Two experiments from the same year, both well-powered, both asking whether AI makes people worse at thinking, reach opposite conclusions.
In one, writers who practiced with an AI tool improved more on a later writing test than writers who practiced without it — and more than writers who got personalized feedback from professional editors. That is Lira, Rogers, Goldstein, Ungar, and Duckworth, in a manuscript compiled in August 2026, whose abstract poses the question as bluntly as anyone has: does using AI make you stupid?
In the other, people solving logic puzzles who requested AI help performed worse once the help was taken away. Their unassisted performance was overestimated when predicted from their earlier AI-assisted performance. That is Wu, Belem, Fu, Steyvers, and Smyth, published at HCOMP '26, whose Bayesian latent ability model found that greater independent problem-solving effort predicted larger gains.
Substitution versus demonstration. Same technology, opposite consequences, and the difference lives in the task design, not the model.
The lazy resolution is that one study is wrong. The better one is that the two studies were not measuring the same thing the AI was doing.
In the writing study, merely being shown an AI-generated cover letter — with no practice opportunity at all — improved skill as much as practicing with the original tool. That tells you where the learning came from: the AI functioned as a high-quality example. In the logic puzzles, the AI supplied the reasoning step itself, and the reasoning step was the skill being trained.
Substitution versus demonstration. Same technology, opposite consequences, and the difference lives in the task design, not the model.
A third experiment complicates the picture further. Karny, Mayer, Ayoub and colleagues ran a dynamic decision-making game with four conditions varying when AI assistance appeared. Immediate performance improved, and they found no skill degradation or carryover effects — evidence that there are classes of tasks where assistance can be provided without risking user autonomy.
Wu's team also found that lower-cost assistance produced more frequent AI use. How much friction you build into the request button is a real variable, and cost is its most honest proxy.
Then there is the part users cannot feel. Shaw and Nave, at Wharton, ran three preregistered experiments using an adapted Cognitive Reflection Test — 1,372 participants, 9,593 trials — randomizing AI accuracy through hidden seed prompts. Participants could not see which version they had.
Accuracy rose 25 percentage points above the no-AI baseline when the assistant was accurate. It fell 15 points below baseline when the assistant erred. A 40-point spread, driven entirely by information participants had no access to. Effect size, Cohen's h = 0.81.
Confidence went up either way. Including after errors. Their write-up is candid that the paper remains a preprint, posted to SSRN in February 2026 and unreviewed as of August 2026, so hold the theory loosely. The behavioral signature is the part worth keeping.
The obvious fixes do not work. Time pressure in Study 2 and per-item incentives with feedback in Study 3 moved baseline performance but did not break the pattern. Paying people to be careful and telling them how they did was insufficient against a confident wrong answer.
Higher trust in AI, lower need for cognition, and lower fluid intelligence predicted more surrender. That reads easily as a story about gullible people. It more plausibly describes a disposition to check that varies across individuals and that nobody screens for when handing someone a tool.
Self-report is a poor instrument here, for a reason Skulmowski documents under two names. The AI-based placebo effect: users overestimate their own ability while or after being assisted. The ghostwriter effect: users decline to disclose AI involvement even when they know the AI contributed. Wu's overestimation finding — that post-AI performance is inflated when predicted from AI-assisted performance — is the same shape.
Confidence, in other words, carries no diagnostic information about whether the skill is still there. It may carry negative information.
The measurement problem runs deeper. Biswas, Erlei, and Gadiraju reanalyzed a human-AI decision dataset — 240 participants, 7,200 trials, three task domains — and found that 67.3% of trial-level belief changes were exactly zero, with 76.4% smaller than five percentage points.
They call the decomposition a belief update gate: separate whether a report moves at all from how much it moves when it does. That distinction changes the story. The pooled within-trajectory slope of 0.494 rises to 0.949 among records with nonzero movement.
That second number is not evidence of near-Bayesian learning, and the authors say so, because it conditions on movement having happened. Nor can observed non-movement distinguish genuine belief inertia from small unexpressed updates or plain rounding. "People are stubborn with AI feedback" and "people updated a lot, quietly" produce identical data.
Konigsberg's construct of cognitive sovereignty names what is at stake across all of this: the ability to notice when your thinking is being displaced, to stay connected to how your beliefs and judgments actually formed, and to tell genuine reasoning apart from the feeling of having reasoned. Critical thinking, metacognition, intellectual autonomy, and epistemic agency each touch part of it. None covers it alone.
Guingrich, Mehta, and Bhatt put the mechanism in one phrase — the feeling of knowing without the labor of judgment. They also note the asymmetry: in documented cases, people are more likely to seek out and act on an LLM's suggestion than on the same suggestion from another person, even an expert.
Ownership has been examined too. Bok, Byon, and Pedersen's work on psychological ownership and AI decision-making, in Sport, Business and Management, takes up who feels the choice was theirs once an assistant is in the loop.
Practical implications, then. Make assistance cost something, because request cost reliably changes request frequency. Design tools to demonstrate rather than answer, since the example-condition writers learned as much as the practice-condition writers. Test for transfer with the AI removed, because that is the only measurement that distinguishes skill from performance.
And treat user confidence as noise. It rose in people who were being led to wrong answers, and it rose after they arrived.