Case Study 2: AI Proctoring — Fairness Under the Webcam

The Context

When the COVID-19 pandemic forced universities online in March 2020, institutions faced an immediate problem: how do you maintain academic integrity when students are taking exams at home? Within weeks, millions of students around the world were being monitored by AI proctoring software they had never heard of, operating under terms of service they had never read, analyzing them in ways they did not understand.

This case study examines the experiences of students at a large public university — let us call it State University — that adopted an AI proctoring system called ExamEye (a composite based on documented issues with real proctoring platforms) for all online examinations beginning in fall 2020.

How ExamEye Works

ExamEye operates on the student's personal computer. Before the exam begins, the student installs the ExamEye browser extension. The software then:

  1. Verifies identity by comparing the student's face to their student ID photo using facial recognition.
  2. Locks down the browser so the student cannot open other tabs, applications, or websites during the exam.
  3. Monitors via webcam throughout the exam, recording video and analyzing it in real time for "suspicious behavior."
  4. Monitors audio through the computer's microphone, flagging significant sounds.
  5. Tracks eye movements and head position, flagging extended periods of looking away from the screen.
  6. Generates a "suspicion score" for each exam session, which the instructor can review alongside the flagged video clips.

The software does not determine whether a student cheated. It flags behavior for human review. Instructors receive a dashboard showing each student's suspicion score and can view the specific moments that triggered flags.

Five Student Experiences

Darius — "The Software Couldn't See Me"

Darius is a junior majoring in biology. He is Black. During his first AI-proctored exam, the identity verification step repeatedly failed. The software could not match his face to his student ID photo. He tried repositioning his laptop, adjusting his lighting, and sitting in different parts of his apartment. After 20 minutes of failed attempts — while the exam clock was already running — he called technical support. The support agent suggested he "point a desk lamp directly at your face." He did. The light was uncomfortably bright and caused him to squint for the duration of the exam.

"I started the exam 20 minutes late, stressed, with a lamp blinding me," Darius said. "I got a C. On the same material, I had gotten a B+ on the in-person midterm."

Darius's experience was not unique. Research on facial recognition systems — including the foundational work of Joy Buolamwini and Timnit Gebru at MIT, published in 2018 — has documented significantly higher error rates for people with darker skin tones, particularly darker-skinned women. These error rates extend to the facial recognition components of proctoring software.

Maria — "My Kids Are Not Cheating Devices"

Maria is a 36-year-old nursing student taking online classes while caring for three children, ages 4, 7, and 11. Her husband works nights. She does not have a home office. She takes exams at the kitchen table.

During a pharmacology exam, her four-year-old wandered into the kitchen asking for a snack. ExamEye flagged this as "third party in testing area." Her seven-year-old turned on the television in the adjacent living room, and the sound was flagged as "unauthorized audio." When Maria turned her head to tell her son to turn down the volume, the software flagged the head movement as "extended gaze away from screen."

Her exam session received one of the highest suspicion scores in the class. Her professor — who had 247 exam sessions to review — saw the high score, looked at the first two flagged clips, and emailed Maria asking her to "explain the presence of unauthorized individuals during the examination."

Maria wrote a two-page response explaining her living situation, her children's ages, and the fact that she had nowhere else to take the exam. Her professor accepted the explanation, but the experience left her humiliated. "I am working harder than anyone in this class," she said. "And the software treated me like a criminal because I am a mother."

James — "It Flagged Me for Thinking"

James has ADHD. When he is thinking through a problem, he looks around the room, taps his desk, and occasionally mouths words to himself. These behaviors — which are documented accommodations through the disability services office — were flagged extensively by ExamEye: "frequent gaze aversion," "repetitive physical movements," and "unauthorized verbalization."

James had registered with the disability office and received accommodations for extended time. But the proctoring software had no mechanism for adjusting its behavioral monitoring for students with disabilities. The flags were generated automatically, and his instructor — reviewing hundreds of flags — had no way to know which students had accommodations that might explain flagged behavior.

After the exam, James received an email from the academic integrity office. "Do you know what it feels like to have your disability flagged as cheating?" he asked during a subsequent meeting with the dean. "I was thinking. Thinking looks different for me than it does for you."

Anika — "I Stopped Studying and Started Performing"

Anika does not have a disability, does not share her apartment, and the facial recognition worked fine for her. Her complaint is different: the surveillance changed how she took the exam.

"I was so focused on not triggering the software that I could not focus on the test," she said. "I was afraid to look up. I was afraid to scratch my face. I timed my eye movements. I was performing 'test-taker' for the camera instead of actually thinking."

Anika's experience aligns with research on the psychological effects of surveillance. A 2021 study of students taking AI-proctored exams found significantly higher test anxiety and lower self-reported performance compared to unproctored conditions. The Panopticon effect — the idea that being watched changes behavior, even when the watcher is not actually observing at any given moment — was alive and well in Anika's apartment.

Tyler — "I Cheated and the Software Didn't Catch Me"

Tyler is included in this case study because his experience matters for the full picture. Tyler used a second device — a phone, placed out of the webcam's view — to look up answers during an AI-proctored exam. He received no flags and scored an A.

Tyler's success does not mean the software is useless. It means the software catches a specific kind of cheating (overt, visible behavior) while missing other kinds (subtle use of secondary devices, pre-written notes on paper out of frame, assistance from another person in a different room). Students motivated to cheat can often do so despite proctoring, while students behaving innocently are the ones most likely to be flagged.

The University's Response

After receiving complaints from students and faculty, State University convened a task force to review its proctoring policies. The task force included students, faculty, disability services staff, IT professionals, and a representative from the software company.

The task force findings included:

  • False flag rate: Over 60 percent of high-suspicion-score exam sessions, upon instructor review, involved no evidence of cheating. The software was generating more noise than signal.
  • Disparate impact: Students of color, students with disabilities, and students in non-traditional living situations were flagged at significantly higher rates.
  • Instructor burden: Most instructors did not have time to review all flagged sessions carefully, leading to either blanket dismissal of flags (undermining the software's purpose) or snap judgments based on superficial review (risking false accusations).
  • Student well-being: Student surveys revealed that AI proctoring was the single most-cited source of exam anxiety, ahead of difficulty of the material.

The task force recommended:

  1. Eliminate AI proctoring for most exams, reserving it for high-stakes licensing exams where no alternative exists.
  2. For remaining proctored exams, provide alternative testing arrangements for students with documented accommodations.
  3. Invest in assessment redesign — moving toward open-book, project-based, and oral assessments that reduce the incentive and opportunity to cheat.
  4. If proctoring is used, require instructors to review all flagged sessions within 48 hours and contact students directly before involving the academic integrity office.
  5. Publish transparent data on flag rates by demographic group and audit the software for bias annually.

The university adopted most of these recommendations. AI proctoring use dropped by 70 percent within one year.

Discussion Questions

  1. The Fairness Question: Darius, Maria, James, and Anika each experienced harm from the proctoring software, but the mechanisms were different. Identify the specific type of harm (bias, environmental assumptions, disability, psychological) in each case. Which do you think is most urgent to address?

  2. Tyler's Case: Tyler cheated and was not caught. Does this undermine the case for proctoring, or does it suggest the software needs to be improved? Can any proctoring system truly prevent determined cheating?

  3. The Panopticon in Practice: How does Anika's experience illustrate the Panopticon effect discussed in Chapter 12? What is the cost when surveillance changes behavior in ways that reduce the quality of the surveilled activity (in this case, test-taking)?

  4. Alternative Assessments: The task force recommended moving toward open-book and project-based assessments. What are the advantages of these approaches? What are the limitations? Are there subjects or contexts where they would not work?

  5. The Company's Role: ExamEye generated revenue during a period of massive demand for online proctoring. Does the company have an ethical obligation to test its software for bias before selling it? Should there be regulatory requirements for proctoring software similar to those for medical devices?

  6. The Root Cause: AI proctoring was adopted because universities needed a way to maintain academic integrity during online exams. Is the real problem the technology, or is it the assessment model (high-stakes, closed-book, timed exams) that creates the need for surveillance in the first place?

  7. Your Position: After reading this case study, where do you stand on AI proctoring? Under what conditions, if any, is it acceptable? Write a 100-word position statement.