United KingdomEducation2023 – ongoingUniversity of Reading
GPT-4 audit of examination marking
Reading's examination audit found that 94% of AI submissions escaped detection
Problem
Unsupervised examinations allowed students to submit AI-written answers. Reading needed to test whether its marking procedures would detect them.
Approach
Researchers obtained university approval for a quality assurance exercise. They submitted GPT-4 answers through university accounts into real psychology examinations across 5 modules. Markers were unaware of the exercise.
Results
Takeaway
The audit exposed a weakness in existing assessment controls. It did not establish a reliable way to detect AI-written work.
Categories
- Industry
- Education
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Generative AI
- Adoption
- Off-the-shelf tool or platform
- Goal
- Quality and risk
- Evidence
- Independent study
Sources
Source summaries
Researchers tested GPT-4 submissions in Reading's normal examination workflow. They reported that 94% escaped detection and that AI answers received higher grades on average than real student submissions.
Full details are in the original sources.