I Don't Want AI to Do Everything for Me
I'm more interested in AI that helps me think better. The difference between delegating thinking and strengthening my own, with examples from research, planning, critique, writing and decisions.
When I talk about artificial intelligence, the most common question is what it can do instead of me. Write the text, prepare the quote, put together the plan, answer the e-mail, do the analysis. It can do all of that, at least at first glance.
But I keep finding that "do everything" is the worst way I use this tool. Not because the result wouldn't be usable, but because along the way I lose the thing that matters most to me in my work: understanding why something is the way it is. If AI makes a decision instead of me, I still have to defend it. If I don't know what it rests on, I haven't made it; I've just taken it over.
I'm more interested in a different use. AI that helps me think better, not AI that thinks instead of me. The difference sounds small. In practice it determines almost everything.
Delegating thinking versus strengthening it
Let me try to describe the difference simply. When I delegate thinking, I hand over a task and take the result. When I strengthen my own thinking, I hand over a problem, my current solution and a question: where am I wrong?
The first way gives me something that looks finished. The second gives me a list of holes, alternatives and questions I hadn't seen. The first is faster. The second is more useful whenever the decision means something.
I've written before about using AI in development; there it's largely about code. Here I'm interested in the wider picture: research, ideas, planning, comparisons, writing, business decisions. The same pattern shows up everywhere.
Research: a faster start, not a final answer
With a new topic, AI helps me get a map quickly. Which concepts matter, where the main differences between approaches are, what I should read. It shortens the beginning, which I used to spend juggling ten browser tabs.
I don't use it as a source of facts, though. If I need a figure that can change, or a claim a decision will rest on, I go to the primary source. AI tells me where to look; it doesn't tell me what's true. When I forget that difference, I get a convincingly written thing nobody has checked.
The most useful research question for me is: what don't I know about this topic yet that I should know before I decide? The answer is often better than the answer to what the right solution is.
Ideas: quantity is cheap, choosing isn't
AI can list twenty ideas for a feature, an article or an offer in a minute. Most will be average, some bad, one or two perhaps interesting. That's not a flaw; brainstorming with people works the same way.
The problem starts if I treat twenty ideas as work done. The work is only beginning: which idea solves a real problem, which one can I actually build, which one fits what I want to do. The tool doesn't know that, because it doesn't know my constraints, my customers or my taste.
So I prefer to use AI the other way round. I give it one idea I like and ask it to try to knock it down. What assumptions are in it? For whom wouldn't it work? What would have to be true for it to make sense? If the idea survives that, it's worth more time.
Critique: the best use I know
If I had to pick the one use that gives me the most, it's critique of my own work. I write a text, a plan or a proposal and ask for the harshest possible review. Where is the argument weak? What will the reader misunderstand? What's missing? Where am I confident without reason?
This works because in my own work I see what I meant to write, not what I wrote. An outside view shows the difference. With a colleague that's valuable, but I can't ask every time. A tool I can.
One limitation I have to keep constantly in mind: AI likes to criticise what's easy to criticise, and likes to praise unless I steer it otherwise. So I ask narrow questions. Not "what do you think of this", but "which three claims in this text would a reader find easiest to refute".
Planning: it helps structure, not decide
When planning a project, AI helps me see steps, dependencies and things I've forgotten. If I describe what I want to do, it can tell me which step is missing, what should be done first and where things usually get stuck.
I don't leave it the decision about what matters in the plan, though. That depends on goals only the person running the project knows. A plan made entirely by the tool looks complete and often doesn't correspond to anything we'll actually do.
So I always write the plan myself first, even if it's rough. Then I hand it over for review. The result is my plan with fewer holes, not someone else's plan that I first have to understand.
Comparisons: a good framework, suspicious scores
When comparing two tools, approaches or providers, AI quickly builds me a framework: which criteria to compare on, what the typical strengths and weaknesses are. That's useful, because it saves the first hour of work.
The scores in the table I take with a pinch of salt. Product data changes, differences between versions matter, and the tool likes to present them with the same confidence whether or not they've been checked. In a comparison a decision will rest on, I verify every important cell myself.
What a comparison really gives me is a list of questions I need to settle. What do we actually need? Which criterion is decisive? Those are questions for me, not for the tool.
Blind spots: ask what you can't see
One use I value more and more is finding my own blind spots. I describe a decision I've already made and ask: what would someone who disagrees say? Which risks have I overlooked? What would go wrong a year from now?
It's not pleasant reading, but it's useful. Not because the AI is right, but because it forces me to think about things I skipped in my enthusiasm. Often I find the decision holds. Sometimes I find it doesn't.
Writing: a draft yes, the voice no
With writing, AI helps me get past the blank screen and organise thoughts that are scattered. It can suggest a structure, point out repetition, shorten a paragraph that runs too long.
I don't let it write the text for me when that text is supposed to say something. The result is always smooth and always slightly empty. It lacks what makes a text worth reading: a stance, a concrete experience, a clear "no". This post, too, was written by first writing down what I think and only then asking for a review.
If a text could have been written by anyone, AI will write it too, and the reader will notice.
Business decisions: a sparring partner, not an adviser
For decisions that cost something, I use AI as a sparring partner to play through the options. What happens if I choose A? What are the consequences of B in a year? What should I know before I decide?
It doesn't make the decision. And I don't let it make the decision easier by confidently recommending one option. When it does that, I ask what the recommendation rests on. The answer often shows it rests on general principles, not on my situation. For general principles I don't need a tool; I need it to make my thinking about the concrete situation more thorough.
The questions I use most often
Since all of this comes down to questions in practice, here are the ones I ask the tool most often. They're nothing special, but each is framed so that the answer strengthens my thinking rather than replacing it.
Which assumptions are in this proposal that I haven't written down? Who would disagree with this, and why? What would have to happen for this decision to be wrong? Which three claims in this text would a reader find easiest to refute? What's missing for this plan to be feasible? Which step have I skipped? What would someone who doesn't know this field ask? What would someone who knows it better than I do ask?
None of these questions asks the tool to do something instead of me. Each asks it to show me something. The difference in the result is considerable: after a conversation like that I have a better plan that is still mine, not someone else's plan I first have to understand and defend.
I notice something else, too. When I use the tool this way, I stay in the role of the one who decides, and I get a little better at that role with every conversation. When I use it as a contractor, that role slowly slips out of my hands, and one day I find I know less about my own project than the tool I'm building it with. That seems like a bad deal to me, however much time it saves.
What's left for me
If I add all this up, I use AI for what makes it good: speed, breadth, tireless questioning. For myself I keep what the tool can't do: judging what matters, taking responsibility for the decision, and understanding what the decision rests on.
That's not a principle born of caution. It's the simple observation that my work is worth as much as my judgement is worth. A tool that strengthens judgement is precious to me. A tool that replaces it makes me dispensable. I'd rather have the first.
