"An 83 Without Studying?"... AI Glasses Shake Up Exam Rooms

Gemini Gets 83% of SAT Questions Correct

German Researchers Test 720 Questions
"Could Threaten the Fairness of In-Person Exams"

As cheating using smart glasses equipped with artificial intelligence (AI) features has emerged as a problem around the world, a study has found that AI answered more than 80% of questions on the U.S. Scholastic Assessment Test (SAT) correctly.

Meta's AI glasses. The photo is not directly related to the article. Yonhap News.

According to Yonhap News Agency on Nov. 1, a paper titled “Large Language Models and Vision-Language Models as a Threat to the Integrity of In-Person Exams: Assessment Using Smart Glasses and Computer Vision,” presented at the 17th International Conference on Computer Supported Education (CSEDU 2025) by researchers Rupert Urbanski and Ralph Peters of Martin Luther University Halle-Wittenberg in Germany, reported the results of tests assessing AI models’ ability to solve questions based on the SAT.

The questions used in the experiment were from previously released SAT tests and comprised a total of 720 questions: 396 reading questions and 324 math questions. Rather than connecting actual smart glasses to AI and having it take the test, the researchers artificially recreated the image distortion that could occur when viewing an exam paper through smart glasses.

When Google’s “Gemini 1.5 Pro-002” was given the questions processed in this way, its overall accuracy reached 82.78%. Specifically, it scored 87.96% on math questions and 78.54% on reading questions. OpenAI’s “GPT-4o” also answered 75.69% of all questions correctly, scoring 74.69% on math questions and 76.52% on reading questions.

When the models were given only clear questions without image distortion, GPT-4o achieved an accuracy rate of 87.22%, while Gemini scored 84.58%.

Based on their findings, the researchers noted that “smart glasses could be used as an input device to relay exam questions to AI,” adding that “smart glasses without a display function are virtually indistinguishable from ordinary glasses to someone unfamiliar with them.”

However, the study did not test the glasses in an exam room while participants wore functioning smart glasses. Although the exam questions were distorted, only one question appeared on each page, so the experiment did not fully reflect the real test-taking situation, in which AI would have to identify which question to solve among multiple questions.

Meanwhile, cheating using smart glasses has already occurred repeatedly in South Korea. Starting with this year’s College Scholastic Ability Test, education authorities plan to deploy dedicated personnel at every test site to identify smart glasses. They also plan to treat cheating with smart glasses as a deliberate and planned disruption of the proper administration of a national examination, and to actively file complaints or request investigations so that offenders can face criminal penalties, including obstruction of the performance of official duties by fraudulent means or business obstruction.

This content was produced with the assistance of AI translation services.