Looking for Answers Can Expose the Question

We usually imagine that looking for an answer starts with knowing the question.

Sometimes it does. If you know exactly what you need, describing it clearly is useful. You can say what has to be true, what cannot change, and what a good result needs to do.

But there is another kind of problem.

Imagine you're rearranging a room in your house. You know the room. You know what has to stay, what gets in the way now, how much space you have, and roughly how you want the room to feel. But if someone hands you a blank sheet of paper and asks you to describe the exact layout you want, you may not know.

That doesn't mean you know nothing about the answer.

Put three plausible layouts in front of you and you may have opinions immediately. That blocks the window. This wastes the corner. I hadn't thought about putting the desk there.

The examples didn't give you knowledge of the room that you were missing. They made differences visible that were difficult to consider while everything was still abstract.

That can make the usual advice—be more specific—surprisingly unhelpful. More detail only helps if you know which details matter.

Sometimes you don't know yet.

Something changes when you can see the possibilities

Generative AI makes it unusually cheap to create things to think against.

Instead of trying to describe the finished answer from a blank page, you can make several plausible versions concrete enough to compare. A vague reaction can become specific. Two choices that sounded nearly identical in words can feel completely different when they are sitting next to each other.

You still haven't magically discovered a perfect answer. But you may have discovered something smaller and more useful: a difference that matters.

That difference gives the next question somewhere to go.

Maybe the problem was never the color; it was the hierarchy. Maybe what sounded like a request for simplicity was really a request to make one action unmistakable. Maybe none of the examples are right, but one of them reveals a relationship that had never occurred to anyone before.

This is one reason generative AI can be useful even when its answers are wrong.

It can make possibilities concrete cheaply enough that we can think against them.

Research on difficult-to-describe preferences has started to show a related effect. In experiments published at ICLR in 2025, people using an interactive language-model system produced more informative descriptions of their preferences in two of three studied tasks, and reported encountering considerations they had not anticipated.¹

The interesting part is not that the model somehow knew what those people wanted.

It gave them something to work against.

And you don't have to choose one of the examples.

Sometimes none of them are right. One might simply show you something you hadn't thought about before—a relationship, an arrangement, a distinction worth exploring.

You can throw away the answer and keep what it taught you.

Then the next search can begin from there.

The search can shape what you want

There is a catch.

If the examples help you notice what matters, then the examples you are shown matter too.

Imagine going back to the room. This time, every suggested layout puts the desk against the same wall. After seeing ten versions of that arrangement, moving the desk somewhere else may stop feeling like an obvious possibility—not because it would be a bad idea, but because it never made it into the choices you were reacting to.

Generative AI can do the same thing. Ten images can look different while sharing the same basic assumptions. A polished version can make one direction feel more inevitable than it really is.

Researchers have seen this problem in visual ideation. In a CHI 2024 experiment, people working with AI-generated images became more fixated on an initial example and produced fewer ideas with less variety and originality than the comparison group.²

That changes how I think the search should work.

If the first batch all seems to come from the same family, don't just ask for ten more. Change the direction. Ask for something that makes a different assumption. Keep the original problem nearby so the generated versions do not quietly replace it.

And when something catches your attention, ask what actually caught it.

Was it the whole design? Or was it the way one thing was grouped? The amount of empty space? A relationship between two parts? A possibility you had not considered?

The more precisely you can name what survived, the less of the generated answer you have to carry forward.

None of this makes the search neutral. The examples can still influence what feels normal, attractive, or possible. The point is to remember that the menu is being generated too.

Sometimes the search comes first

This does not mean every problem should begin with a pile of AI-generated options.

If you already know a requirement, say it. If something cannot change, protect it. If the question depends on a fact, go find the fact. Exploration is not a substitute for knowing what is already knowable.

But some questions become clearer only after you have something concrete to push against.

You start with a rough sense of the problem. You look at a few possibilities. One exposes a difference you had not noticed. You keep that difference and throw the rest away. The next search is better because the question is better.

Eventually, another round stops teaching you anything. Then it is time to leave the search and try something in the real world.

But before that point, looking for an answer can do something more useful than finding one.

It can help you discover what you were really asking.