Novice programmers using unrestricted AI coding assistants produce code at the same rate as scaffolded users. The productivity numbers look identical. But when the AI is removed and the programmers must maintain what they built, unrestricted users fail 77% of the time. Scaffolded users fail 39%.
Kazemitabaar and colleagues (arXiv:2602.20206) call this epistemic debt — the gap between apparent competence (code works, produced on time) and actual understanding (can the programmer explain, modify, and debug it). The debt accumulates silently during AI-assisted work and comes due when the AI is unavailable.
The mechanism is the distinction between cognitive offloading and cognitive outsourcing. Offloading delegates tasks that don't contribute to learning — formatting, boilerplate, syntax lookup. Outsourcing delegates tasks that do — design decisions, algorithm selection, error reasoning. Unrestricted AI encourages outsourcing because the path of least resistance is to accept the generated code without engaging with why it works.
The fix is metacognitive friction: an “Explanation Gate” that requires the learner to teach back the concept before integrating the generated code. The learner must explain what the code does and why it solves the problem. This converts outsourcing back to offloading — the AI still generates the code, but the learner must understand it to proceed.
The scaffolded group's 39% failure rate is not zero. Metacognitive friction reduces epistemic debt; it doesn't eliminate it. The debt has a floor set by the difficulty of the maintenance task itself.
The general observation: tools that increase visible productivity while decreasing invisible understanding create debt that compounds. The debt is invisible because the metric (productivity) doesn't measure the liability (comprehension). Friction — slowing down the tool at the point where understanding is required — converts debt into equity.