The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks

Synthesis: Across three pre-registered studies (N=2,691), this paper documents systematic miscalibration in how people perceive their own AI usage. The authors find that people not only use AI for cognitively simple tasks even when it provides no meaningful efficiency benefit, but also systematically misperceive both how much they use AI and how much it helps them.

Two Forms of Miscalibration

1. Self-estimate miscalibration: People on average underestimate how often they actually use AI — they believe they use it significantly less than objective measures show.
2. Efficiency-gain illusion: People overestimate the time and effort savings that AI assistance provides, believing tasks are faster and easier with AI even when objective metrics show no difference.

The Overreliance Feedback Loop

Critically, the authors identify a session-level carryover effect: prior AI use in a session leads to further AI adoption, which in turn entrenches the miscalibration about time savings. This creates a self-reinforcing feedback loop — initial AI use begets more AI use, and the illusion of efficiency makes it harder for users to self-correct.

Connection to Over-Reliance

This study provides a cognitive mechanism for the Over-Reliance phenomenon: the efficiency-gain illusion explains why students and other users continue to reach for AI even when it demonstrably doesn't help — and in fact may harm learning outcomes, as documented in related work. The findings complement Evidence of a Cognitive Shift in AI Education: How Students Are Rethinking Human Intelligence? by identifying the perceptual biases that drive behavioral change.

What this means for practice

• Learners. Treat your own estimate of how much you use AI as unreliable evidence: participants predicted they would use AI in 33% of tasks but the actual rate was 47%, and they expected AI to save 55.7 seconds on a task where it saved 7.5 seconds.
• Learners. Judge AI by measured results rather than by how easy it felt to use, because prior AI use made participants more likely to opt for AI again (β = 0.54, p < 0.001) rather than better calibrated.
• Instructors. Log behavior instead of asking students to self-report AI use: the actual usage rate exceeded the stated rate by 14 percentage points, and on easy task variants there was an 18% difference between the 20% predicted rate and the 38% actual rate.
• Instructors. Set explicit expectations for simple tasks, since 21-54% of participants used AI on tasks where the average predicted rate was only 2.3%.
• Instructors. Teach metacognitive calibration about AI use as part of AI literacy: users need more than technical knowledge of AI capabilities and limitations, because simply telling people that AI may not save time does not correct an illusion that is perceptual and self-reinforcing.

Limitations

• Three pre-registered studies with N = 2691 Prolific participants used tasks designed to be completed in under 5 minutes, so the findings do not speak to cognitively demanding work.
• Study 1 and Study 2 rely on between-subjects aggregates rather than comparing the same individual's stated and actual AI use, and the prediction and completion samples were not completely disjoint.
• Study 3 measured AI adoption immediately after the exposure phase, limiting conclusions about longer-term behavioral change.
• The experiments did not directly control for participant motivation or incentives, and the binary AI-use variable is a rough proxy that does not capture how participants actually used AI.

Citation

Yu, S., Cheng, M., Jabbar, A., Sucholutsky, I., Collins, K. M., Jurafsky, D., & Hawkins, R. D. (2026). The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks.