fire-protection systems engineers. A test proctor at an examination center in Gwangju caught a candidate in his 40s after noticing a suspicious light reflected in the lenses. An investigation found that he had connected an AI application he developed to the glasses to test whether answers would appear on the lenses. In June, prosecutors summarily indicted the man on charges of violating the National Technical Qualifications Act. It marked South Korea’s first criminal prosecution for examination misconduct involving AI glasses.

Misconduct involving AI glasses had been detected since last year. During the Test of Proficiency in Korean (TOPIK) administered in South Korea last year, three Chinese candidates were caught wearing smart glasses. The devices used cameras mounted on the frames to scan questions and display answers on the lenses. Two candidates were also caught entering testing rooms with AI glasses during two separate Test of English for International Communication (TOEIC) examinations held in May. Their scores were invalidated, and they were barred from taking the examination for four years.

The Ministry of Education sent an official notice to municipal and provincial education offices explicitly designating AI glasses as prohibited items in testing venues and instructing proctors to watch for candidates repeatedly touching the temples of their glasses. Chung-Ang University College of Medicine also revised its student assessment guidelines last year to require checks for thick eyeglass frames and repeated handling motions. Yet on-site countermeasures have already fallen behind technological advances. AI glasses use frame-mounted cameras to recognize questions, while generative AI delivers answers through lens displays or miniature speakers. The latest products weigh around 50 grams and feature temples as slim as those on ordinary horn-rimmed glasses. Under these conditions, repeatedly

The Waning Viability of Answer-Based Assessment

What universities now require is a new body of research into learning assessments aligned with the pace of generative AI development. Assignments that require students to identify and submit predetermined answers, along with examinations that test their ability to reproduce memorized content, fall squarely within generative AI’s strongest capabilities. When professors pose established questions and grade only the final answers, student performance inevitably hinges more on access to AI and the ability to conceal traces of its use than on the depth of their knowledge. Rigorous assessment should instead test students’ ability to formulate hypotheses under unfamiliar conditions, identify errors in AI-generated responses and logically defend their chosen solutions.

The educational impact of AI varies according to course design. The Organisation for Economic Co-operation and Development (OECD) found that students who delegated assignments to general-purpose AI produced more polished submissions, but the learning gains disappeared—or their scores declined—when they were tested without access to AI. In an experiment involving 194 physics students at Harvard University, the group using an AI tutor designed according to pedagogical principles achieved more than twice the learning gains of the group receiving active-learning instruction. Both blanket prohibition and unrestricted delegation of answers to AI, however, failed to guarantee educational benefits.

From Finding Answers to Designing Questions

Universities overseas are increasingly introducing courses that require students to use AI for data analysis and independently verify errors in its output. A representative example is a master’s-level geoscience course at Germany’s University of Jena, featured this year in the international journal Communications Earth & Environment. Students entered multiple primary sources into AI systems and were then required to identify contradictions among the literature, passages supported by weak evidence and points on which authors offered conflicting interpretations. They cross-checked the AI-generated content against the original texts to verify the facts and documented the evidence used to accept or discard each response. To review this verification process, faculty members required students to submit their prompts, reference materials and even the responses they had rejected. Grading criteria were also expanded from the accuracy of final answers to the sophistication of the questions, the credibility of the sources and the ability to verify AI-generated results.

New York University’s Stern School of Business has also introduced a format in which an AI voice agent subjects students to a continuous series of questions about their projects. The AI examiner sequentially asks about the project’s objectives, data selection, model design and causes of failure, then randomly presents a case from the course and requires students to defend their judgment on the spot. A total of 36 students completed examinations averaging 25 minutes over nine days, and multiple AI models cross-graded the conversation transcripts. The University of Sydney likewise combines open assessments that permit AI use with secure assessments conducted under supervision to verify individual competence. As the longstanding practice of judging students by completed answers erodes, the ability to explain and defend one’s reasoning in real time is emerging as a decisive benchmark of academic performance.