ExamSoft’s supervisory software has a face detection problem

Gabe Teninbaum, a professor at Suffolk University School of Law, is calling on ExamSoft to fix a serious bug in its testing software: failure to recognize faces. It is a problem that can delay candidates – or prevent them from starting exams – and, according to reports, disproportionately affects people with dark skin tones.

ExamSoft software records students as they complete remote exams and monitors for signs of academic dishonesty. Teninbaum’s report addresses an ExamSoft feature called ExamID, which aims to verify that test participants are who they say they are. The first time a student connects to the exam portal, he uploads a photo of himself (his “base image”); they are then asked to take another selfie before starting future exams, which the software checks on their original photo.

Research has found that facial recognition algorithms consistently make more mistakes in identifying black faces than white ones. And while these studies did not specifically focus on ExamSoft, it does not appear to be an exception. In September, several non-white candidates told the New York Times that the software was unable to identify them due to “bad lighting” – a problem that Teninbaum, who has fair skin, was unable to replicate.

Earlier this fall, Teninbaum decided to find a solution. He believes that such mistakes add undue stress to an already stressful period of time. “These are students who are about to take a high-risk exam with a lot at stake, and that is very undesirable,” said Teninbaum in an interview with The Verge.

“Whenever you go for an exam, you just want to focus on it,” he added. “You don’t want to feel that you have these additional challenges.”

Teninbaum also believes that optics is important; schools owe this to marginalized students not to trust a category of software that is known to be discriminatory. “Students deserve to feel that their institution is doing what it can to protect their rights, interests and dignity,” he says.

In its report, which will be published in The Journal of Robotics, Artificial Intelligence, and Law, Teninbaum describes the workaround he found.

He suggests that schools give each student an identical generic-based image. He then proposes that they should ask ExamSoft to enable “deferred identification”, a feature built into the software that allows students to proceed with exams even if identification fails. This feature is basically hidden – it is not mentioned anywhere on the ExamSoft website (at least, not that I could find). Only ExamSoft can activate it.

Together, these adjustments will cause ExamSoft to incorrectly identify all test participants. But they can still proceed with their exams – ExamSoft will send the selfies to the school after that, and instructors can manually check them all. “We know who our students are,” says Teninbaum. “We can be sure that students are who they say they are and avoid subjecting students to this type of challenge.”

He also suggests that ExamSoft make the “deferred identification” feature accessible to customers. “The reporter asks ExamSoft to create a resource by which institutions can simply turn on / off, thereby bypassing ExamID until the time when the technology matures into one that does not discriminate,” says its report.

Teninbaum hopes that these changes will last beyond the COVID-19 pandemic and can help students feel more comfortable in remote classes. “It will be a growing problem as people get more and more online for their education,” he says.

Still, it solved only part of the problem. Students experienced a series of hiccups with ExamSoft’s monitoring software. More than 3,000 people who used the platform to take the California order exam in October had their videos flagged for possible rule violations – almost 36% of candidates who took the exam online. Users have reported audio problems and other technical flaws as well.

A group of six U.S. senators – including Richard Blumenthal, Elizabeth Warren and Corey Booker – wrote an open letter to ExamSoft in December, highlighting the potential harm to black and disabled students, among several other concerns.

ExamSoft did not immediately respond to a request for comment.

Source