Learning as a Feedback Loop (and the AI Partner)

Education is, at heart, a control system. A signal goes in, something is processed, output is produced — and then the critical step: the output is compared against a reference, and the error drives the next iteration. Run that loop open — input without measuring the error — and no stable learning occurs, in machines or in students. AI tools have changed the dynamics of this loop profoundly, and whether they accelerate your learning or quietly destroy it depends entirely on which part of the loop you hand over.

The Cybernetic View of Learning

• Input: lectures, documentation, worked examples, other people's code.
• Process: the attempt to compile the concept mentally — building the schema that predicts how the thing behaves.
• Output: code that runs (or doesn't), an argument that holds (or doesn't), a design that survives contact with requirements (or doesn't).
• Feedback — the load-bearing step: compiler errors, test failures, peer critique, marking, and honest self-assessment. This is where the error signal is generated, and the error signal is what learning is. No comparison, no correction; no correction, no learning.

The Open-Loop Trap

Now insert a generative model into the diagram. If the AI produces both the code and the explanation and the assessment of whether it's good, the student becomes a passive conduit between the prompt box and the submission box. The loop still spins — but it spins inside the model, not inside the learner. The "productive struggle" — the effortful, slightly painful comparison of your prediction against reality — is precisely the step that writes long-term memory and builds mental schemas, and it has been bypassed. The output may be excellent; the learning is zero. This is the mechanism behind the skill-atrophy concerns explored at length in Combatting Academic Misconduct Stemming from Misuse of GenAI, and it is why institutional policy is converging on "AI may assist the input, never replace the evaluation" (see GenAI Needs in Higher Education).

The trap has a tell: if you cannot predict what the AI's output will do before you run it, you are outside the loop.

Prompting for Growth: A Computer Scientist's Guide

The fix is not abstinence; it is keeping the evaluation step. Three patterns that preserve the loop:

• Don't ask for answers; ask for code reviews. Write the code yourself, get it working, then paste it and ask: "What are the edge cases or structural flaws here?" You did the processing; the model sharpens the feedback signal. This is the loop improved, not bypassed.
• Use the model as an interactive rubber duck. Explain your architectural design to it and ask it to find holes in your logic. The act of articulating forces the mental compile; the model's probing is bonus feedback. (The classic rubber duck never talked back; this one does, which is mostly an upgrade.)
• Test-driven learning. Write the unit tests for a concept before asking the AI to help implement a helper function. The tests are your prediction — your stake in the ground about how the thing should behave. When the AI's implementation meets your tests, you have verified its output against your understanding, which keeps you the architect (see TDD & BDD, and the guardrails in AI-Assisted Development).

A Self-Check

Score yourself honestly. Every "loop open" answer is a place where a future exam, interview, or production incident will present the invoice.