# Will AI Make Your Employees Worse at Their Jobs? What a 26,811-Student Study Means at Work

- Canonical: https://spiai.ai/blog/ai-deskilling-employees
- Published: 2026-10-31
- Author: SpiAI
- Description: AI deskilling is real: a 26,811-student study found better homework and worse exams. How a small business can use AI without eroding staff skills.

---
A study that circulated widely on X in late October followed 26,811 secondary students in central China for 30 months. As [Laurence Moroney summarized it](https://x.com/lmoroney/status/2113712896081617321): "Within six months of starting to use AI, students were finishing homework in about 45 minutes instead of 64, and their homework scores rose 18%. Over the same stretch, their monthly closed-book exam scores fell 20%."

The line that should stop a business owner came from a reply further down the feed. [@KKrishuGupta wrote](https://x.com/KKrishuGupta/status/2113521953902678510): "AI made the output better and the learning worse. The same thing is happening at work, we just don't have exams there."

That is the real risk for a service business with 5 to 75 people. The quotes look better, the emails read cleaner, the reports arrive faster. Nobody notices that the person producing them understands the work a little less each month, until the day the AI is wrong and nobody catches it.

## TL;DR

- A CEPR study of 26,811 students found AI raised homework scores 18% and cut homework time 30%, while closed-book exam scores fell 20% within six months.
- The losses concentrated among students whose fast, high-scoring homework suggested they were outsourcing the work. Students who kept spending similar time saw small losses.
- Workplaces have the same pattern and no exams. Better-looking output can hide shrinking judgment, especially in newer staff.
- Design AI use so people still do the thinking that builds skill. AI drafts, your team approves, and the approver has to actually know the work.

## What did the study actually find?

The paper is "The Generative AI Learning Penalty: Evidence from Chinese Secondary Education" by David Strömberg, Victor Lei and Yanhui Wu, published as a CEPR discussion paper. It used administrative records for students in grades 7 to 12 across nine subjects, linking weekly homework scores and completion times to monthly closed-book exams and entrance exams. Because students adopted AI at different times, the authors compared each student's results before and after adoption against students who had not yet adopted.

The headline numbers, from the paper's abstract and the authors' VoxEU column:

| Measure | After AI adoption |
|---|---|
| Homework completion time | Down about 30%, from about 64 to about 45 minutes |
| Homework scores | Up 18% |
| Monthly closed-book exam scores | Down 20% within six months |
| High-school entrance exam scores | Down 24% at full effect, about two years later |
| College entrance exam scores | Down 18% at full effect |

Two further findings matter most for employers. First, roughly 80% of AI users behaved in a way consistent with outsourcing their homework: exceptionally short completion time combined with high scores. Their losses were the largest. AI users who kept spending about the same time on homework as non-users saw small losses. Second, among AI users the usual relationship flipped: higher homework scores went with worse exam results. The authors note that a sudden jump in homework quality became a warning sign.

## Why does this matter for a small business?

Students are not employees, and a classroom in China is not a plumbing company in Ohio. The authors studied learning, not job performance, and we should not stretch their numbers further than they go. But the mechanism is familiar to anyone who has trained staff.

Skill in most service work comes from doing the hard middle part yourself: reading the messy customer email and figuring out what they actually want, building the quote line by line, working out why the numbers do not reconcile. AI is very good at doing exactly that middle part. When it does, the output improves immediately and the learning quietly stops.

In a small business, that shows up in three places:

- **New hires.** Someone six months into the job who has always had AI write their customer replies may never build the instinct for when a reply is wrong.
- **The review step.** "AI drafts, your team approves" only works if the approver knows enough to catch a mistake. A rubber stamp is not a review.
- **Your succession plan.** The knowledge in your senior people's heads gets passed on through doing the work together. If juniors skip that, it does not get passed on.

## Which work should staff still do without AI?

Not all work builds skill, and not all skill is worth keeping. Sort tasks into three groups.

**Let AI do it.** Formatting, summarizing a long thread you will still read, first drafts of routine messages, transcribing notes. Nobody's career depends on reformatting a proposal.

**Let AI draft, but make people reason first.** Quotes, diagnoses, customer complaints, scope decisions. Have the person write down their own answer or reasoning before they look at the AI's. Then compare. This is the workplace version of keeping homework time the same, the group in the study with small losses.

**Do it by hand, at least while learning.** The core judgment of your trade: pricing a job, reading a site, reconciling accounts, deciding what to tell an upset client. New staff should do these without AI until they can explain their reasoning out loud.

## How to keep skills while using AI

1. **Write "think first" into the workflow.** For judgment tasks, the person records their own answer before opening the AI draft. It takes a minute and it is the single biggest protection.
2. **Make juniors explain AI output.** In review, ask "why is this right?" If they cannot answer without the AI, the review was not real.
3. **Run occasional closed-book checks.** The study's authors suggest short in-class quizzes. At work, that can be a monthly ten-minute walk-through: price this job, read this email, what would you do?
4. **Track time as well as quality.** The study's warning sign was work that got much faster and much better at once. If someone's output jumps overnight, look at how it is being produced.
5. **Pair seniors and juniors on the hard cases.** Keep the apprenticeship model for the work that matters most, with the AI as a tool both can see, not a shortcut one hides.
6. **Rotate the reviewer.** If the same senior person approves everything, nobody else learns what a good approval looks like.

## Is AI deskilling inevitable?

No, and the study itself shows why. The students who kept engaging with their homework saw small losses. The damage came from outsourcing, not from the tool. Other discussions on X this month made the related point that people with real expertise get the most out of AI, because they know what to ask for and can spot when the answer is wrong.

That is the opportunity for an owner-led business. You can design how AI is used in a way large organizations struggle to. You know every person on your team, you see their work, and you can decide where the thinking has to stay human.

## What to watch for

- Output quality jumps suddenly while time spent drops sharply.
- Staff cannot explain why a draft they approved is correct.
- Mistakes reach customers that an experienced person would have caught.
- New hires are faster than veterans at producing work but slower to handle anything unusual.
- Nobody on the team can do a core task when the AI tool is down.

## Verdict

Verdict: Buy AI for your team, with a skills plan attached. The evidence says the risk is real when people outsource the thinking, and small when they keep engaging with the work. Keep AI on formatting and first drafts, require people to reason before they read the AI's answer on judgment calls, and keep new staff doing core tasks by hand until they can explain them. AI drafts, your team approves, and the people approving should still be able to do the work themselves.

**Want AI in your workflows without hollowing out your team's skills?** In a free 30-minute assessment we sort your team's work into what AI should do, what it should draft and what people need to keep doing themselves. [Book a free assessment](https://spiai.ai/)

## FAQ

### What is AI deskilling?

AI deskilling is the loss of skill that can happen when people let AI do the parts of a task that used to build their expertise. Output may look better while the person's own understanding weakens.

### What did the 26,811-student AI study find?

A CEPR discussion paper by Strömberg, Lei and Wu found that after adopting AI, students' homework scores rose 18% and homework time fell about 30%, but monthly closed-book exam scores fell 20% within six months, with larger long-run drops on entrance exams.

### Does the study prove AI makes employees worse at their jobs?

No. It studied secondary students' learning, not job performance. It does show a mechanism worth taking seriously at work: when people outsource the thinking, output improves while understanding declines.

### Which employees are most at risk of AI deskilling?

Newer staff who are still building core skills. Experienced people already have the judgment to check AI output, while juniors may never develop it if AI always does the hard part.

### How can a small business prevent AI deskilling?

Have staff write their own answer before reading the AI's on judgment tasks, ask them to explain AI output in review, run short closed-book check-ins and keep new hires doing core tasks by hand while they learn.

### Should we ban AI for new employees?

A ban is usually unnecessary. Let new staff use AI for formatting and summaries, but require them to do core judgment tasks themselves until they can explain their reasoning without it.

### How do I know if someone is over-relying on AI?

Watch for work that suddenly becomes much faster and much better at the same time, approvals that the person cannot explain and mistakes an experienced reviewer would have caught.