Learn With the Tool Beside You

The answer arrives before the question has quite settled in your mind.

You asked an AI tool to help interpret an unfamiliar report. It produces a smooth summary, identifies three apparent findings and suggests what they mean for your project. A task that felt large now fits on the screen.

There is genuine relief in that. There is also a choice. Will you use the answer as an entrance into the report, or as a reason never to open it?

The difference may be invisible in the first document you submit. Both routes can produce a convincing page. It becomes visible when someone asks where the main claim came from, whether the comparison is fair, or what would change if a key assumption were wrong.

This is the learning question at the center of AI-assisted work: what has become easier for you to do, and what have you become more able to do yourself?

We need both answers. A useful tool does not have to teach every time it helps. But if an entry role is supposed to develop a person, someone must check that development rather than infer it from improved output.

Help can be real

In a study of 5,172 customer-support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined a staggered AI rollout. Estimated issues resolved per hour rose by 15 percent on average, with larger gains among less experienced and lower-performing agents. Performance during tool outages was consistent with some learning beyond live assistance, though those estimates were noisy and chat mix could differ. This was one workplace setting, with managers and ongoing practice, not a randomized test of replacing teachers or a study of economy-wide hiring.¹

That evidence gives us a reason to resist a simple anti-tool rule. Access to guidance can help someone do work they are still learning. A beginner should not have to earn the right to useful assistance by struggling alone for an arbitrary period.

Think of a new worker faced with a problem they have not seen before. A suggestion can reveal relevant information or show a possible response. If the worker examines it, checks it against the case and sees the result, the interaction may become part of learning. The support can make a difficult task accessible enough to attempt.

The important words are examines, checks and sees the result. Those are actions we can design into a role. They should not be assumed simply because an AI product describes itself as a coach.

A manager can ask how the tool’s suggestion enters the workflow. Does the worker retain the chance to judge it? Can they reject it? Do they learn whether the action helped? Is someone available when the case is outside the tool’s reliable range? The answer is an account of the work, not an opinion about technology in general.

Finishing may still conceal a gap

Judy Hanwen Shen and Alex Tamkin randomized 52 Python-experienced programmers learning an unfamiliar library. Those given AI chat help scored lower on an immediate unaided comprehension quiz; the completion-time difference was not statistically significant. Most had substantial coding experience; the experiment tests unfamiliar-library learning rather than isolating career beginners. It lasted about an hour and did not measure long-term skill. Patterns in how participants prompted the tool were exploratory, not proof of a best prompting method.²

The result cautions us against using a finished task as evidence that its maker learned the underlying work. It does not tell us to forbid all assistance.

A person can arrive at a correct output without encountering the decision that would have taught them something. If the tool supplies the entire method, the learner may only need to recognize that the final thing looks plausible. That may be enough for a harmless convenience. It is a poor foundation for later responsibility if the person cannot tell when the method fails.

This is why a learning assignment should have a stated purpose. Are you trying to understand a new concept, produce a document on time, or demonstrate a skill you already have? These purposes can coexist, but they sometimes pull in different directions. The quickest route to a finished document may give the least practice in the part you need to learn.

Naming the purpose lets you choose the help rather than accept whatever the tool offers first.

Another task gives a different view

A 2026 working paper by Guillermo Cruces and colleagues studied adults in Argentina completing a business problem. AI improved assisted performance, especially in the lower-education group; that group showed some immediate carry-over on unaided follow-up questions. Of 1,795 eligible adults randomized, 1,174 completed the experiment. Participants needed a computer, and the education split was based on progress through tertiary study. The authors explicitly caution that the follow-up does not establish durable learning or transfer to a new task.³

These studies ask different questions. One examines performance while people do their jobs and during occasional outages. Another tests immediate understanding of unfamiliar programming material. Another follows an assisted business exercise with questions about the same problem. We should not line up the results as if they were measurements from the same instrument.

Together, they make the practical task more precise. We need to inspect the learner’s work under the conditions that matter. An improvement with assistance can be valuable. An immediate explanation without assistance tells us something else. Handling a different problem later is a further test.

A workplace can welcome the first improvement while making time to examine the others. There is no need to turn mixed evidence into indecision. It can guide a better question: which ability does this role need the learner to keep when the situation changes?

Start with a claim you can check

Return to the unfamiliar report. Before asking for a complete summary, choose a specific question. Perhaps you need to know whether the report’s participants resemble the people your organization serves. Or whether a reported improvement concerns speed, quality or both.

Open the relevant section yourself. Write a rough answer in plain words, with the part you do not understand marked as a question. This can be a few lines. The point is to give yourself a first view that the tool can help you examine.

You might then ask the tool to explain a term, locate a possible inconsistency in your interpretation or suggest a question you should ask of the method. Treat its response as something to verify against the report. If it points to a section or gives a quotation, check that the section exists and the quotation says what is claimed.

This is a proposed study routine, not a technique established by the experiments above. Its value for you should be judged by what you can explain and correct afterward.

You may discover that your first answer was wrong. Good. The correction has a place to land because you know what you previously thought. You may discover that the tool was wrong, or that the report does not contain enough information to answer the question. Those are useful outcomes too.

A professional summary sometimes needs to say that the evidence does not settle the decision. AI can help phrase that sentence, but it cannot make the missing evidence exist.

Keep the difficult part in the exercise

Tools tend to make the easy-to-request result readily available. Learning sometimes depends on staying with a less comfortable step.

If your goal is to understand why two quantities cannot be compared, asking for the final recommendation may let you avoid the comparison. Ask instead for help identifying the denominator, the period and the population. Then decide whether the quantities belong together, and explain your decision.

If your goal is to learn a small piece of code, you might first predict what it will do, run it in an appropriate safe setting and compare the result with your prediction. If the tool repairs it, examine the change. Can you explain why the new version behaves differently? Can you recognize a similar problem in a fresh example?

If your goal is communication, compare two possible messages against a clear audience and purpose. Which one makes a promise the sender cannot support? Which omits information the reader needs? A tool’s confident choice is a reason to inspect the text, not a substitute for your own decision.

In each case, preserve the part you are trying to learn. Let the tool remove incidental burdens where that helps. Formatting a practice document may be incidental. Deciding what belongs in its conclusion may be the whole point.

The boundary will change as you improve. You need not keep practicing a basic operation long after you can handle it reliably, merely to prove that your work is pure. Move the practice to the next meaningful difficulty.

Check yourself without turning it into punishment

After a period of assisted work, close the tool and explain the idea in your own words. Use a clean example if the learning goal calls for it. Identify a likely error and how you would notice it. Mark what remains uncertain.

This is information for choosing the next step. It should not become a ritual in which you try to expose yourself as an impostor.

You may understand the general idea but need the documentation for exact syntax. That is ordinary in many kinds of work. You may remember a procedure but have no good explanation of when to use it. That gap deserves a different kind of practice. You may discover that your main difficulty is interpreting the question, not carrying out the answer.

Be specific about the gap. I am bad at this is too large to repair. I cannot yet explain why this comparison is misleading gives you something to study.

A supervisor can use a similar check without making the learner perform helplessness in public. Ask for an explanation in an agreed review, offer a changed example and discuss what support remains appropriate. If the result shows that more teaching is needed, use it to plan teaching.

An assessment becomes unfair when it quietly changes the rules. If normal work allows documentation, an entirely closed-book test may measure something different. If the job requires independent action during an outage, some unassisted assessment may be necessary. Explain which condition matters and why.

The aim is readiness for real responsibilities, including the supports the job will actually provide.

Protect the information and the person

Learning with AI does not grant permission to put anything into a tool. Use the systems and materials your workplace has approved. Keep confidential records, client details and other people’s private information out of an outside service unless the proper permission and safeguards exist.

For personal practice, a public or invented example is usually easier to manage. You can learn how to summarize a complaint without pasting a real customer’s private message. You can practice a calculation with made-up records, clearly labeled, rather than a file copied from work.

Do not rely on a tool to decide whether sharing is allowed. That decision belongs to the relevant rules, owners and authorized people. When uncertain, ask before uploading.

There is also a question of how learning is monitored. A company could use every prompt as a basis for judging a worker, without explaining what is being recorded or how it will be interpreted. A fair learning arrangement should tell people what records are kept and who can see them. A beginner needs enough safety to ask a basic question without fearing that it will become a permanent verdict.

We want people to reveal gaps early. Designing a system that punishes every visible gap works against that goal.

The workplace has to leave time

A manager might be pleased when a new tool helps a beginner finish sooner. It is tempting to fill every saved minute with another task. If all the gain becomes more throughput, however, the role may leave no time to check what the person has learned.

The employer should make an explicit choice about learning time. A small part of the schedule can be reserved for reviewing a difficult case, inspecting an error or trying a task under a changed condition. The amount depends on the role. The commitment should be real enough to survive an ordinary busy week.

Expecting people to do all of this after paid hours moves the training cost back onto the newcomer. Those with more free time can comply. Those with care duties, a second job or limited energy may find themselves judged for a constraint the schedule created.

The tool can be powerful and the job can still be poorly designed. That is an institutional problem with an institutional remedy: decide what the role is supposed to develop, fund the practice and review the result.

A beginner can ask about these conditions before accepting a role. What will I be expected to do with AI? What do I need to understand without it? Who checks work before it affects a customer? When will I get feedback? These are questions about doing the job well, not signs of resistance to new technology.

The best outcome is neither permanent dependence nor an unnecessary contest with the machine. It is a person who can use strong tools, recognize their limits and carry a growing share of the responsibility. That person deserves a route into work where those abilities can develop on paid time.

Sources and limits