All insights
Focus6 min read

Fifteen Minutes to Give Up

Vishal Sachar

Co-Founder & CEO of CLRT

You rolled the assistant out to everyone, and the numbers you chose to watch went the right way. Drafts arrive sooner, tickets close faster, and the adoption chart climbs every month. Nothing on that dashboard is false. It is simply measuring one condition, your people working with the assistant switched on, and it has nothing to say about the other condition, the moment the assistant cannot help: the problem it has no context for, the meeting where the answer has to come from someone's own head, the decision nobody has framed yet. A run of randomised experiments, in a preprint revised this week, measured that second condition directly. The finding is not only that people got more answers wrong. It is that, after a single short session of help, they gave up sooner.

0.711
Share of unaided test problems solved after a short session of AI help, against 0.77 for controls, in the 585-person replication built to rule out an earlier confound (arXiv preprint)
Liu et al., arXiv, 2026
0.082
Share of reading test problems the helped group skipped once the assistant was removed, against 0.01 for controls (P = 0.008), where skipping carried no penalty (arXiv preprint)
Liu et al., arXiv, 2026
61%3
Of the helped group in the arXiv preprint's replication who said they mainly took answers directly, the group that later solved 0.65 against 0.77 for controls
Liu et al., arXiv, 2026

The authors are unusually candid about their own design. Researchers at Carnegie Mellon, UC Berkeley, Oxford, MIT and UCLA recruited adults online, gave them fraction problems and later reading comprehension questions, and randomised each person into one of two conditions. One group worked with an AI assistant in a sidebar that already held the answers and handed them over on request. The other worked alone. After a run of practice problems the assistant was removed without warning and everyone faced the same short test. In the first experiment, of 305 people, the helped group solved 0.57 of the test problems against 0.74 for controls, but the authors note that uneven exclusions may have inflated that first gap, and built the replication to remove it. With that confound addressed, 585 people still showed a significant gap, 0.71 against 0.77, and a reading-comprehension experiment of 168 people using SAT practice questions found 0.76 against 0.89. In the authors' words, the effects emerged after a ten to fifteen minute session.

FIG. 01Unaided solve rates fell in all three randomised experiments after a single short session of AI help; the authors flag the first experiment's gap as possibly inflated by uneven exclusions. Liu et al., arXiv preprint v6, 13 October 2026.
01Persistence, not accuracy

The solve rates are the less interesting half. Every participant had a skip button, and every participant was told plainly that a wrong answer cost nothing and that pay did not depend on the score. Under those conditions, skipping is not a judgement about difficulty. It is a decision not to try. In the first experiment the helped group skipped 0.20 of the test problems against 0.11 for controls, and in the reading experiment, where answers given in under five seconds also counted as skips, 0.08 against 0.01. In the larger replication the skip gap pointed the same way but was not statistically significant, and it is worth saying so, because the authors do. What the experiments isolate is not a loss of knowledge, since nobody forgets fractions in a quarter of an hour. It is a loss of willingness to sit with a problem that does not yield straight away, and that willingness is the thing every organisation quietly depends on.

FIG. 02Skipping carried no penalty, yet the helped group skipped more test problems; the gap was significant in experiments 1 and 3 and not in the replication, and in the reading test, answers given in under five seconds also counted as skips. Liu et al., arXiv preprint v6.
02What speed hides

That is why this belongs under focus rather than training. Deep work, the kind a leader is paid for, is mostly the act of staying with an unresolved problem past the point where it is comfortable: the pricing question with no clean answer, the restructuring with three bad options, the strategy memo that will not come right on the third draft. Persistence is the raw material of that work, and the authors note it is among the strongest predictors of long-term learning. An always-on assistant is built for the opposite instinct. It removes the discomfort the moment it appears, which is exactly what makes it feel productive, and exactly what a speed metric rewards. A rollout measured only on speed is therefore counting the benefit and structurally blind to the cost, because the cost only appears in the hours when the assistant cannot help.

03Answers versus hints

The more useful finding sits inside the replication. Asked afterwards how they had used the assistant, 61% of the helped group said they had mainly taken answers directly. Their later solve rate was 0.65 against 0.77 for people who never had help. The 27% who had used it for hints or clarification solved 0.76, almost exactly the controls' rate, and skipped less often than the answer takers. The authors are careful that this split is self-reported and cross-sectional, so it shows an association rather than a cause, and anyone citing it should be equally careful. But the pattern is consistent with everything else in the paper, and it moves the question somewhere more practical. The assistant did not erode persistence by existing. It did so when it was pointed at the struggle itself, used to remove the effort rather than to direct it.

FIG. 03In the replication, people who took answers directly fell behind controls while hint users did not; the split is self-reported and shows association, not cause. Liu et al., arXiv preprint v6.

This is the part a rollout plan almost never contains. Giving every employee the same assistant with the same defaults, and then measuring turnaround, is a decision to let the most convenient use win, and the most convenient use is asking for the answer. For a great deal of work that is exactly right: nobody needs to persist through reformatting a deck. For some work it is the opposite of right, because the effort is the point, whether that is a junior building the judgement your seniors will one day rely on, or a leader thinking through a problem with no precedent. The decision about which is which cannot be left to each person at the moment of temptation. It is a design decision about where to point the tool: where it should answer, where it should only prompt, and where it should stay out of the room.

Every dashboard measures your people with the assistant switched on. The cost appears in the hour it cannot help.

A deeper dive

The authors' proposed mechanism is worth taking seriously because it explains why the effect appears so fast. Their argument is that an assistant which always answers conditions people to expect an immediate answer, and so denies them the experience of working through a difficulty on their own. One way to read that is as calibration: each time effort pays off, the point at which a problem feels worth abandoning moves a little further out, and when the effort is removed, that point moves back in, apparently within a single sitting. The caveats are real, and the paper states them. These were online participants doing fractions and reading questions, not professionals doing their own work. They were tested immediately after the assistant was removed, so nobody yet knows how long the effect lasts, or whether it compounds or fades with months of daily use. The paper is an arXiv preprint, revised on 13 October, that the authors mark as published at COLM 2026. A separate three-month field trial of 133 patent lawyers, an unreviewed working paper conducted and funded by Google, complicates the picture from the professional end: after three months of AI-assisted drafting, the treated lawyers were better at unaided judgement overall, but the whole gain went to seniors; juniors showed no average gain, their scores splitting into more poor and more good.

The second-order trap is that the obvious fix does not scale. The hint users suggest a better pattern exists, and the authors call it worth pursuing, but they also describe user-facing remedies, such as Socratic prompting or limited use time, as "band-aids", because the assistant offers a temptation to offload at scale. An organisation that responds by asking people to use AI more thoughtfully has handed the design to the moment of least resistance. The other trap is in measurement. Every metric a rollout reports, adoption, time saved, tickets closed, is taken with the tool switched on, so the dashboard will look healthiest precisely where persistence is eroding fastest. The only honest instrument is to observe some work with the assistant switched off, deliberately and periodically, on the tasks where judgement is supposed to be growing. Almost nobody does this, because it looks like a step backwards. It is the only way to know whether the people you pay for judgement still have the habit of producing it, or have quietly learned to wait for the answer.

Work with CLRT

This is the work CLRT does. We do not hand everyone the same assistant and report the time saved. We map where AI should answer, where it should only prompt, and where the effort has to stay human because the effort is how your people become the judgement you rely on, and we design the measurement that tells you whether that judgement is growing or quietly thinning. If you want to see where your own time and your team's should be pointed first, start with CLRT Ascent at ascent.clrtstudio.com, then bring us the work you cannot afford your people to stop struggling with.

Vishal Sachar

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.

Start here

Skip the reading. See where your leverage leaks.

Ascent is our free diagnostic. Ten minutes, and you have the one workflow worth building first.