Optimized but Unowned: Why the Plan an AI Wrote for You Is Easier to Quit

The plan was better than anything I would have written. Cleaner structure, more realistic timelines, a fallback for the weeks that go sideways. I abandoned it in nine days. It took me a long time to understand that those two facts were related.

I write software for a living, so I came to language models with an engineer's instinct: if a task can be automated well, automate it. Goal-setting looked like a task. It has inputs — where you are, where you want to be, what you can spare. It has an output — a plan. And the output a good model produces is genuinely strong. It asks better clarifying questions than most people would think to ask themselves.

So why does the plan you asked for die so much faster than the messy one you wrote at a kitchen table?

Researchers looking at this have started using a phrase that stuck with me: optimized but unowned. Goals generated on someone's behalf by an AI — even well-crafted, personalized ones that the person rates highly at the moment they receive them — get followed through on less than goals the person struggled to write themselves. The struggle wasn't inefficiency. The struggle was doing something.

What the Struggle Was Actually Producing

We tend to think of goal-setting as an information problem. You don't know what to do, so you find out what to do, and then you do it. Under that model, a better plan should mean better follow-through, and outsourcing the planning should be pure gain.

But the twenty minutes you spend staring at a blank page is not primarily information retrieval. It is a set of small, uncomfortable acts:

• You have to admit what you actually want, which is often not the respectable version.
• You have to look at what you have already failed at, because that's the evidence you're reasoning from.
• You have to choose, which means killing the four other things you also wanted.
• You have to name a number or a date, which converts a wish into something that can be wrong.

None of that shows up in the final document. All of it shows up in you. By the time you've written "three sessions a week" in your own handwriting, you've already run a private simulation of Tuesday evening, and some part of you has agreed to it. The plan is a receipt for a decision that already happened.

When a model writes it, the document is better and the receipt is missing.

Effort Justification, and Why We Value What Costs Us

There's a long thread in social psychology here. The classic experiments on effort justification found that people who went through something difficult to join a group later rated that group more favourably than people who joined easily. The going through was itself a source of the valuing.

The furniture version of this is better known: people assign higher value to things they assembled themselves than to identical pre-assembled objects. Not because self-assembly improves the shelf. Because effort attaches you to the outcome.

There's a mechanism underneath both, and it's slightly uncomfortable: we infer what we believe from what we did. Having spent a difficult hour on something, the mind resolves the dissonance in the cheapest available direction — this must matter to me. It's not a flattering account of human motivation. It is a reliable one.

Now consider what happens with a plan you didn't build. There's no invested cost to justify. Nothing needs explaining. When Wednesday arrives and you don't feel like it, the internal argument is over before it starts, because there's no evidence on the other side. Quitting is free.

Worse, a well-made external plan gives you an exit that feels reasonable: the plan was unrealistic for my life. When you wrote it yourself, that same sentence indicts you. When a model wrote it, it indicts the model. We are very good at finding the door that doesn't cost us anything.

The Older Research This Sits On Top Of

None of this should have surprised anyone, because self-determination theory has been making the argument for decades.

Deci and Ryan's work identifies three needs that sustain motivation: autonomy (I chose this), competence (I can do this), and relatedness (this connects me to people I care about). Of the three, autonomy is the one that outsourcing quietly damages.

Their research distinguished between goals that are genuinely yours and goals that are what they called introjected — goals you've swallowed from outside, that sit inside you without ever being digested. Introjected goals still produce action, for a while, powered by guilt or obligation. They just don't survive friction. The moment something costs you, an introjected goal loses the argument to almost anything.

An AI-generated plan is a remarkably efficient way to manufacture an introjected goal. It arrives fully formed, sounds authoritative, and is calibrated to what you said you wanted. It feels like yours. It has the shape of yours. But nothing in you had to move to produce it.

There's also a competence cost, which gets less attention. Part of what makes a plan stick is having built the model of yourself that produced it — knowing that you overcommit on Mondays, that you'll skip anything scheduled after 8pm, that you need the gym bag by the door. Get the plan handed to you and you never build that model. So when the plan breaks, and every plan breaks, you have nothing to repair it with. You go back and ask for another one.

Why "Make Me a Plan" Is the Wrong Prompt

The failure isn't the tool. It's which part of the work you handed over.

Goal-setting has roughly four stages: figuring out what you actually want, deciding what to pursue, designing how to pursue it, and adjusting when reality disagrees. A language model is genuinely excellent at the third stage and can be useful in the fourth. Stages one and two are exactly where the ownership is generated — and they are the ones "make me a plan" quietly absorbs, because you cannot design a plan without having chosen a goal, so the model chooses for you and you agree, quickly, because what it picked was reasonable.

That quick agreement is the whole problem. Agreement is not the same act as decision. Decision has a cost. Agreement doesn't.

I notice this most on the days I'm in a hurry. When I want a plan by the end of the coffee, I ask for the plan. When I have an actual half hour, I end up arguing with the model instead, and the arguing version is the one still alive a month later.

Co-Writing a Goal Without Giving Away the Deed

This isn't an argument for refusing the tool. It's an argument about sequence. Four steps, and the order is the entire point.

1. Write it badly, by hand, before you open anything. Ten minutes. What you want, why you want it, what you've already tried, what you think will go wrong. Do not organise it. The badness is not a problem to be fixed later — the unstructured version is the honest one, and it is the artifact you're protecting.

2. Choose before you consult. Pick the one thing. Write the number and the date yourself. This is the step people skip, and it is the only irreplaceable one. If you let the model help you decide what to pursue, you have given away the part that would have held you.

3. Now bring it in — as an interrogator, not an author. Paste your messy version and ask it to work on it, not to replace it. Prompts that keep ownership on your side of the table:

• "Here's my goal and my plan. What are the three most likely reasons I'll quit in week three?"
• "What am I assuming here that I haven't checked?"
• "Ask me five questions that would make this plan better. Don't rewrite it."
• "What does someone who succeeded at this do that isn't in my plan?"

Notice that every one of those returns something you then have to decide about. That's the test. If the output can be adopted without a decision, you've outsourced too much.

4. Rewrite the final version yourself. Not copy — retype, in your own words, including the parts you're taking from the model. This sounds like superstition. It isn't. Retyping forces you to pass every line through your own judgment, and you will quietly drop two or three things you never actually agreed with. Those are the lines that would have broken first anyway.

A test for whether you still own it: put the plan away and describe it out loud to someone. If you find yourself explaining why each piece is there, it's yours. If you're reciting steps, it isn't.

Where This Argument Runs Out

I want to be careful not to over-apply this, because there's a version of this idea that becomes an excuse to do everything the hard way for its own sake.

Ownership matters where persistence matters — where the plan will be tested by a bad week and has to survive it. Long-horizon, effortful, identity-adjacent goals: fitness, money, learning something difficult, changing how you work.

It matters far less for tasks that are merely tasks. Nobody needs to feel ownership over a packing list or a project timeline for a thing they were going to do anyway. Use the tool freely there. The distinction is whether the plan needs to survive you not feeling like it.

And there's a real cost on the other side. For some people, the blank page is not productive struggle — it's where the whole thing stops. If a model getting you to a first draft is the difference between starting and not starting, start. Just come back and rewrite it once it's running, so that the version you're following is one you've actually agreed to.

Questions Worth Asking

Is this saying I shouldn't use AI for planning at all?

No. It's saying the ordering matters more than the quality. Use it after you've chosen, as a critic and a stress-tester, rather than before you've chosen, as an author. The same model, in the same conversation, produces very different durability depending on which side of your decision it sits.

What if the AI's plan is genuinely better than mine?

It probably is, on paper. But a plan is not judged on paper — it's judged on whether it's still running in March. A slightly worse plan you'll actually follow beats a better plan you'll abandon, and this is one of the few places in life where that trade is so lopsided it isn't really a trade.

Does this apply to plans other people make for me — a coach, a manager?

Partly, and the research on introjected goals says the same thing. But a human coach usually does something a prompt doesn't: they make you defend your choices, they notice when you're agreeing too fast, and there's a relationship attached that supplies its own motivation. That last part is a genuinely different force, and a model doesn't have it.

How do I know if a goal is really mine?

One question, asked honestly: if nobody ever found out whether I did this, would I still want it? Goals that survive that question tend to be autonomous. Goals that evaporate under it were probably assembled from other people's expectations — and a language model, trained on all of us, is very good at producing exactly that kind of goal.

I still use the model. I just make it wait now, which is a strange thing to say about software. The plan on my desk is worse than the one it would write, and I'm eleven weeks into it, which is nine weeks longer than I've ever managed with a better one.