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

Abstract: People are increasingly turning to AI assistance for simple tasks, e.g., arithmetic, spell-check, and answering simple questions. But does AI assistance actually save users time and effort? We investigate people's propensity to use AI for cognitively simple tasks and assess whether their reliance is well-calibrated. Across three pre-registered user studies (N = 2691), we find that people frequently choose to use AI even when doing so is inefficient (i.e. provides no meaningful time or effort savings). We identify systematic miscalibration at two levels: (1) a self-estimate miscalibration where people on average believe that they are using AI less than they actually are, and (2) efficiency-gain illusions where people overestimate how much time and effort savings AI use affords. We also identify a session-level carryover effect where a participant's prior AI use leads to further AI adoption and entrenches their miscalibration about time savings. Our results shed light on the mechanisms and biases underlying people's choice of whether to use AI as well as the risk of an overreliance feedback loop.

• The paper demonstrates a 14-percentage point gap between predicted (33%) and actual (47%) AI usage, revealing significant miscalibration.
• It finds that anticipated efficiency gains (55.7 seconds saved) are illusory as actual gains are minimal (7.5 seconds) and interface friction may add time.
• Exposure to AI assistance increases subsequent reliance, driving a feedback loop that could lead to cognitive deskilling and automation bias.

The Efficiency-Gain Illusion in AI-Assisted Simple Tasks: Calibration, Usage, and Behavioral Feedback

Study Motivation and Design

The paper "The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks" (2605.22687) investigates metacognitive miscalibration and behavioral biases in AI-assisted workflows for simple, cognitively elementary tasks. The premise is that the public discourse posits LLMs and AI assistants as productivity amplifiers, yet empirical evidence demonstrates that such assistance may not yield actual gains in time or effort—potentially leading to both inefficiency and cognitive deskilling.

To empirically dissect these issues, the researchers conducted three large-scale, pre-registered studies with $N=2691$ participants. Tasks were drawn from a taxonomy spanning four cognitive categories: information seeking, information processing/synthesis, procedural guidance/execution, and content creation/transformation, with easy and difficult variants per task type. The study designs systematically measured both predicted and behavioral AI usage, as well as subjective and objective efficiency outcomes.

Figure 1: Study overview and task taxonomy; measurement contrasts for AI-assisted versus independent modes across three studies.

Self-Estimate Miscalibration: Underestimating AI Use

A core finding is that individuals exhibit robust miscalibration regarding their propensity to use AI assistance.

• The population-level predicted rate of AI usage across tasks was 33%, yet the observed behavioral usage rate was 47%, a statistically significant gap of 14 percentage points ($p<0.001$).
• The miscalibration was more pronounced for easy tasks (predicted: 20%; actual: 38%; difference: 18 percentage points).
• The actual usage rate exceeded predictions even for tasks where almost no one anticipated delegating to AI, e.g., basic arithmetic, recall, or simple reasoning.

  Figure 2: Behavioral AI usage rates exceed predicted rates for both easy and difficult tasks, with pronounced gaps for trivial items.

This demonstrates a population-level underestimation of reliance on external cognitive support, revealing poor metacognitive monitoring of assistive technology usage, consistent with the planning fallacy and biases in self-assessment literature.

Efficiency-Gain Illusions: Overestimating Time and Effort Savings

Contrary to the prevailing narrative, the studies show that AI adoption for simple tasks does not confer meaningful gains in time or effort.

• Participants predicted that AI assistance would decrease completion time by 55.7 seconds; in reality, the actual savings were only 7.5 seconds on average.
• For easy tasks, AI use often incurred additional time due to chat interface friction—specifically, prompting and processing time dominated any minimal gains from task automation.
• Participants were well-calibrated regarding independent task completion times (predicted: 99 s; actual: 93.7 s), but markedly overestimated AI-assisted completion speed (predicted: 43.3 s; actual: 86.2 s).
• Effort (NASA-TLX) was also over-predicted for independent completion but well-calibrated for AI-assisted modes (predicted: 2.66; actual: 2.36 independent; predicted: 1.76; actual: 1.74 AI-assisted).

  Figure 3: Contrasts of predicted versus actual time and effort for independent and AI-assisted conditions; time savings from AI are illusory, especially for simple tasks.

This establishes the "speedup illusion" and "offloading illusion," wherein the perceived resource savings from AI delegation exceed the realized gains—sometimes inverting the efficiency rationale altogether when interface friction outweighs computational speed.

Carryover Effects: Behavioral Feedback and Entrenchment of Miscalibration

Exposure to AI-assisted completion in prior tasks increases both the likelihood of subsequent AI adoption and the degree of miscalibration regarding AI's efficiency benefits.

• Participants who completed tasks with AI in the exposure phase had a subsequent AI usage rate of 44.5%, compared to 27.7% among those exposed to independent completion.
• AI-exposed participants showed greater agreement with the speedup illusion, operationalized by reverse-coded ratings; prior exposure entrenched distorted time savings beliefs and induced an overreliance feedback loop.
• The effects were robust across easy and difficult task variants and persisted across both first and second test trials.
• Independent completion exposure did not significantly reduce subsequent AI usage rates compared to baseline controls.

Figure 4: Exposure to AI in early task phases increases subsequent AI adoption and amplifies miscalibration regarding completion speed.

This demonstrates a behavioral bias akin to inertia and habit formation, wherein assistive technology usage cascades and entrenches distorted resource rationality judgments.

Task-Level Effects and Individual Variation

Analysis at the task granularity reveals that:

• For many easy variants, the predicted AI usage rates were near zero, yet actual adoption was substantial (see Figure 5).
• Across participants, distinct clusters emerge: some never use AI, others always do, and a subset selectively delegates based on perceived difficulty (Figure 6).
• AI assistance occasionally reduces completion time for more complex variants (e.g., multi-item addition or multi-sentence writing tasks), but for trivially simple items it generally increases completion time.
• Across nearly all task types, AI assistance reduces subjective effort, but this reduction is modest and often overestimated in prediction.

  Figure 5: Actual AI usage rates exceed predicted rates for most easy task variants.

  

  Figure 6: Individual participant clustering in AI usage across task difficulty; selective and habitual use patterns.

  

Figure 7: Task-level analysis showing heterogeneity in time and effort differentials between AI-assisted and independent conditions.

Practical and Theoretical Implications

The empirical findings challenge prevailing assumptions about the utility of LLM-based AI in everyday workflows, particularly for simple, rapidly solvable tasks. The main implications include:

• High rates of unproductive delegation may lead to cognitive deskilling, reduced confidence, impaired independent performance, weakened knowledge retention, and diminished learning capacity, effecting a spectrum of individual and societal risks.
• AI-specific miscalibration is distinct from generic delegation errors—participants expect more from AI than from human assistance, driving irrational offloading behaviors.
• Behavioral feedback loops may develop, wherein initial exposure further entrenches miscalibrated beliefs and increases subsequent reliance, risking automation bias and overreliance.

The study recommends interventions at the level of belief calibration, rather than surface-level exposure controls, to mitigate inefficient AI usage. Such interventions may include metacognitive prompts, feedback on actual efficiency outcomes, or friction-modulation mechanisms in user interfaces.

Future Directions and Limitations

Key limitations for further study include:

• The focus on tasks solvable within five minutes constrains generalizability to more complex workflows; subsequent research should delineate the threshold complexity at which AI assistance yields net resource savings.
• AI capabilities are dynamically evolving; miscalibration patterns may shift over time and require longitudinal investigation.
• Motivation and incentive structures may influence completion behaviors in crowdworking settings, necessitating context-sensitive analysis.
• Within-subject metacognitive monitoring metrics were not directly computed; future work should leverage longitudinal within-subject designs and direct metacognitive self-reports.

Conclusion

This paper rigorously demonstrates that users underestimate their reliance on AI and overestimate the associated efficiency gains for simple cognitive tasks. Behavioral feedback loops exacerbate miscalibration and overreliance, with practical consequences for skill retention and cognitive agency. Addressing miscalibration at the belief level is critical for optimizing human-AI interaction paradigms, and ongoing research is needed to examine how calibration, experience, and dynamic AI capabilities interact to shape decision making in assistive workflows.