💻 computer science

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

Through three pre-registered studies involving 2,691 participants, this paper reveals that people frequently and inefficiently rely on AI for simple tasks due to a dual miscalibration where they underestimate their own usage while overestimating the time and effort savings, a bias that is further reinforced by a session-level carryover effect.

Original authors: Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky, Robert D. Hawkins

Published 2026-05-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky, Robert D. Hawkins

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Magic Shortcut" That Isn't Magic

Imagine you are trying to cross a river. You see a bridge (doing the task yourself) and a magical flying carpet (using AI). Most people assume the flying carpet is always faster and easier. They think, "I'll hop on the carpet; it will save me time and energy!"

This paper is like a group of researchers who decided to actually time the journey. They found a surprising truth: On simple, short trips, the flying carpet often takes longer than walking, and people are completely fooled into thinking it's a shortcut.

The researchers call this the "Efficiency-Gain Illusion." It's a mental trap where we overestimate how much AI helps us and underestimate how often we actually use it, even when it's a bad idea.

The Three Experiments (The Story So Far)

The researchers ran three large studies with nearly 3,000 people to figure out why we keep hopping on the "flying carpet" even when it slows us down.

Study 1: The "I Won't Do It" Lie

The Setup: They asked people two things:

"If you had to do a simple task (like adding two numbers), would you use AI?"
Then, they actually made people do the tasks and watched what they did.

The Discovery: People are terrible at predicting their own habits.

The Prediction: People said, "Oh, I'd only use AI maybe 20% of the time for easy stuff. I can handle that myself."
The Reality: When actually doing the tasks, people used AI 38% of the time.
The Analogy: It's like someone saying, "I'm not a sugar addict, I'll only eat one cookie." But when the cookie jar is right there, they eat three. We claim we are independent, but when the AI button is right in front of us, we click it way more often than we admit we would.

Study 2: The "Speedometer" Glitch

The Setup: They asked people to guess how long a task would take with AI versus without it. Then, they timed the actual results.

The Discovery: People have a broken speedometer in their heads.

The Illusion: People predicted AI would save them about 56 seconds on a task.
The Reality: AI actually only saved them 7.5 seconds. In fact, for very simple tasks (like "What color is white mixed with black?"), using AI actually made the task slower by about 10 seconds.
Why? The "friction" of using AI is too high for simple jobs. You have to:
1. Type the prompt.
2. Wait for the AI to think.
3. Read the answer.
4. Copy and paste it.
- The Analogy: Imagine you need to walk to the mailbox next door. You decide to drive your car instead. You have to walk to the car, start the engine, drive one block, park, and walk back. The car (AI) is powerful, but for a 10-second walk, the car is a massive waste of time. Yet, we think the car is a "super-speed" tool.

Study 3: The "Snowball Effect"

The Setup: They split people into two groups. One group did the first few tasks using AI. The other group did them alone. Then, they gave everyone new tasks and saw what they chose.

The Discovery: Using AI once makes you more likely to use it again, even if it's a bad idea.

The Effect: People who used AI at the start were much more likely to use it again on the next simple tasks.
The Trap: Not only did they use it more, but they also became more convinced that AI was faster, even though the data showed it wasn't.
The Analogy: It's like trying a new video game controller. If you use it for the first level, you get used to holding it. Even if the old controller was actually better for the next level, you keep using the new one because you're already holding it. You convince yourself, "This new controller must be faster," even if you're just slower. This creates a loop where we get stuck relying on AI, believing it helps, while it actually makes us less efficient.

The Two Main Mistakes We Make

The paper identifies two specific ways our brains get tricked:

The "I Won't Use It" Blindspot: We think we are rational and will only use AI for hard things. But in reality, we use it for easy things too, often without realizing how often we do it.
The "Magic Speed" Hallucination: We imagine AI is a turbo-boost that instantly solves problems. In reality, for simple tasks, the time it takes to talk to the AI (typing, waiting, reading) is longer than just doing the math or writing the sentence ourselves.

The Bottom Line

The paper concludes that we are stuck in a feedback loop of over-reliance.

We use AI too much on simple tasks.
We think it saves us time (it doesn't).
Because we think it saves time, we use it again.
This makes us worse at doing things on our own and keeps us trapped in the illusion that AI is a magic shortcut, even when it's just a detour.

The researchers suggest that to fix this, we need to realize that doing things ourselves isn't always the "hard" option. Sometimes, the "easy" button (AI) is actually the hard way to go for simple jobs.

Technical Summary: The Efficiency-Gain Illusion in AI-Assisted Simple Tasks

Problem Statement

As Large Language Models (LLMs) become ubiquitous for tasks ranging from coding to personal advice, a prevailing narrative suggests they dramatically boost productivity. However, emerging evidence indicates that AI assistance does not always save time or effort. This paper investigates the decision-making mechanisms behind human reliance on AI for cognitively simple tasks (e.g., arithmetic, spell-checking, factual recall). Specifically, it addresses whether users accurately calibrate their reliance on AI and whether their usage patterns align with the principle of resource rationality—delegating tasks only when internal cognitive costs exceed the costs of external support. The core problem is the potential existence of systematic miscalibrations where users overestimate AI's efficiency gains, leading to inefficient usage even on tasks that can be completed quickly without assistance.

Methodology

The authors conducted three large-scale, pre-registered user studies with a total sample size of $N = 2,691$ participants recruited via Prolific. The studies utilized a between-subjects design to avoid cognitive dissonance and carry-over biases affecting self-reported preferences.

Task Construction: The researchers constructed 24 tasks derived from the Taxonomy of User Needs and Actions (TUNA), covering four categories: Information Seeking, Information Processing, Procedural Guidance, and Content Creation. Each task had "easy" and "difficult" variants, all designed to be completable in under five minutes independently.
Study 1 (Self-Estimate Miscalibration): Participants completed four tasks with the option to use an AI chatbot (GPT-4o). Their actual behavioral choice to use AI was compared against a "predicted rate" derived from a separate sample of participants who stated their preference for independent vs. AI-assisted completion.
Study 2 (Efficiency-Gain Illusions): This study compared predicted versus actual time and effort.
- Prediction Sample ( $N=600$ ): Participants predicted the time and mental effort (measured via NASA-TLX scale) required for independent completion versus completion with AI or another human participant.
- Completion Sample ( $N=1,001$ ): Participants were randomly assigned to complete tasks independently or with AI. Actual completion times were recorded via hidden timers, and effort was self-reported.
Study 3 (Carryover Effects): Participants underwent an "exposure phase" where they completed two tasks either independently or with AI (varying by difficulty). They then entered a "test phase" to measure subsequent AI adoption rates and their calibration regarding time/effort savings.

Key Contributions and Results

1. Self-Estimate Miscalibration

The study identified a significant gap between predicted and actual AI usage.

Finding: The population-level actual rate of AI use was 47%, significantly higher than the predicted rate of 33%.
Nuance: This miscalibration was most pronounced for easy task variants. Participants predicted a 20% usage rate for easy tasks but actually used AI 38% of the time. For difficult tasks, the gap was smaller (46% predicted vs. 57% actual).
Implication: People systematically underestimate their own propensity to rely on AI for simple tasks, even when they claim they would prefer to work independently.

2. Efficiency-Gain Illusions

Participants held distorted beliefs about the benefits of AI, termed "efficiency-gain illusions."

Speedup Illusion: Participants predicted AI would save an average of 55.7 seconds per task. In reality, AI assistance saved only 7.5 seconds on average. For easy tasks, AI actually slowed completion by 10 seconds compared to independent work due to interface friction (prompting, response generation, and processing).
Offloading Illusion: Participants overestimated the mental effort required to complete tasks independently (predicted 2.66 vs. actual 2.36 on a 7-point NASA-TLX scale) while accurately estimating the low effort of using AI. This led to an overestimation of effort savings.
AI Specificity: These illusions were specific to AI. When participants predicted assistance from another human, the expected time savings were significantly smaller than those predicted for AI, indicating a unique bias toward AI efficiency.

3. The Overreliance Feedback Loop

Study 3 demonstrated that prior exposure to AI exacerbates miscalibration and increases future reliance.

Carryover Effect: Participants who used AI during the exposure phase were significantly more likely to use AI in the subsequent test phase (44.5%) compared to those who completed tasks independently (27.7%).
Reinforcement of Bias: Exposure to AI did not correct users' beliefs; instead, it increased their agreement with the "speedup illusion" (the belief that tasks are faster with AI).
Inefficiency Persists: Even after exposure, those who chose AI in the test phase spent more time (7.06 seconds longer) and reported higher effort than those who worked independently.

Significance and Claims

The paper claims to provide evidence of specific cognitive biases in human-AI interaction that challenge the narrative of AI-driven efficiency.

Mechanism of Overreliance: The authors argue that the "efficiency-gain illusion" creates a feedback loop. Users habituate to AI use because they believe it saves time, even when it does not. Because they are unaware of this miscalibration, they lack the incentive to change their behavior.
Cognitive Risks: The study highlights risks associated with this overreliance, including cognitive deskilling, reduced confidence in independent performance, and impaired learning abilities.
Intervention Point: The authors suggest that addressing these miscalibrations is a critical intervention point. They posit that mere exposure to AI does not improve calibration; rather, interventions at the belief level are necessary to encourage independent thinking and promote healthier, more efficient AI use.
Resource Rationality: The findings suggest that current human-AI interactions often violate resource rationality, as users delegate tasks to external tools even when the internal costs of performing them are lower than the costs of using the tool.

The paper concludes that while AI has potential as a "thought partner," its current deployment on simple tasks is often counterproductive due to skewed mental models regarding efficiency gains.