The give-up reflex: what ten minutes of AI help does to your persistence

Overview

This is about a quieter cost of reaching for AI than a wrong answer: what waiting for the answer does to you. A randomised trial published in April 2026 found that about ten minutes with an AI assistant was enough to leave people solving fewer problems on their own afterwards — and, tellingly, giving up on them sooner. The worry most people name is long-term skill fade. The finding is that the reflex sets in almost immediately. The point you’ll be able to act on: a small change to how you start a hard task that keeps the reflex from forming.

The content

Here’s the obvious read: over-reliance is a slow burn — lean on AI for months or years, and your own edge quietly dulls. That worry is real, and 2026 has plenty of it; a June Nature feature rounds up the early evidence of AI “deskilling” across medicine and software. But the sharper finding is how fast it starts. In a set of randomised controlled trials (N = 1,222 across three experiments, on maths problems and reading comprehension), people using an AI assistant did better with it — and then, when it was taken away and they worked alone, did worse than people who’d never had it. In the first experiment they solved fewer of the same problems (0.57 versus 0.73) and, the part that matters here, skipped more of them: a skip rate of 0.20 against 0.11 for the unassisted group — roughly double. They gave up sooner, and the direction held, though more weakly, when the researchers ran it again. Call it the give-up reflex.

The mechanism the authors name is the whole lesson: AI “conditions people to expect immediate answers, thereby denying them the experience of working through challenges on their own.” Sit with a hard problem and you build a tolerance for being stuck; hand it straight to a model and you don’t. That tolerance isn’t a soft virtue — persistence is one of the strongest predictors of how well people learn. So the trade isn’t just this problem’s answer for a few saved minutes. It’s a rep you needed, quietly skipped.

One honest limit before you generalise: these were short reasoning tasks in a lab, not your quarterly analysis or a contract you’re marking up, and the exact numbers won’t transfer to your desk. What transfers is the mechanism. The tool on your desk is the same one training you to expect an answer the moment you feel stuck — which means the fix isn’t willpower or going cold turkey. It’s changing the order. Put your own attempt before the model’s: a few minutes forming a rough answer, or a precise account of where you’re stuck, does two things at once. It keeps the persistence rep, and it hands you your own draft to judge the AI’s answer against — instead of anchoring on whatever it says first.

Try it

On the next genuinely hard task today — the kind where your own judgement is the point, not a formatting chore — set a timer for five to ten minutes and attempt it before you open the AI. Don’t try to finish. Try to get far enough to hold a position: your best guess at the answer, the approach you’d take, or the exact spot you’re stuck and why. Then bring the model in, and read its response against what you already had.

Where it breaks: don’t do this for routine or low-stakes work — boilerplate, a first-pass email, a reformat — where there’s no hard problem to persist at and the timer is just friction. The reflex only costs you on the problems where the struggle was the point. If you want to know how much you’ve been offloading without noticing, a May 2026 study built exactly that measure and found people can’t feel it — usage counts and gut sense both miss it, and reliance jumps when you’re rushed.

Additional reading

• AI Assistance Reduces Persistence and Hurts Independent Performance (Liu et al.) (6 April 2026) — the randomised trials behind this module: N = 1,222 across three experiments, the ~10-minute onset, the lower solve rate and higher skip rate once AI was removed, and the “conditions people to expect immediate answers” mechanism. Note the tasks were maths and reading, not professional knowledge work.
• Is AI ruining our skills? Early results are in — and they’re not good (Lenharo, Nature) (18 June 2026) — a news feature surveying the 2026 “deskilling” evidence, including clinicians whose own detection skills declined after getting used to an AI aid. Wider context, not a single study.
• Offloading Score: Measuring AI Reliance Through Counterfactual Workflows (Padmakumar et al.) (28 May 2026) — a controlled study (n = 40) showing reliance rises under time pressure (+43%) and that usage counts and self-report don’t detect it; the source for the “you can’t feel it” line in Try it.

Editor’s note

I really force myself to resist the urge sometimes to do this with contracts. Where there’s a clause I don’t immediately follow, I’m tempted to hand it straight to the model. I have to catch myself, stop, and form my own read first, even a rough one. A few minutes of my own effort is what lets me tell whether an answer is any good after I’ve handed it to a model. Hold the time even when you’re rushed. I know it’s exactly when you want to skip it, but that’s the space where you build the judgement muscle.