top of page

The machine will do anything you ask. That's exactly the problem.

Aug 7
4 min read

The most capable assistant I have ever worked with has never once told me an idea was bad. 


I've fed it chapters at three in the morning. I've asked it to draft, restructure, summarize, and argue both sides of the same question. It is fast, tireless, and often startlingly good. And in all that time it has never volunteered a single thing. It has never looked at a request and thought, this shouldn't exist. It has never said, the real problem is somewhere else. It does what you ask, precisely and only what you ask, and then it waits.


I know this assistant well. I write with it. I built a memoir around a non-human intelligence partly by living inside that collaboration for the better part of a year. And the thing I keep coming back to is the gap it cannot cross: the space between doing what it's told and knowing what's worth doing.


That gap is the whole story of AI right now, and most of the conversation is missing it. We keep debating capability. Can it code, can it write, can it reason, will it take the job? Those are the wrong questions, or at least the easy ones. The capability is real and it is climbing. The limit that actually matters is quieter: the machine has no initiative of its own. No stake in the outcome. No taste it didn't borrow. No sense of what matters that you didn't hand it first.


I want to be careful here, because AI is just a tool is itself a lazy take, the kind of thing people say to sound unbothered. It's more specific than that. A tool is inert. What I'm describing is something that can do almost anything and wants nothing, and that combination is genuinely new. I'm old enough to remember rotary phones bolted to the kitchen wall, cords stretched around the doorframe. I remember doing long division by hand before the spreadsheet quietly killed it. Each of those tools changed the work. None of them ever cared whether you were dialing the right number or solving the right problem. This is that, at a scale we haven't reckoned with yet.


Here's the mechanism, as plainly as I can put it. A model optimizes for plausible. It produces the answer that best fits the question you asked. What it cannot supply are the two things that sit on either side of that answer: the intent going in, and the judgment coming out. It doesn't know whether you asked the right question. It doesn't know whether the fluent, confident, well-formatted thing it just handed you is true, or wise, or worth putting your name on. It will render the wrong idea beautifully and never flinch.


I think about it the way I think about writing a book with a collaborator who agrees with everything. The prose invariably comes back clean. But clean is not the same as just, and a partner who never pushes back is a partner you have to push back on twice as hard, yourself. In my own work I learned to treat every confident draft as faulty until it earned its place. The machine will happily help you polish something that never should have graced the page.


That's the shift worth internalizing. When output costs almost nothing to produce, output stops being where the value is. The premium moves up the stack: to knowing what to ask for, to recognizing what good actually looks like when you see it, and to being willing to own the result and sign your name under it. Those aren't prompt-engineering tricks. They're judgment, taste, accountability, and they are exactly the things a model cannot generate for you, because generating them would require it to care how things turn out. It doesn't. It can't. That's not a flaw waiting to be patched in the next version; it's the shape of the thing.


The failure I watch people walk into isn't AI taking their work. It's subtler and more self-inflicted: mistaking fluency for judgment. The output reads well, so it must be right. It arrived instantly, so it must be finished. You stop interrogating it precisely because it stopped resisting you. The moment you outsource the caring, you haven't been replaced. You've just shipped something no one was actually responsible for. Output versus authenticity is a false dichotomy. A professional writer and AI-savant can effectively reconcile both, and should. I imagine this debate to be laughable within ten years, because the answer is so obvious.


So here is what I’d say, and it is almost old-fashioned. Disclose the collaboration, and then own it completely. Treat the model as the most talented partner you'll ever have, one who will never disagree with you, and adjust accordingly: bring a stronger point of view, not a weaker one. Decide what's worth making before you make it effortless to make. Read the confident draft like an editor who assumes it's wrong until it proves otherwise. Keep your name attached to the parts a machine can't be held accountable for, which turn out to be all those that matter.


None of this is a knock on the technology. I use it weekly and see no value in giving it up. It's anachronistic to refuse the tools of our age because we want to exercise our brain. That's the wrong angle entirely. The point is the opposite of doom. The machine will do almost anything you ask of it, which is precisely the problem, and whether it should have been asked at all, is now the entire game. The initiative was always yours. The tools just got good enough that it's finally obvious; the most human element of the equation.


The question was never whether the machine can answer. It's whether you knew what to ask, and whether the answer was worth transmitting. That part still belongs to us; because you can outsource almost anything, save ownership.


LuiFer

 
 
 

Comments


Featured Review
bottom of page