evidoria

← Back to browse

Good practice Imported

GPT-4-Powered Formative Feedback for Large Accounting Classes

South Africa · Pretoria · See the South Africa profile · See the Pretoria profile

Evidence: Descriptive / self-reported Top 73% 40/100 · Ask Evidence Copilot about this practice

Pretoria accounting lecturers built a GPT-4 web tool giving essay-question feedback engineered around Nicol & Macfarlane-Dick's principles. A study of 75 scripts across five assessments over two years checked how well AI feedback matched those principles.

75 scripts
Answer scripts evaluated
15 scripts
Scripts per assessment
5 assessments
Assessments covered (two academic years)

Details

Maturity
Pilot
Promoter
University of Pretoria, Department of Accounting
Period
2022–2024
Keywords
Generative AI, formative feedback, higher education, accounting

Context

Giving individualised written feedback on discussion and essay-style questions is one of the hardest things to do well in a large, competency-based university course, where one lecturer may face hundreds of scripts per assessment. Accounting academics at the University of Pretoria built a no-code web application (on the Bubble.io platform) that routes students' submitted answers to GPT-4 and returns feedback engineered to follow Nicol and Macfarlane-Dick's widely cited seven principles of good feedback practice.

Objectives

The researchers set out to evaluate how consistently AI-generated feedback embodied the seven principles of good feedback practice, such as clarifying what a good answer looks like and helping students self-correct rather than simply supplying the right answer, rather than measuring downstream grade improvements.

Activities

The team drew a purposive sample of 75 answer scripts (15 per assessment) spanning five distinct accounting assessments across two academic years, chosen to cover a range of student performance levels and topics, and published the analysis in Assessment & Evaluation in Higher Education.

Results

Because the evaluation is a small purposive sample from a single course rather than a randomised trial or a multi-course rollout, the findings should be read as a proof-of-concept for prompt-engineering AI feedback to a pedagogical standard, not as settled evidence of impact on learning outcomes.

Implementation
Indicative cost
Low (< €50k)
Time to results
Short (< 1 year)
Staffing & skills
University of Pretoria Department of Accounting academic staff (tool design and prompt engineering)

Conditions for success

  • A no-code platform (Bubble.io) to keep development accessible to academic staff without dedicated engineering support
  • Explicit prompt engineering grounded in an established pedagogical feedback framework (Nicol and Macfarlane-Dick's seven principles)

Common failure modes

  • Evaluated on a single purposive sample within one course at one university; no evidence yet of use beyond this pilot

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.

Similar practices you may find useful