evidoria

← Back to browse

Good practice Imported

Proctorio at Vrije Universiteit Amsterdam — A Cautionary Case in AI Exam-Proctoring Bias

Netherlands · Amsterdam · See the Netherlands profile · See the Amsterdam profile

Evidence: Observational / pre–post Top 98% 13/100 · Ask Evidence Copilot about this practice

When Proctorio's face-detection anti-cheating software repeatedly failed to recognise Black student Robin Pocornie during a 2020 exam at VU Amsterdam, her complaint led to a landmark 2022 human-rights finding on algorithmic discrimination and a nuanced 2023 final ruling.

~30% %
Lower facial-recognition accuracy cited for darker skin tones
Proctorio at Vrije Universiteit Amsterdam — A Cautionary Case in AI Exam-Proctoring Bias

Details

Promoter
Vrije Universiteit Amsterdam
Period
2020-2023
Region (NUTS)
NL32
Keywords
exam proctoring, algorithmic bias, higher education, facial recognition

Context

In September 2020, Robin Pocornie, a Black student at Vrije Universiteit Amsterdam (VU), sat a remote exam using Proctorio, an AI anti-cheating tool that verifies identity via webcam and face-detection. The software repeatedly failed to recognise her face, logging 'face not found' and 'room too dark' errors even after she shone a lamp directly at herself — an experience she said other students of colour also reported.

Activities

In July 2022, Pocornie filed a complaint with the Netherlands Institute for Human Rights. In December 2022 the Institute issued an interim finding that she had presented sufficient evidence of algorithmic discrimination, shifting the burden of proof to VU.

Results

In its final judgment of October 2023, the Institute concluded VU had not discriminated through its use of the software itself, but had discriminated in how it handled her complaint; it explicitly stated the ruling 'does not rule out' that Proctorio or similar AI software could cause discrimination in other cases.

Conclusions

The case sits inside a well-established evidence base — echoing Buolamwini and Gebru's Gender Shades research, with one Dutch report citing roughly 30% lower recognition accuracy on darker skin tones — and stands as a cautionary example of an AI tool reaching national-university scale before its bias risks and complaint-handling procedures were adequately tested or governed.

Implementation

Indicative cost
Medium (€50k–€500k)
Time to results
Medium (1–3 years)

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful