Artificial Intelligence · Learning

Imagine Asha learning spreadsheets. Yesterday, she could not write a lookup formula. Today, an AI assistant produces one, fixes the error, and explains the result with the confidence of a teacher holding the answer key.

The sheet works. Asha celebrates. Tomorrow, faced with a slightly different sheet, she opens the chat again.

Asha is fictional. The question is real: did she learn the skill, or did she complete a task with help? Both can be useful. They just deserve separate entries in the progress report.

Research on coding, mathematics and workplace AI suggests we should measure more than the quality of the finished output. Here is what those studies actually found, where their conclusions stop, and a practical way to check your own learning.

Two goals hiding inside the same chat

Before asking for help, finish this sentence: “At the end of this session, I want to be able to…”

If the answer is “send this correctly formatted document,” delegating routine formatting may be sensible. If the answer is “understand how this formula works,” receiving a perfect formula leaves a second job unfinished: demonstrating that understanding yourself.

Try keeping two measures: what you completed with assistance and what you can explain or do independently. A polished submission answers the first question. A fresh attempt with the chat closed starts to answer the second.

This distinction also matters at work. Our guide to AI in an everyday workday explores where assistance fits; here, the focus is the skill you want to take home after the work is done.

What the coding experiment found

In a 2026 preprint, Anthropic researchers studied 52 developers learning Trio, an unfamiliar Python library. Participants were randomly assigned to complete coding tasks with or without an AI assistant, then took a quiz. The AI group averaged 50%, compared with 67% without AI: a 17-percentage-point gap. The roughly two-minute speed advantage was not statistically significant. See the research paper and Anthropic’s study explanation.

Participants who asked conceptual questions or requested explanations tended to perform better than those who delegated heavily. Those small interaction-pattern comparisons were observational, however; they do not establish that a particular prompt caused better learning.

The boundaries matter. This was a small study of an unfamiliar library and immediate assessment, published as a preprint. It does not establish lasting skill loss, effects across every occupation, or a universal rule against AI. It does show why “the code ran” is an incomplete learning measure.

A classroom study adds an important distinction

A 2025 PNAS study involved nearly 1,000 mathematics students at one high school in Turkey. Researchers compared no AI with a general AI interface and a tutor designed with teacher input.

Both AI groups performed better during assisted practice. When students took an exam without assistance, the general-interface group scored about 17% lower than the control group. That is a relative difference, unlike the percentage-point comparison above. The specially designed tutor largely avoided this penalty, but did not show a statistically significant exam advantage over the control.

The tutor combined hints with teacher-supplied correct solutions and common mistakes. It was more elaborate than adding “teach me” to an ordinary chat. The trial concerns a particular school, subject and implementation; it cannot settle every question about AI education. The journal’s August 2025 correction addresses an author affiliation, not the findings.

Confidence is useful. It is also worth checking.

A Microsoft Research and Carnegie Mellon study presented at CHI 2025 surveyed 319 knowledge workers, collecting 936 examples of AI use. Greater confidence in AI was associated with less reported critical thinking; greater confidence in one’s own ability was associated with more. Participants also described work shifting toward checking and integrating AI output.

These were self-reports, not a randomized experiment or a measurement of permanent cognitive decline. The useful question for a reader is modest: when an answer feels convincing, what evidence would let you check it?

Our AI answer fact-checking guide provides a companion routine for claims and sources. For a skill, ask yourself to produce an explanation, prediction or working example as well.

Try a 15-minute learning check

This is an editorial exercise you can adapt, not a validated intervention or a promise to reproduce the studies. Choose one small, low-stakes skill: a spreadsheet formula, a short function, or a paragraph in a language you are practising.

1. Minutes 0–3: make your own attempt. Write down the goal and the point where you get stuck. An unfinished attempt gives you something concrete to discuss.
2. Minutes 3–6: ask for one hint. Show the attempt and ask the assistant to identify the next useful question. If the response supplies the full solution, cover it and request a smaller hint.
3. Minutes 6–9: explain your next step. Say why you chose it and what result you expect. Check the explanation against your course material, documentation or a known example.
4. Minutes 9–13: close the chat and change the problem. Use different values or a slightly different requirement. Attempt it without the previous solution in view.
5. Minutes 13–15: record the gap. Note what you could do, what needed help, and one question for your next session.

For fictional Asha, the changed problem might use a different lookup column and include a missing match. Her self-check is whether she can explain what the formula should return in both cases. Getting stuck is useful information; it tells her what to practise next.

Try another short unaided example the next day. Keep the result separate from how fluent the original AI explanation felt. A smooth explanation can be enjoyable without being your own explanation yet.

A prompt that leaves you work to do

I am learning [specific skill]. Here is my attempt: [attempt]. My goal is to solve a new example independently. First ask me one question that helps locate my misunderstanding. Give one hint at a time and wait for my response. Ask me to explain the next step before revealing a full solution. When we finish, give me a similar problem with one meaningful change, withhold its solution, and ask me to predict the result. Flag uncertain claims so I can check them against [course notes or official documentation].

Treat that as a request, not a guarantee. An assistant may give away answers, miss an error or approve faulty reasoning. Use an answer key, instructor feedback, documentation or a small executable example to check the important parts. A chatbot congratulating you is pleasant; it is not an assessment system.

When should you just delegate?

You do not need to turn every small chore into a seminar. Decide which capabilities matter for your goals and which tasks you are comfortable outsourcing. You might delegate formatting while practising the analysis, or request boilerplate code while personally tracing the logic you need to understand.

When you rely on an agent to act as well as explain, the ability to check its work still matters. Our AI agent reliability guide looks at that verification problem.

Questions readers often ask

Does this mean AI makes people less intelligent?

No such broad conclusion follows from these studies. They examine particular learning outcomes or reported thinking under particular conditions. Permanent changes in intelligence are a different claim.

Is a hint always better than a complete worked example?

The evidence here does not establish that rule. A complete example may be something you need to inspect. After reading it, try explaining the choices and solving a changed example to see what you can do yourself.

What if I have a deadline?

Separate the immediate deliverable from your learning goal. Record which parts required help, check the output, and schedule a small independent attempt afterward. Follow any course or workplace rules about permitted assistance and disclosure.

For your next session, choose one thing you want to be able to do with the chat closed. Put that goal above the prompt box. Then check it before calling the session finished.

Editorial note: Prepared with AI assistance and checked against the linked primary sources on 15 October 2026. Asha is an invented illustration, and the learning routine is an editorial suggestion. The hero image is an original AI-generated illustration. Read our editorial policy or send a correction.