Online proctoring software disproportionately flags students with darker skin for cheating because the underlying AI frequently fails to recognize their faces.
AI-driven exam monitors are twice as likely to mark students with dark skin as "suspicious" compared to their light-skinned classmates. This automated bias creates a digital hurdle that leads to false accusations and heightened anxiety for students of color, particularly women.
This isn't just a technical glitch; it’s a threat to a student’s academic record. If your child is using software like Respondus for a high-stakes exam, a "suspicious behavior" flag can trigger an investigation or a failing grade before a human ever looks at the footage. The psychological toll of being "watched" by a system that doesn't see you correctly adds a layer of stress that can actively tank test performance and long-term confidence.
Remote learning forced schools to adopt automated surveillance tools at a massive scale, often without vetting them for equity. Researchers were concerned that these tools—trained on data sets that historically underrepresent people of color—would carry the same racial biases found in facial recognition used by law enforcement. They analyzed students in STEM courses to see if these "neutral" algorithms were creating a tiered system of scrutiny based on skin tone and sex.
The numbers show a stark divide in how the AI treats different students.
- Black students averaged about six "suspicious behavior" flags per exam, while white students averaged just over one.
- The software failed to detect a face entirely for dark-skinned students 22% of the time, compared to only 8% for light-skinned students.
- Women with the darkest skin tones faced the highest risk, suggesting the AI struggles with the combination of gendered features and skin tone.
- Students with the darkest skin were flagged for review more than twice as often as their lighter-skinned peers.
The software isn't just "detecting" cheating; it is effectively punishing students for having skin tones the algorithm wasn't designed to process. When a student is flagged for "suspicious behavior" because the AI lost track of their face, the burden of proof shifts to the student to prove they weren't cheating. This turns a standard test into an interrogation by algorithm, where "innocent until proven guilty" is replaced by "suspicious until manually cleared."
The sample size for the darkest skin tone category was small, with only 41 students represented. While the findings are statistically significant, the data comes from a single institution (University of Louisville) using one specific software package, Respondus Monitor. Skin tones were also manually categorized by researchers from photos, which involves a degree of human subjectivity.
- If your child is taking an exam using automated proctoring software, set up high-intensity, front-facing lighting to help the sensor track their face more clearly and reduce "missing face" flags.
- If your school uses "priority scores" to rank student behavior, ask the administration if instructors are required to manually review every flag before a student is penalized or contacted.
- If your child has darker skin, talk to them about the possibility of being flagged so they aren't blindsided or discouraged by a technical error during the middle of a high-pressure test.
- If your student is flagged for "suspicious activity," request the raw footage and the specific reason for the flag immediately to verify if the issue was a face-detection failure rather than actual behavior.
Automated proctoring is a flawed tool that places an unfair burden on students of color and can compromise their academic success. Parents should treat "suspicious behavior" flags as technical errors first and disciplinary issues second. Demand human oversight for every automated decision to ensure your child isn't penalized by a biased algorithm.
Yoder-Himes, Deborah R., Asif, Alina, Kinney, Kaelin et al. (2022). Racial, skin tone, and sex disparities in automated proctoring software. Frontiers in Education. doi:10.3389/feduc.2022.881449 — frontiersin.org