When universities first confronted Generative AI, the immediate institutional reflex was to draft policy documents filled with rigid “Do’s and Don’ts.”
Do use it to fix grammar. Don’t use it to write essays. Do cite your prompts. Don’t let it touch your code.
This approach fundamentally misunderstands the nature of the technology. A “Do’s and Don’ts” list treats Generative AI like a financial calculator, a tool with fixed, predictable inputs and outputs. But AI is not a calculator. It is an active cognitive collaborator. Primitive rulebooks become obsolete within months, leaving both faculty and students trapped between defensive compliance and underground usage.
To lead a university-wide transformation, we must replace static rulebooks with an operational pedagogical framework. We need a clear, structural workflow that defines how humans and algorithms co-create knowledge: The Sandwich Model.
The Architecture of the Sandwich
The Sandwich Model divides any academic or engineering assignment into three distinct layers, placing human agency explicitly at the beginning and the end of the process.
1. The Top Slice: Human Intention (Framing the Problem)
Before a student types a single character into a prompt window, they must establish the problem space.
Generative AI is a mirror; if you feed it a shallow, ungrounded prompt, it returns a shallow, generic answer. The “Top Slice” requires the student to define the objectives, set boundary conditions, select the governing theoretical frameworks, and formulate a targeted hypothesis. If a student does not understand the underlying fundamentals of freight logistics, systems engineering, or economic theory, they cannot write a prompt that yields meaningful results.
The machine does not initiate the inquiry. Human intention frames the problem.
2. The Filling: AI Generation (Absorbing the Heavy Lifting)
Once the problem is rigorously framed, the AI handles the computational and generative heavy lifting.
This is the middle layer where the algorithm acts as an accelerator. It generates multiple code variations, drafts preliminary summaries, runs iterative scenario models, or highlights connections across vast datasets. Tasks that once took hours of mechanical labor are executed in seconds.
Crucially, this phase is treated not as the final product, but as raw cognitive material.
3. The Bottom Slice: Human Verification (Critical Judgment & Ethics)
The most important layer of the model and where authentic learning is evaluated is the bottom slice. The student steps back into the loop as the lead auditor and decision-maker.
Every output generated in the middle phase must be interrogated:
- Accuracy: Where are the synthetic hallucinations or faulty mathematical assumptions?
- Logic: Does the code handle rare edge cases, or does it collapse under real-world constraints?
- Taste and Synthesis: Is the argument compelling, or is it merely plausible-sounding fluff?
- Ethics: What hidden biases or systemic risks are embedded in the algorithm’s recommendations?
If a student cannot defend why an AI-generated solution works, identify where it breaks, and justify its ethical implications, they have not completed the assignment.
Literacy Over Automation
There is a persistent misunderstanding in policy debates that integrating AI means trying to automate 80% of our educational programs.
That is entirely the wrong metric. The objective is not to maximize the percentage of work done by machines, but to achieve 100% AI literacy and critical awareness across our entire student body and faculty.
The Sandwich Model changes the goalposts of assessment. We stop grading students on their ability to perform routine, low-level execution. The machine has commoditized that. Instead, we grade them on their Top Slice (the clarity and theoretical depth of their framing) and their Bottom Slice (their capacity for verification, critical taste, and strategic judgment).
By embedding this framework into our curricula, we might ensure that AI does not dull human intellect. It elevates our students from passive generators of routine work into confident, critical managers of intelligent systems.
Tom Van Woensel 