A short email should feel like a short email.

But it is easy now to make it weirdly complicated. Open the AI tool. Write the prompt. Read the answer. Notice the tone is off. Add context. Read again. Copy the part that works. Fix the phrase you would not actually say. Check the fact. Rewrite the ending.

The original job may have taken ninety seconds. The shortcut took four minutes.

A new preprint from researchers at Stanford, NYU, MIT, and Princeton’s AI Lab studies that gap. The paper calls it the “efficiency-gain illusion.” Across three preregistered studies with 2,691 participants, people often chose AI help for simple cognitive tasks even when it gave them little or no meaningful time or effort savings.

In one study, participants expected AI to save about 55.7 seconds on average. The measured saving was 7.5 seconds. For the simplest task variants, the AI-assisted route was slower.

Put a fence around that claim. The paper looked at short, bounded tasks: arithmetic, spelling, recall, and brief rewriting in controlled study settings. It does not settle every argument about AI at work. A complicated project with messy inputs is a different animal.

But small work is everywhere. And small work is where extra steps can quietly eat the gain.

The answer is only one part of the cost. The full loop is prompt, wait, read, check, correct, copy, and decide whether the output can be trusted. Sometimes that loop pays for itself. Sometimes it is the checkout line.

Garry Tan put the blunt version in a public comment surfaced by Digg: “The mistake: Using AI for simple tasks.” His better fit was complex, multi-stage work with people and moving parts.

That is useful for normal work, because most people do not need an ideology about AI. They need routing judgment.

Use the tool when the work is hard to start but easy to check. Use it when the setup can be reused. Use it when the task is repetitive, the material is scattered, or the first draft is the blocker.

Slow down when the task is already clear and short. Slow down when you will have to verify every line anyway. Slow down when a wrong answer creates more cleanup than the original job.

The study also found a habit effect. After using AI on earlier tasks, participants became more likely to use it again later, and their sense of the time savings became less accurate.

That is the habit to watch.

Some tasks deserve a helper. Some deserve a template. Some deserve a checklist. Some deserve five quiet minutes and no new interface.

A checkout line is fine when the cart is full.

For one item, look around first.

What to watch

  • How much time would the task take without opening another tool?
  • Will the answer need checking line by line?
  • Does the AI setup get reused, or is this a one-off detour?
  • Could a wrong answer create cleanup work?
  • Are you choosing the tool because it helps, or because reaching for it became automatic?

Source note

Primary source is a 2026 preprint on AI efficiency-gain illusions for simple tasks. Treat as preprint/user-study evidence, not a claim about every AI workflow.