EPIQ Foundation
Back to the research library

Autonomy & decisions·Overview

AI Assistance Reduces Persistence and Hurts Independent Performance

Liu, Christian, Dumbalska, Bakker, Dubey — CMU, Oxford, MIT, UCLA; preprint arXiv, kwiecień 2026

Ten minutes of help — and you stop trying on your own.

A team from Carnegie Mellon, Oxford, MIT and UCLA pooled three randomised experiments with 1,222 adults. The tasks were deliberately elementary: fifteen fraction problems of rising difficulty and a reading-comprehension battery. Half the participants worked alone; the other half had a GPT-5 assistant in a side panel, preloaded with each solution.

The key move: for the final tasks the assistant vanished without warning. The assisted group then solved fewer problems than the control group and abandoned the task far more often. The effect appeared after roughly ten minutes of exposure and replicated across all three trials, reading included.

This is causal evidence, not a correlation — and it concerns not only skill but the willingness to make an effort. The help did not move competence onto the tool; it weakened the will to try alone. Caveat: a preprint, not yet peer-reviewed; short laboratory tasks. For personal use the lesson is still hard to wave away: cognitive debt is taken on in a quarter of an hour.

From the original

Although AI assistance improves performance in the short-term, people perform significantly worse without AI and are more likely to give up.

Although AI assistance improves performance in the short-term, people perform significantly worse without AI and are more likely to give up.

Liu i in., CMU / Oxford / MIT / UCLA 2026
Year
2026
Original language
EN
Source
Liu, Christian, Dumbalska, Bakker, Dubey — CMU, Oxford, MIT, UCLA; preprint arXiv, kwiecień 2026

An overview in the EPIQ team’s own words, not a translation. The full text remains with the publisher — link below. Quotations are short, verbatim and attributed, within the right of quotation.

AI Assistance Reduces Persistence and Hurts Independent Performance · EPIQ Foundation