💻 computer science

Cognitive offloading and the speedup illusion in human-AI interaction

A large-scale study reveals that users suffer from a "speedup illusion" when interacting with AI, significantly overestimating the time savings and underestimating the effort required for AI-assisted tasks compared to independent work, despite actual completion times remaining unchanged.

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

Published 2026-05-25

📖 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" Illusion

Imagine you are trying to clean your room. You have two options: do it yourself, or hire a super-fast robot to do it for you. Most people assume the robot will finish the job in half the time. They think, "If I let the robot do the heavy lifting, I'll be done in 10 minutes instead of 20."

This paper investigates whether that assumption is true. The researchers found that while the robot (AI) does make the job feel easier, it often doesn't actually make it faster. In fact, for simple tasks, using the robot can sometimes take just as long as, or even longer than, doing it yourself.

The authors call this the "Speedup Illusion." It's a mental trick where our brains confuse "feeling less tired" with "finishing sooner."

How They Tested It

The researchers ran a massive experiment with over 1,200 people. They split the group into two main scenarios:

The Guessers: These people were asked to predict how long a task would take. Some guessed how long it would take them to do it alone; others guessed how long it would take if they used an AI assistant or even a very smart human friend.
The Doers: These people actually had to do the tasks. Some did them alone; others used an AI chatbot to help.

The tasks ranged from simple things (like "Name one Olympic medalist") to harder things (like "Summarize a long article" or "Solve a logic puzzle").

What They Discovered

1. The "Robot" is Faster in Our Heads, but Not in Reality
When people guessed, they were very confident that AI would save them a huge amount of time. They thought AI would be much faster than a human friend, too.

The Reality: When they actually did the tasks, the time it took to finish with AI was almost the same as doing it alone. For simple tasks, there was zero time saved.
The Twist: People were actually pretty good at guessing how long they would take to do a task alone. But they were terrible at guessing how long it would take with AI. They consistently underestimated the time AI would need.

2. The "Effort vs. Time" Disconnect
Here is the most interesting part: Even though the clock showed the same amount of time passing, people felt like they worked much less hard when using AI.

The Analogy: Think of carrying a heavy box.
- Doing it alone: You carry the box up the stairs. It takes 2 minutes, and your muscles burn.
- Using AI: You ask a machine to carry the box. You just have to type a command and wait. It still takes 2 minutes (because the machine needs time to think and type back), but you feel like you did almost nothing.
The Result: Because the work felt so easy, people's brains tricked them into thinking the task was also faster. They confused "low effort" with "high speed."

3. The "Thinking" Trap
The study found that people who dislike thinking (who prefer to avoid mental effort) were the most likely to believe the Speedup Illusion. They were the most eager to hand over the task to AI, believing it would be a massive time-saver, even when it wasn't.

4. Why Did AI Sometimes Slow Things Down?
For some complex tasks, using AI actually took longer than doing it alone. Why?

The "Translation" Cost: The user had to spend time writing a clear prompt (instructions) for the AI.
The "Decoding" Cost: The AI would sometimes give a long, confusing, or overly detailed answer. The user then had to spend extra time reading, understanding, and fixing that answer.
The Net Result: The time saved by not doing the work was eaten up by the time spent talking to the AI and fixing its mistakes.

The Dangerous Loop

The paper warns of a "feedback loop."

We think AI is a magic time-saver (even when it isn't).
We use it because we want to save time.
We feel less tired (less effort), so we think we made a good choice.
We use it even more, even for simple tasks where it doesn't help.
Over time, we might lose our own ability to do things quickly on our own because we've gotten out of the habit.

The Bottom Line

The paper concludes that AI is great at reducing mental effort, but it is not a guaranteed time machine.

We need to stop assuming that because something feels easier, it is also faster. Sometimes, the "shortcut" is just a different path that takes the same amount of time, but feels much more relaxing to walk. The key is to be realistic: AI helps us feel less stressed, but it doesn't always give us more free time.

Technical Summary: Cognitive Offloading and the Speedup Illusion in Human-AI Interaction

Problem Statement
Large Language Models (LLMs) are increasingly viewed as tools to boost human productivity by accelerating task completion. However, a critical gap exists in understanding whether users are well-calibrated in estimating the time savings offered by AI. While resource-rational theories suggest humans offload cognition when internal costs exceed external tool costs, this decision relies on accurate mental models of both self and tool capabilities. The authors hypothesize an "asymmetric miscalibration": users possess accurate metacognitive monitoring of their own independent completion times but lack privileged access to the internal workings of LLMs, leading to a systematic underestimation of AI-assisted completion times. This phenomenon, termed the "speedup illusion," suggests users may overestimate efficiency gains, particularly on simple cognitive tasks where AI assistance might not actually save time.

Methodology
The authors conducted a preregistered, large-scale behavioral study ( $N = 1,237$ ) using a between-subject design to minimize carryover biases. The study comprised two distinct samples:

Prediction Sample ( $N=401$ ): Participants predicted completion times for 24 tasks under three hypothetical conditions: independent completion, AI assistance (GPT-4o), and assistance from another highly intelligent human. They also indicated their preferred mode of completion.
Completion Sample ( $N=836$ ): Participants actually completed the same 24 tasks either independently or with AI assistance. A hidden timer recorded actual completion times, and the NASA-TLX scale measured subjective cognitive effort.

Task Construction
The 24 tasks were categorized into four cognitive domains based on a taxonomy of LLM use:

C1: Information Seeking
C2: Information Processing & Synthesis
C3: Procedural Guidance & Execution
C4: Content Creation & Transformation
Each category included six tasks spanning two difficulty levels (e.g., retrieving one fact vs. ten facts). The analysis focused exclusively on high-quality, correct responses to ensure valid time and effort measurements.

Key Results

The Speedup Illusion: Participants significantly underestimated the time required for AI-assisted tasks. While actual independent completion times were accurately predicted, the actual time for AI-assisted tasks was approximately 57.8 seconds longer than predicted. This miscalibration persisted across all task categories and difficulty levels.
Asymmetry in Calibration: In contrast to the AI condition, there was no significant difference between predicted and actual times for independent completion, indicating accurate self-calibration.
Task Difficulty and Efficiency: AI assistance only resulted in faster completion times for difficult tasks (saving ~26 seconds), not easy tasks. For easy tasks, independent and AI-assisted completion times were statistically equivalent.
Dissociation of Time and Effort: A critical finding is the decoupling of objective time and subjective effort. While AI assistance did not consistently reduce completion time, it significantly reduced subjective mental effort (NASA-TLX scores) across all tasks. Participants felt the process was less demanding even when it took the same amount of time.
Predictors of Miscalibration: Individuals with a lower "Need for Cognition" (more averse to thinking) were more susceptible to the speedup illusion, believing AI would save more time. Conversely, prior AI usage frequency and experience did not predict the magnitude of calibration error.
Prompting Dynamics: Analysis of interactions revealed that cognitive effort is often spent on prompt construction and post-response processing (reading/understanding verbose outputs) rather than the model's generation time. In some cases (e.g., logic problems), the time spent processing the AI's output exceeded the time saved, contradicting user expectations.

Key Contributions

Identification of the Speedup Illusion: The study empirically demonstrates that users hold a biased mental model where they expect AI to speed up tasks even when it does not, specifically on simple cognitive tasks.
Time-Effort Dissociation: The research highlights that efficiency gains in human-AI interaction are not solely defined by time reduction. AI consistently reduces subjective effort, which may drive user behavior even when objective time savings are absent.
Calibration as a Mechanism: The paper frames calibration not just as a user skill but as a central mechanism in human-AI interaction, suggesting that misalignment between expectation and reality drives adoption patterns.

Significance and Claims
The authors argue that the "speedup illusion" creates a risk of overreliance. Because users conflate reduced effort with reduced time, they may delegate tasks to AI even when it offers no temporal advantage. This behavior could initiate a negative feedback loop: miscalibration encourages AI use, which leads to further miscalibration and potential cognitive deskilling (loss of independent capability). The paper concludes that narratives around AI productivity must move beyond simple time metrics to include the subjective experience of effort. Understanding and correcting these calibration gaps is presented as a necessary intervention to prevent disruptive levels of offloading and to foster more resource-rational human-AI collaboration.