You ask AI to write the email. It is clearer than the version you would have written in a hurry.
You ask it to explain a difficult idea. It gives you an explanation you can follow.
You ask it what you should do. It presents a tidy set of options, weighs them, and offers a recommendation.
All three requests can be useful. But they are not quite the same transaction.
In the first, you may be delegating expression. In the second, you may be receiving instruction. In the third, you may be handing over part of your judgment.
The difference becomes harder to notice when the interface is identical: type a question, receive an answer.
Of course we should use useful tools
There is no virtue in spending an afternoon formatting something that can be formatted in seconds.
AI can make information more approachable, help organise a messy project, and give someone a place to begin when a blank page has defeated them.
For somebody who struggles with reading or writing, assistance can expand participation. For an experienced person, it can remove work that contributes little to the outcome.
Keeping every task manual would be a strange way to defend human capability.
The useful question is what kind of assistance preserves or develops the capability you care about.
A finished answer is not the same as a learned skill
In a field experiment involving nearly a thousand high-school mathematics students, researchers compared a broadly accessible GPT-4 assistant with a tutor designed to protect learning.[1]
Both improved performance while students had access to them. But students using the unrestricted assistant performed worse than the control group when the assistance was removed. The learning-focused tutor largely mitigated that problem.
This was a particular educational setting, not proof that every use of AI weakens the mind. It does show why the design of assistance matters. Doing better with help and becoming better without help are different outcomes.
You can recognise the distinction in ordinary life. Following a solved example may feel clear. Solving a new problem alone reveals whether the understanding is yours.
Preserve the first attempt
If you want to become better at an activity, give yourself a first attempt before asking the machine to complete it.
Write the rough paragraph. Sketch the explanation. Work through the problem far enough to find where you get stuck.
Then ask for feedback, a hint, or a competing interpretation.
This does not require turning every email into a lesson. It means deciding which activities are practice and which are simply work to finish.
A coach preparing a class might use AI to organise timings after deciding what students need to learn. A student might request a question that tests understanding instead of a paragraph to submit. A manager might draft the difficult point in their own words, then ask whether it is clear and fair.
Choose the kind of help that fits the purpose. Finishing a task and practising a skill can call for different requests.
Change the request, change the practice
If the skill matters to you, try changing what you ask for:
- Instead of “Write my argument,” ask, “Here is my argument. Where is the reasoning weakest?”
- Instead of “Solve this problem,” ask, “Give me one hint, then wait for my attempt.”
- Instead of “Tell me what to do,” ask, “What assumptions am I making, and what evidence could challenge them?”
- Instead of “Explain it again,” ask, “Give me a new example to solve, then assess my explanation.”
These requests put more of the work back in your hands. They do not guarantee good teaching or accurate feedback. Check important claims, and treat a fluent critique as something to examine rather than automatically obey.
The useful test comes afterwards: close the tool and try. Can you explain the idea, make the next decision, or complete a similar task?
Keep responsibility attached to the decision
AI can help you compare choices. It does not live with the consequences.
That matters when deciding what kind of business to build, how to respond to somebody you care about, or which commitments to make.
The machine can generate reasons. You still need to decide whether they are true, relevant and yours.
Before accepting a recommendation, try explaining the decision without repeating the answer's phrasing. What are you prioritising? What are you giving up? What evidence would change your mind?
If you cannot explain those things, the tidy recommendation may have concealed an unfinished decision.
Be careful with the human part
A message of condolence, an apology and a performance review are all writing tasks. They are also acts within a relationship.
Assistance can help you find words. But there is a difference between finding words for something you mean and generating a feeling you have not taken time to understand.
Before asking for an apology, identify what you did and what you intend to change. Before asking for a difficult conversation script, decide what the other person deserves to hear from you.
You can delegate polishing. Accountability requires your participation.
Learn enough to check
An output you cannot evaluate leaves you dependent on the apparent confidence of its presentation.
Sometimes the right response is to learn more. Sometimes it is to involve a qualified person. Sometimes it is to use a simpler method you can verify.
You do not need expertise in everything. You do need a way to recognise when the stakes exceed your ability to check the answer.
That may be the most useful skill AI asks us to develop: knowing which assistance can be trusted for which purpose, and why.
Try this for a week
Choose one capability you want to retain or develop.
For that activity, make a first attempt, ask AI for critique rather than completion, and then try a similar task without it. Notice what you can now explain or do independently.
For another task that contributes little to your learning, delegate freely and check the result.
At the end of the week, compare the results. Did assistance help you understand the work, or mainly help you avoid it? Did the time you saved go somewhere you intended?
You do not need a rule against using AI. You need a reason for how you use it. Let it carry the work you have chosen to delegate, while making room to practise what you still want to call your own.
After the tool has helped you, what are you more capable of doing?
Sources
[1] Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. Study abstract. Published correction.
The study concerns the tested GPT-4 tools and mathematics setting. The applications to writing, coaching and judgment are GOOD QUESTION’s editorial interpretation.