Understanding the Rise of Generative AI in Higher Education Assessments
In the rapidly evolving landscape of higher education, generative artificial intelligence (GAI) tools like ChatGPT have transformed how students approach learning and assessments. Since its public release in late 2022, ChatGPT has demonstrated remarkable capabilities in generating human-like text, code, and even complex calculations, prompting educators worldwide to rethink traditional evaluation methods. In Singapore, a global leader in educational innovation, universities are at the forefront of integrating AI responsibly while safeguarding academic integrity.
Singapore Management University (SMU), one of the nation's top institutions, exemplifies this proactive stance. With its emphasis on experiential learning and industry relevance, SMU has developed comprehensive frameworks for GAI use, allowing students to leverage tools like ChatGPT for idea generation and research under strict guidelines. This approach aligns with national efforts, where the Ministry of Education (MOE) and institutions like National University of Singapore (NUS) and Nanyang Technological University (NTU) permit AI in assignments provided proper citation and transparency.
However, the challenge lies in assessments. Traditional exams and assignments, once reliable indicators of student mastery, are now vulnerable to AI assistance, raising concerns about critical thinking and originality. A 2025 Higher Education Policy Institute (HEPI) study revealed that 88% of students use GAI for assignments, underscoring the urgency for redesign.
Spotlight on SMU's Groundbreaking Research
Leading the charge is Dr. Michelle L. F. Cheong from SMU's School of Computing & Information Systems. Her seminal paper, "ChatGPT's Performance Evaluation in Spreadsheets Modelling to Inform Assessments Redesign," published in the Journal of Computer Assisted Learning on May 5, 2025 (DOI: 10.1111/jcal.70035), provides empirical evidence on ChatGPT's limitations.
Cheong's work is particularly relevant for Singapore's higher education sector, where quantitative skills are prized. SMU, known for its business-focused programs, integrates spreadsheets extensively, making this research directly applicable. The study not only benchmarks AI performance but offers actionable redesign strategies, positioning SMU as a thought leader.
Dr. Cheong emphasizes, "Educators need new information on how well ChatGPT performs to redesign future assessments in this new paradigm." Her findings reveal that while AI excels in routine tasks, it falters in higher-order thinking, empowering instructors to target those areas.
Methodology: Rigorous Testing with Bloom's Taxonomy
Cheong categorized assessment questions using the revised Bloom's Taxonomy, a framework outlining six cognitive levels: (1) Remembering, (2) Understanding, (3) Applying, (4) Analyzing, (5) Evaluating, and (6) Creating. Two original quizzes were developed:
- Financial calculations quiz: Scenarios like computing discounts and payments for buying computers.
- Monte Carlo simulation quiz: Estimating COVID-19 infection risks in queues using probabilistic modeling.
These multi-step, linked items mimic authentic coursework. ChatGPT was prompted in four engineering variations:
- Zero-Shot-Baseline: Direct question, no guidance.
- Zero-Shot-Chain-of-Thought (CoT): Adds "Let's think step by step."
- One-Shot: Includes one solved example.
- One-Shot-CoT: Example plus step-by-step reasoning.
Performance was scored objectively on numerical accuracy and logical correctness, repeated for consistency.
Key Findings: Where ChatGPT Succeeds and Fails
The results were telling. With baseline prompts, ChatGPT achieved high accuracy up to Level 3 (Applying), correctly handling basic formulas and data setup. Errors emerged at Level 4 (Analyzing), often miscalculating despite identifying the right method—e.g., wrong probability distributions in simulations.
Prompt enhancements extended capabilities: Zero-Shot-CoT reached Level 5 (Evaluating) reliably, while One-Shot variants boosted mid-levels. However, no configuration mastered Level 6 (Creating), where novel model design is required. Overall, accuracy dropped sharply with complexity, highlighting AI's brittleness in advanced reasoning.
| Bloom's Level | Baseline Accuracy | Best Prompt Accuracy | Example Error |
|---|---|---|---|
| 1-3 (Low-Mid) | High | High | Rare |
| 4 (Analyzing) | Low | Medium | Incorrect linkages |
| 5-6 (High) | Very Low | Low | Failed innovation |
These insights validate anecdotal educator concerns and provide data-driven baselines.
Redesign Strategies Tailored to Cognitive Levels
Cheong proposes level-specific redesigns to mitigate GAI while fostering skills:
- Levels 1-2: In-class ChatGPT sessions to generate answers, then critique errors. Students learn prompt crafting and verification.
- Levels 3-4: Collaborative peer projects, e.g., building EV adoption spreadsheets, emphasizing analysis where AI stumbles.
- Levels 5-6: AI-prohibited tasks focusing on creation/evaluation, like designing original simulations from real data.
SMU's Centre for Teaching Excellence (CTE) echoes this with AI-resistant practices: process-tracked drafts, personal reflections, in-class viva voce, and authentic tasks like mock client consultations.SMU CTE Guidelines
For spreadsheets specifically, require code explanations, error debugging, or integration with class-specific datasets.
Singapore's Broader Higher Education Landscape
Singapore universities report few AI misconduct cases—less than a handful at SMU over three years—thanks to transparent policies.
An OECD survey highlights Singapore teachers' high AI adoption—3 in 4 use it, double the global average—driving innovative pedagogies.
Explore academic career advice for roles in AI-enhanced teaching.
Stakeholder Perspectives: Educators, Students, and Industry
Educators praise Cheong's work for empowering redesign without bans. SMU students, surveyed informally, appreciate guided AI use for brainstorming but stress human oversight for depth. Industry partners value graduates skilled in AI collaboration, aligning with Singapore's Smart Nation initiative.
Challenges include equity—prompt engineering favors advanced users—and detection tool reliability. Solutions: Faculty training via SMU CTE workshops and student AI literacy modules.
Real-World Case Studies and Statistics
In Cheong's financial quiz, ChatGPT nailed discount calcs (Level 2) but bungled multi-scenario analysis (Level 4). The Monte Carlo task exposed flaws in risk modeling, mirroring real analytics pitfalls.
Stats: 92% of students use AI per HEPI; Singapore's low cheating (ST report) vs. global rises. Cheong's prior study on permitted ChatGPT assessments showed improved learning when integrated thoughtfully.
- AI boosts productivity but risks rote reliance.
- Redesigned assessments raise critical thinking by 20-30% in pilots.
Read the full paper for datasets.
Future Outlook: AI as Ally in Singapore Higher Ed
Looking ahead, Singapore's Research, Innovation and Enterprise 2030 (RIE2030) plan invests S$37 billion in AI, including education. SMU's research paves the way for hybrid models: AI for basics, humans for innovation.
Implications: Enhanced employability via AI fluency; global benchmarks for unis. Challenges like evolving AI (GPT-4o) require ongoing research.
Discover Singapore higher ed jobs or global opportunities.
Photo by - Landsmann - on Unsplash
Actionable Insights for Educators and Institutions
To implement:
- Audit assessments via Bloom's mapping.
- Pilot Cheong's prompts for baselines.
- Integrate AI literacy in curricula.
- Use tools like Peerceptiv for feedback.
Rate professors on Rate My Professor for AI-savvy insights. For career growth, check higher ed career advice.