Keep your skills growing while AI does part of the work

6 min read1 quick quiz1 checklistSources cited

When you learn a new skill, use AI to explain and to check your work. Do not use it only to produce the answer. A 2026 experiment found that people who handed the work to an AI assistant learned less than people who worked on their own.

What the research found

Judy Hanwen Shen and Alex Tamkin ran a randomized controlled trial. This kind of test assigns people to groups by chance. The paper appeared on arXiv in January 2026. Anthropic also published a summary on January 29, 2026.

The study had 52 software engineers, mostly junior. Each one learned a Python library that was new to them. One group could use an AI assistant. The other group could not. Then everyone took a quiz.

• The AI group scored 50% on the quiz. The hand-coding group scored 67%.
• The AI group finished about two minutes faster. This gap was not statistically significant, so it may be chance.
• The largest gap appeared on debugging questions.

The researchers also grouped how people used the assistant. Three habits kept quiz scores high (65% or more on average). People asked for code and then asked questions to understand it. People asked for code and an explanation together. People asked only concept questions and fixed their own errors.

Three other habits kept scores low (below 40% on average). People handed over all the code writing. People began with questions and then handed over everything. People used AI to debug and verify code without building their own understanding.

What this study cannot tell you

Read the limits before you change your habits.

• The sample was small. Each habit group held only 2 to 7 people. The authors say the habit analysis does not prove cause.
• The people were software engineers, not virtual assistants. We apply the result to your work by analogy.
• The task was one new library. The quiz came right after the task. The study does not show long-term skill.
• The tool was a chat assistant. The authors call the findings preliminary.
• The authors work at Anthropic, which makes AI assistants. Independent studies would add confidence.

The authors also note that their earlier work looked at productivity on familiar tasks. Faster work on tasks you know and slower learning of new skills do not contradict each other.

A second signal: confidence and checking

Hao-Ping Lee and colleagues surveyed 319 knowledge workers. The paper appeared at CHI 2025 and is listed by Microsoft Research. Workers who trusted the AI more reported less critical thinking. Workers who trusted their own skill more reported more critical thinking.

This survey is self-reported. It shows a link, not a cause. It still gives one practical lesson. Your own skill is what lets you question the output. This idea connects to our guide on five checks for AI-assisted client work.

Example: two ways to learn a new tool

This example is made up for illustration. It is not a measured result.

Mara is a virtual assistant. A client asks her to build an email automation in a platform she has never used. She has until Friday.

• Path A. Mara asks the AI to write every step. She follows the steps and sends the result. She finishes on Wednesday.
• Path B. Mara builds the first step herself and notes where she gets stuck. She asks the AI to explain the trigger concept. She then asks for a draft with reasons. She writes two sentences on how the flow works. She finishes on Thursday.

On the next Monday, the automation fails. Mara on Path B can find the broken trigger. Mara on Path A must start from the beginning with the AI. Path B cost one extra day. It also gave Mara a skill she can bill for again.

Start here: a learn-first routine

Use this routine when the skill is new and clients will ask for it again.

Tradeoffs

The routine costs time on day one. Use full delegation when you already know the task, when the risk is low, or when you will never repeat the task. Use the routine for skills that clients pay for again and again.

This advice applies to any reader. If your client hours are tight or your connection is unstable, pick one skill each month for learn-first mode. Do not try to do it for every task.

Sources

1. Shen, J. H. and Tamkin, A. “How AI Impacts Skill Formation.” arXiv:2601.20245, submitted January 28, 2026 (v2 February 1, 2026). arxiv.org/abs/2601.20245
2. Anthropic. “How AI assistance impacts the formation of coding skills.” January 29, 2026. anthropic.com/research/AI-assistance-coding-skills
3. Lee, H.-P. et al. “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” CHI 2025, via Microsoft Research. microsoft.com/en-us/research. Pages read October 12, 2026.