The Widening Gap. Who AI carries, and who it carries away.

Under the surface, the class split in two. Eleven students showed no metacognitive difficulties: they read suggestions, kept the ones matching a plan they already had, and ignored the rest. Ten students struggled, and the researchers watched the assistant amplify the struggle rather than fix it: suggestions interrupted half-formed thoughts, plausible completions pulled them onto the wrong problem, and generated code let them advance without understanding what they were advancing through.

The acceptance numbers tell the same story softly. Students with difficulties accepted 34.1% of Copilot suggestions on average against 24.5% for those without, but the spreads overlap heavily and the paper reports no significance test on that gap, so treat it as texture. The sharper observation is what got accepted: three of the fastest finishers sat near 30% acceptance too, and the difference was that everything they took was something they already wanted.

Scope before you generalize: 21 novices, one institution, one 35-minute problem, and a course that had used AI tools from day one. This is a study of people still learning to program, watched closely, not a trial on professionals.

In 2024, with AI, 10 of 21 students struggled with the same problem, and 9 of those 10 reached a working solution anyway, by leaning on the tools. The researchers, watching their words and actions, judge that most of them would not have gotten there alone, and that most finished believing they understood more than they did. The abstract's phrase is 'an illusion of competence'; the conclusion says students already poised to succeed used GenAI to accelerate while struggling students were left with that illusion.

Hold the claim to its actual size. This is a comparison across two cohorts six years apart, not a randomized contrast, and the study measured no learning outcome; the illusion is an interpretation of think-aloud and eye-tracking data, which the paper itself concedes. What it documents cleanly is the disappearance of a signal: failure used to be legible to the person failing. Now the checkmark is green either way.

A 2026 comment in npj Artificial Intelligence gives that missing step a name: AI outputs are results, and they become responses only when a person evaluates them and takes responsibility for them (the authors call it the 3R principle). A framing piece with no data, but a precise one.

A companion study by an overlapping team puts the same shape on different instruments. Across a semester of an introductory course (54 enrolled, 40 analyzed), students with higher self-efficacy, lower fear of failure, or higher prior grades used AI less, and later in the problem-solving process. Heavier, earlier AI use went with lower programming self-efficacy (on a 1-to-7 scale) and higher fear of failure (on a 1-to-5 scale).

Two things keep this honest: the authors explicitly refuse the causal reading, since the traits that predict AI use also predict grades on their own, and self-regulation showed no relationship with AI use at all, a null that most summaries of this paper quietly drop. The authors' own headline is the optimistic one: at least some students use GenAI to support, not replace, their thinking.

The skill the accelerating students had is worth naming, and the paper reaches back to 1994 for the name. Marvin Minsky argued that expertise is substantially negative: knowing what not to do, what not to accept, what not to think. An expert has seen more bad ideas than a novice has seen ideas. Reading a suggestion and recognizing it as wrong in two seconds is negative expertise applied to code review, and it is precisely the asset a novice has not accumulated yet. The tool prices it in anyway.

The camps completed the same amount of assigned work. The difference showed up in what happened next: 44% of the taught-the-process campers went beyond the assignment and built more, against 17% of the control camp, over twice the self-initiated work (p = .033). The control camp's productivity tracked how often they called for help; the experimental camp's did not, which the authors read as independence. And self-efficacy, equal at the start, ended significantly higher in the experimental camp (p = .0005).

High schoolers at summer camps, web development, and the two camps were filled by program registration rather than researcher randomization, a confound the paper names itself. What it licenses is modest and useful: awareness of your own problem-solving process is teachable, and teaching it changes behavior. That is the mechanism every intervention in this literature leans on, ours included.

A CHI 2024 framework paper from Microsoft Research supplies the wider lens: working with generative AI is itself metacognitively expensive. Prompting, evaluating output, and deciding how much to trust it are monitoring-and-control acts, and a tool that raises those demands without supporting them is a usability problem. Cite it for the lens, not for an effect size; it is a position paper.

Learners' feel for what they know is informative and poorly calibrated: in the lab, retrieval practice with feedback left learners almost perfectly calibrated on a later test, while learners without feedback were overconfident.

The practical detail sits in the lucky guesses: answers that were correct but rated as guesses mostly evaporated without feedback, and mostly survived with it. In the one study that measured exactly this, feedback roughly doubled what remained of low-confidence correct answers. That is the 'got lucky' button doing its job: it tells the explanation where to work. Thirty people, general-knowledge facts, one lab.

Across many lab studies, errors made with high confidence are more likely to be fixed after feedback than errors made with low confidence, the opposite of what interference theories predicted; the effect is well replicated at short delays.

Two boundaries matter. High-confidence errors are rare, the review says so plainly, so this effect can only ever do a small share of a scheduler's work. And the fix decays: in a week-delay study, corrections fell from .86 of errors to .56, and the errors most likely to resurface were precisely the confident ones. A corrected error is not a closed file. It is the strongest argument in this literature for re-serving an item instead of retiring it.

Nobody has run the study our vote most directly implies: does adding a confidence rating after a retrieval attempt, with feedback held constant, improve delayed retention in adults on technical material? The adjacent evidence is mixed. Prompting judgments during study averages out to no effect in a meta-analysis. The largest realistic test of confidence prompts, on school assessments, found solid Bayesian evidence for no effect on concurrent scores.

What survives is the narrower claim above: the vote is information routing, not magic. It tells feedback where to work, and one more thing the literature does support: self-assessments are most accurate when made right after a real attempt, in the same format as the real test, which is exactly where ours sits.