Skip to main content

Gemini scored 82.78% on distorted SAT images in smart-glasses study

1 NOVEMBER 2026·2 MIN READ·4 SOURCES

Researchers found that Gemini 1.5 Pro-002 scored 82.78% on SAT images altered to simulate a distant view, and warned that smart glasses could relay exam questions to AI.

Gemini scored 82.78% on distorted SAT images in smart-glasses study

Key takeaways · 4

  • 01

    Gemini scored 82.78% overall on the distorted images, while GPT-4o scored 75.69%.

  • 02

    The models' ranking reversed on clean images: GPT-4o scored 87.22% and Gemini scored 84.58%.

  • 03

    The test placed one question on each page, so it did not fully capture the task of finding a question on a real exam page.

  • 04

    South Korea's rules invalidate an exam if a candidate wears or possesses smart glasses during the test.

What the test found

At the 17th International Conference on Computer Supported Education, CSEDU 2025, researchers presented results showing that commercial models maintained high accuracy on exam questions even when image quality was poor.[1] On the distorted SAT images, Google's Gemini 1.5 Pro-002 scored 82.78% overall, compared with 75.69% for OpenAI's GPT-4o.[1] Gemini answered 87.96% of math questions and 78.54% of reading-comprehension questions correctly; GPT-4o answered 74.69% of math questions and 76.52% of reading questions correctly.[1] The researchers warned that smart glasses could pass exam questions to AI without students needing to collaborate with one another.[1]

What the scores do—and do not—show

For the simulated distant view, researchers applied 75-degree perspective distortion, a three-pixel blur and 5% noise to the exam images.[2] On clean, undistorted images, GPT-4o scored 87.22% and Gemini scored 84.58%, reversing the ranking seen in the distorted-image test.[1] The findings therefore show different relative performance under the two image conditions; they do not establish that either model would achieve those scores through glasses in an exam hall.[1] The researchers explicitly said the experiment had not been validated using actual functioning smart glasses during an exam.[1]

Important limits and exam responses

Researchers processed the material so that only one question appeared on each page, unlike a typical page containing multiple questions.[1] They said this setup did not fully capture the challenge of locating which question to solve on a real exam page.[1] The researchers also noted that smart glasses without displays can be virtually indistinguishable from ordinary eyeglasses to people unfamiliar with them.[1] In South Korea, five cases of smart-glasses cheating were confirmed in national qualification exams during the year reported, and test-takers were caught switching glasses after passing pre-exam checks.[1][4] The Ministry of Education planned one dedicated staff member per CSAT test site to detect smart glasses.[1]

South Korea's stated penalties

South Korea's rules make a test invalid if a candidate wears smart glasses and suspend that candidate's eligibility to take the CSAT the following year.[1] The rules also invalidate an exam if a candidate merely possesses smart glasses during the test.[1] The Ministry of Education decided to pursue complaints or investigations seeking criminal punishment for smart-glasses cheating.[1] It was also pushing to extend the CSAT eligibility suspension for cheaters from one year to three years.[1] These measures describe South Korea's response; the researchers' image benchmark itself did not test a live proctoring or enforcement process.[1][2]

Exam administrators can use the findings to assess whether rules and screening procedures account for discreet camera-equipped devices, while keeping the benchmark's limits in view. The test measures model performance on prepared images, not the full process of using smart glasses during an in-person exam.

Why it matters
Story quiz

Bring this to your team — one session a day, built for your industry.

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

For teams

How this developed

  1. 1 November 2026

    Gemini scored 82.78% on distorted SAT images in smart-glasses study

Sources

AI fluency, one session a day, built for your work.