Most of this guide is about getting more out of AI. This page is about the times when getting more out of AI is the wrong goal. Sometimes the point is the work itself. Sometimes the point is the relationship the work is part of. AI is bad at noticing the difference, so this part is yours.

Part one: when the point is to struggle

The reason AI feels like a free upgrade for most tasks is that, for most tasks, the output is the point. A summarised council letter, a tighter complaint email, a shopping list, a holiday itinerary. You wanted the result; AI got you to the result faster. Win.

For some tasks, the output is not the point. The struggle is the point. Outsource the struggle to AI and you have not made the work easier. You have skipped the work entirely.

Kids' homework

Homework is not a quiz on whether the family can produce the right answer. It is reps for the child's brain. The maths exercise is for the maths. The book report is for the reading and the thinking and the figuring out what to say about it. Hand the question to AI and the child gets a polished answer with no maths and no thinking attached.

The rule that holds in our house and in most houses I have seen working: AI is allowed to teach. AI is not allowed to do. A child stuck on a fractions question can ask AI to explain fractions. A child stuck on what a metaphor is can ask AI to explain it with five examples. A child being asked to write a paragraph about a metaphor in their book is being asked to do that themselves, and AI doing it for them is the cheating version. The line is "explain the concept" on one side, "write me the answer" on the other. Most kids understand the line clearly the moment you draw it.

One genuine negative: AI is so cheerful about producing the answer that "explain it" can drift into "give me the answer with a bit of preamble". You have to actually read what came back. If the AI's "explanation" includes the worked solution to the actual question, that is not an explanation. Send the child back to ask again.

Creative hobbies you are doing for the love of it

If you took up watercolours to slow down, do not let AI do the design work for you. The point is the doing. The painting that comes out is a record of an afternoon you spent paying close attention to the light on the leaves. AI cannot give you that afternoon. It can give you a finished image of leaves, which is a different thing entirely.

The same applies to woodwork, knitting, baking, photography, gardening, anything you took up because the activity itself is the reward. AI is fine for the boring research bits ("what is the difference between merino and alpaca for a baby blanket"). It is not fine as the substitute for the doing. If you are not sure which side of the line you are on, ask yourself: would I be disappointed if AI did this for me and the result was identical? If yes, do it yourself.

Fitness and habit goals

AI is fine for a starter plan. "I am 55, sedentary, no injuries, three sessions a week, want to be able to walk five kilometres without stopping. Build me a 12-week plan." The plan that comes back is reasonable, individualised enough, and a much better starting point than nothing. The work, of course, is yours. The plan does nothing on the days you do not do it.

The trap to avoid is asking AI to do the work that is meant to be hard. Logging the workouts, deciding to go, getting out of bed. There is a temptation to spend more time tweaking the AI plan than doing the plan. The tweaking feels productive. It is not. The first plan is good enough for the first month. The doing is the part that compounds.

Learning a language

AI is the best tutor in the history of language learning. It is also the best way to fake the work and learn nothing. The same tool that can drill you on Italian verbs in conversation, correct your pronunciation, and roleplay a barista in Bologna can also translate the entire Duolingo exercise for you in two seconds.

The rule: use AI to practice, not to translate everything for you. The struggle of trying to remember the word, getting it half wrong, being corrected, and trying again is what builds the language in your head. Skip that loop and you have not learned anything; you have just had a conversation through a translator. The longer version, including specific prompts that keep you doing the work, is on the Language Learning page.

Writing in your own voice

If you want to sound like you, draft it yourself first and then ask AI to tighten it. Reverse the order and you sound generic. The first sentence the AI writes is always slightly nobody's: a careful average of how every other person on the internet writes, with the texture sanded off.

The order matters. You first, AI second. Write a rough version with your own words, your own rhythm, your own jokes. Hand it to AI for tightening, for reordering, for cutting the bit that is not pulling its weight. The result still sounds like you, only sharper. The reverse, AI first and then you tweaking, almost always ends up sounding like a polished press release written by nobody in particular.

Part two: AI etiquette at home

The other place AI gets you in trouble is around the people who know you. Not because AI is dangerous, but because the way you use it changes the relationship in small, hard-to-name ways. A handful of habits make this easier.

Be open about it

If you used AI to draft a heartfelt message, say so once. "I drafted this with some AI help and then rewrote it myself." Most people are completely fine with that. Most people are not fine with finding out later. The "finding out later" version turns a thoughtful gesture into a small betrayal, even though the message itself has not changed.

The same goes for work emails, community-group communications, and apologies. A single line acknowledging the help removes the awkwardness entirely. The flip side: if you genuinely rewrote it in your own voice and the AI was a thinking partner rather than a ghostwriter, you do not need to disclose AI any more than you would disclose a thesaurus. The test is whether the words at the end are yours.

Do not let AI write the relationship

Apology messages, sympathy notes, and birthday speeches need your fingerprints on them. The whole point of saying sorry, or saying you are thinking of someone, or marking a birthday, is that you took the time and that the words are yours. A polished AI-written apology that arrives in someone's inbox lands wrong even when the recipient cannot put their finger on why. Something is off.

AI as a thinking partner is fine here. "Help me think about what to say to my brother after the argument." "I want to write a sympathy note to a friend whose father has died, and I am terrible at this. What kinds of things do people find helpful?" Use the conversation to clarify your own thoughts, then write the actual message yourself, in your own words, even if those words are clumsier. The clumsiness is the human bit. Strip it out and you have stripped out the part that mattered.

Read the room

Family group chats, community-group emails, kids' school WhatsApps, the social fabric of all the small groups you belong to. The reason these messages work is that they sound like you. Everyone in the chat has a sense of what you sound like, even if they could not describe it. AI-polished prose can sound off in a way they cannot quite name, and over time it accumulates as a vague feeling that something is different about you.

Send what you would actually say. If you would type two short sentences and a typo, send two short sentences and a typo. Do not run a quick reply through AI to make it sound more articulate; the more articulate version is less yours. Save AI for the messages where you actually want help and where the polish is welcome: a long email you have been putting off, a tricky note to a teacher, a complaint that needs to be measured rather than ranty.

A short version

The questions to ask yourself, when you are not sure whether to reach for AI: is the output the point, or is the doing the point? If the doing is the point, AI is the wrong tool. Are these words from me, or to be from me? If "to be from me" but actually from AI, the answer to "should I send this" is usually no, or at least not without a rewrite. Would the person reading this still feel I had taken the time, if they knew? If yes, fine. If no, take the time.

The hard rules on what you should never put into AI in the first place are on Rules That Matter. The recipes for verifying what AI tells you are on How to Check What AI Tells You. And the longer thinking on AI and creative work, including the difference between using AI well and using it lazily, is on AI and Creativity.