The Efficiency-Gain Illusion: Every Minute Spent on AI May Be Slower Than Doing It Yourself
| Paper Info | | |---|---| | Title | The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks | | Authors | Sunny Yu, Myra Cheng, Ahmad Jabbar, Ilia Sucholutsky, Katherine M. Collins, Dan Jurafsky, Robert D. Hawkins | | Institution | Stanford University | | arXiv ID | 2605.22687 | | Date | May 21, 2026 | | Categories | cs.CY / cs.HC | | Method | Three preregistered user experiments (N=2691) | | Key finding | A dual systematic bias in AI use — people underestimate their own frequency of AI use while overestimating its efficiency gains, forming a self-reinforcing "over-reliance feedback loop" |
⏱️ Prologue: How Many Times Did You Open AI in One Coffee Break?
On the afternoon of May 21, 2026, you're writing an email about rescheduling next week's meeting. You open a ChatGPT conversation and type: "Help me write a polite email telling colleagues the meeting moved from Wednesday to Thursday at 3 PM, explaining it's due to a venue conflict, and apologizing."
The AI generates a polished, complete, grammatically flawless email in five seconds. You copy, paste, tweak two words, and hit send. Total time: about thirty seconds.
Writing it yourself would have taken about ninety seconds.
You feel you saved a minute.
But here's the question you need to stop and ask: of those thirty seconds — opening ChatGPT, typing the prompt, waiting for generation, copy-pasting, tweaking — how much could you actually have saved?
Seven Stanford researchers behind *The Efficiency-Gain Illusion* — including NLP luminary Professor Dan Jurafsky — used data from 2,691 participants to tell you one thing: you may not have saved those thirty seconds at all. You may not even know how often you use AI.
And worse — the more you use it, the deeper the illusion becomes.
📊 Chapter 1: When 2,691 People Took a Simple Test
The researchers designed three independent, preregistered user experiments. The design was elegantly simple: give participants "cognitively simple tasks" — arithmetic, spell-checking, simple Q&A — and let them choose whether to do each themselves or delegate to AI. The actual efficiency data behind every choice (time spent, effort expended) was precisely recorded.
The researchers then compared three numbers: 1. How often people actually used AI — how many tasks they really delegated 2. How often people believed they used it — their retrospective estimates 3. The actual time/effort AI saved — objectively measured efficiency gains
Two systematic biases emerged.
Bias one: self-estimate miscalibration. In post-hoc surveys, people systematically underestimated how often they used AI. They thought "I only used it three or four times"; the actual data said "you used it seven times." The bias pointed in one direction — almost nobody overestimated their AI use; almost everyone underestimated it.
Bias two: the efficiency-gain illusion. People held a systematic overestimation of the time and effort AI saved them. They felt "AI cut my time in half," but actual measurement showed — on many simple tasks — the savings were not significant, and sometimes doing it manually was faster.
Stacked together, these two biases paint a troubling picture: people rely on AI more than they realize, and that reliance is often a bad deal.
🔄 Chapter 2: The Most Dangerous Feedback Loop — "Once Used, Always Used"
The paper's third effect is, in my view, the most important one.
The session-level carryover effect.
In the experiments, researchers observed that once a person used AI on one task, their probability of choosing AI for the *next* task rose significantly. And — more importantly — each additional use of AI further entrenched their bias about "how much time AI saved me."
In plain language: every time you use AI, your brain reinforces the logic of depending on it — even when it objectively isn't saving you time. This isn't just a cognitive bias; it's a positive feedback loop.
The dynamics look roughly like this:
Round one: you hand-calculate a multiplication problem. The AI could give you the answer with one click. You feel it saved you five seconds — in fact maybe only two, because opening the window and typing the question ate the other three. But your brain logs the association "AI = time saved."
Round ten: you can no longer hand-calculate multiplication. Your first reaction to seeing numbers is "ask the AI." You throw thirty simple arithmetic problems at the AI — three seconds to type each, one second to wait — when calculating yourself would have taken about four seconds per problem. You saved nothing. But you still feel like you did.
Round one hundred: you no longer question whether you should ask the AI. You just instinctively type the equation into the chat window the instant the numbers appear.
That's the feedback loop, fully closed.
🧠 Chapter 3: Why Simple Tasks Are the Most Dangerous Trap
The paper deliberately scopes itself to "cognitively simple tasks." This is intentional — and it's the study's most profound design choice.
For complex tasks — say, "help me analyze a concurrency bug in this 3,000-line Python file" — AI's value is obvious. When it can do what you cannot, the cost is worth it no matter how many tokens it burns.
Simple tasks are different.
Arithmetic: 2+3=5. Anyone with elementary education can do this in milliseconds. But typing "what is 2+3" into an AI and waiting for a reply takes at least five to ten seconds total; doing it yourself takes under one second. The AI is ten times slower.
Spell-checking: "recieve" should be "receive." Human visual pattern recognition is extraordinarily efficient at native-language spelling. The AI may need a full token-generation cycle to suggest a correction — you just need one more glance at the word.
Simple Q&A: "What's the capital of France?" You already know. By the moment you know the answer, the AI's response hasn't even started generating.
Using AI on these tasks essentially means replacing your own existing knowledge retrieval with machine computation. And the friction costs of that replacement — opening the app, typing the question, waiting, reading the answer — frequently exceed the cost of simply doing it yourself.
But people don't perceive this gap. Because typing and waiting *feel* like "the machine is working while I rest" — an experience the brain files under "effort-saving" rather than "time-consuming." Our internal clock runs slower than the computer's clock while we wait for AI.
🔍 Chapter 4: The Boundaries of the Study — What the Paper Honestly Says
The researchers themselves candidly acknowledge several things:
First, only "cognitively simple tasks" were tested. For complex tasks, AI's real gains are undeniable. The paper's conclusions do not apply to scenarios where "AI can do things humans cannot." The researchers drew this boundary themselves.
Second, benefits beyond time were not evaluated. The paper measured time and effort savings. But AI may confer other benefits — reduced cognitive load (even at equal total time, giving the brain a break has value), fewer errors (especially during lapses in attention), or the ability to do other things while the AI answers (parallel multitasking). These benefits fall outside the paper's evaluation framework.
Third, the duration of the session-level carryover effect. The paper observed the short-term "used AI, more likely to use AI again" effect. But how long does it last? Minutes, hours, days? No longitudinal tracking data is provided. The "half-life" of this feedback loop is an important unknown.
Fourth, sample generalizability. 2,691 participants is a substantial sample, but the paper doesn't detail the distribution of participants' technical literacy. If participants were mostly college students (common in user studies), applicability to older or non-technical populations may need further validation.
🎭 Chapter 5: Self-Deception in the AI Era — The Most Dangerous Laziness
This paper reminds me of something I often see in coffee shops.
Someone sits in front of a laptop with three AI chat windows open. They type into a window: "Summarize this 800-word news article for me" — wait, read the summary, feel dissatisfied, type "be more detailed" — wait, read, adjust again —
That same 800-word article, read from start to finish, would take them two and a half minutes.
Instead they spent five minutes on the AI, went through two rounds of prompt-tweaking, and the final output wasn't necessarily better than reading it themselves.
During those five minutes, their subjective feeling was "AI is helping me, I'm being efficient." The objective fact: "You spent 150% more time to accomplish the same thing as reading it yourself."
This is a trade you make with yourself — the time you pay gets mentally tagged as 'saved,' because you feel you weren't 'working.'
That is the core of the efficiency-gain illusion. Between the feeling of 'less effort' and the reality of 'less time' lies a huge gap we never perceive. And each time we cross that gap, we build another layer of dependence on AI — even when the gap is fake.
🏁 Epilogue: Ask Yourself One Question Before You Use It
Having read the paper, I've distilled one very simple question to ask yourself every time you reach to open an AI chat window:
"If the AI went down right now, could I do this myself within two seconds?"
For things where the answer is 'yes,' the AI won't be faster than you.
For things where the answer is 'no,' the AI has real value.
The essence of this question is that it forces you to face your own known/unknown boundary honestly. You know what 2+3 equals. You know the capital of France. You know "recieve" is misspelled. But when you reach for the AI, you never first ask yourself "do I already know the answer?"
Not because you need the AI. Because you've gotten used to it.
And habit, it turns out, is precisely the core risk factor this paper identified.
📚 References
1. 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. *arXiv:2605.22687*. 2. Dell'Acqua, F., et al. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. *Harvard Business School Working Paper*. 3. Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at Work. *NBER Working Paper*. 4. Pezzo, M. V., & Beckstead, J. W. (2006). The Cognitive Illusion Controversy: A Methodological Debate in Disguise. *Psychological Methods*. 5. Parasuraman, R., & Riley, V. (1997). Humans and Automation: Use, Misuse, Disuse, Abuse. *Human Factors*.