The forgotten skill: knowing when not to ask AI
Some questions are worse for having been asked of an AI at all. A short list of categories where the human answer beats the model answer reliably.
The most under-used AI skill in 2026 isn't a better prompt or a smarter model choice. It's the discipline of not asking. There are categories of question where the human answer beats the model answer reliably, and asking the AI anyway produces work that's worse for having involved the AI at all. This essay is the short list — five categories where the right move is to close the chat window and think.
1. Questions about your own taste
"What should I name my company?" "What should the cover of my book look like?" "How should this paragraph sound?"
The AI can produce competent answers to all of these. It cannot produce YOUR answer to them, because your answer comes from you. Asking the AI here is outsourcing taste — and once your taste runs on AI defaults, your work stops being yours.
2. Questions about a specific person
"Should I email Sarah back the way I drafted it, or soften it?" "What does my co-founder mean by saying X?"
The AI doesn't know Sarah. The AI doesn't know your co-founder. It will generate a fluent answer based on the abstract category "person who said X", and that answer will be worse than the answer you'd get by thinking for thirty seconds about who Sarah actually is.
The exception: drafting an email tone is fine if you'll edit it through your knowledge of Sarah. Asking the AI to make the final call about Sarah is not fine.
3. Questions where you already know the answer but want permission
"Should I quit this job?" "Should I have the hard conversation?" "Should I ship the imperfect version?"
You already know. You're asking the AI because you want a second opinion that might agree with the easier path. The AI is structurally bad at this — it's trained to be helpful and will often validate whichever direction you frame the question toward. The validation isn't free; it costs you the conviction you'd have built by sitting with the question.
If you want a second opinion, ask a person. The AI's second opinion is shaped by your prompt.
4. Questions about your own work, before you've reread it
"Is this draft good?"
If you ask this without reading the draft yourself once more, the AI's answer becomes your read. You will start to perceive the draft through what the AI said about it. This is bad — you should perceive it through your own re-reading.
After you've read it yourself and formed an opinion, asking the AI for a second pass is fine. Before, it's outsourcing the part where you actually engage with what you made.
5. Questions where the value is in the time spent thinking
"What do I really believe about X?" "What kind of life do I want to be living in five years?" "What's the right way to position my work?"
These are questions where the thinking is the point. The answer at the end matters less than the path you took to get there. An AI shortcuts the path — you get an answer that sounds reasonable in thirty seconds, and you lose the half-hour you would have spent figuring out what you actually think.
If thirty seconds was what the question deserved, you weren't going to think about it anyway. If it deserved half an hour, the AI is robbing you of the wrong things.
What this isn't
It's not "AI bad". AI is a useful tool for plenty of questions — most of the ones you'd ask in a normal day. The five categories above are specific exceptions. Outside them, ask away.
The skill is noticing when you're about to ask a question that falls in one of these categories and pausing. Sometimes the pause leads you to ask anyway, with eyes open about what you're trading. Sometimes it leads you to close the tab and think. Either choice is fine; the unconsidered choice — defaulting to "ask the AI" because that's what you do now — is the one that hollows out over time.
Where this fits
This is the most-cited essay-with-essay link in our Essays cluster, paired with the thinking is yours, the models do the typing. Both come from the same root claim — AI works best when you're deliberate about which parts you let it touch. The tools that respect this — see our comparison piece for ones that try, or oran.chat itself for one shaped around the idea — are the ones worth using.