AI Is Great at Brainstorming. That Might Be the Problem.
There is a particular kind of satisfaction in asking AI for ideas.
You type one sentence.
It pauses for barely a moment.
Then ten ideas appear.
No awkward silence. No staring at an empty document. No meeting where everybody says, “Let’s just throw things out there,” and then nobody throws anything out there.
It feels like creativity on demand.
And sometimes it genuinely helps.
But I’ve started noticing something when I use AI for brainstorming: the first ideas are often good enough to stop me thinking.
That is a strange problem to have.
Before generative AI, the blank page forced you to wander around a little. You remembered something somebody said. You connected two unrelated things. You wrote down a bad idea that accidentally led to a better one.
Now you can skip straight to a polished list.
The question is whether skipping the messy part also skips some of the originality.
Better ideas, but more similar ones
There is research that makes this tension particularly interesting.
In a study published in Science Advances in 2024, researchers asked people to write short stories. Some participants could use ideas generated by GPT-4, while others wrote without AI assistance.
Access to generative AI improved how the stories were rated for novelty and usefulness, particularly for writers who had initially scored lower on creativity.
That sounds like a clear win.
But the researchers found another effect: AI-assisted stories became more similar to one another.
In other words, generative AI could improve individual creative output while reducing the diversity of the group’s output.
I think that distinction matters far beyond short stories.
Imagine a thousand people asking roughly the same model:
“Give me ten SaaS ideas.”
“Give me ten LinkedIn hooks.”
“Give me names for my startup.”
“Give me content ideas for an AI newsletter.”
The model can produce variation.
But it is generating from patterns it has learned, guided by prompts that are often remarkably similar.
If we all take the first acceptable answer, we shouldn’t be surprised when the internet starts feeling strangely familiar.
Don’t ask for ideas. Create a search space.
I still brainstorm with AI.
I just try to avoid treating “generate 20 ideas” as the whole process.
Suppose I’m thinking about a new productivity product.
I could ask:
“Give me 20 ideas for a productivity app.”
I’ll probably get some combination of task management, habits, focus, calendars, goals, automation and collaboration.
Nothing necessarily wrong with those.
But I haven’t given the model much reason to go anywhere else.
A better starting point might be:
“Here are three things that frustrate me about productivity tools I’ve actually used. Generate ideas that attack each frustration from a different direction. Do not suggest task managers, habit trackers or calendar apps.”
Now there are constraints.
Then I can deliberately make the search stranger:
“What assumption do all these ideas share?”
“What if notifications were forbidden?”
“What would a logistics company do with this problem?”
“What would the opposite of these ideas look like?”
“Combine the most practical idea with the weirdest one.”
“Which five ideas feel most predictable? Remove them and continue.”
That is a different relationship with the model.
I’m not asking it to hand me creativity.
I’m using it to move around a space of possibilities.
Your unfair advantage is the stuff the model didn’t live
There is another ingredient I think matters even more: your own experience.
AI has access to patterns across an enormous amount of language.
It does not have your Tuesday.
It wasn’t in the meeting where everybody misunderstood the same feature.
Get Joshua Idunnu Paul’s stories in your inbox
Join Medium for free to get updates from this writer.
It didn’t hear the customer explain the bizarre workaround they invented.
It didn’t spend three hours debugging the problem that turned out to be one configuration flag.
It doesn’t know which part of your industry everybody quietly complains about but has accepted as normal.
Those details are useful because they are specific.
If I bring them into a brainstorming session, I’m giving the model material that is much less generic than:
“Give me business ideas for developers.”
This is why I think one of the worst ways to use AI creatively is to arrive empty-handed every time.
Bring something.
An observation.
A constraint.
A failed attempt.
A customer complaint.
A belief you want challenged.
Two unrelated ideas you suspect might connect.
Even a bad idea you can’t quite stop thinking about.
Then let AI push against it.
Generate first. Judge later.
Another mistake is asking AI to open and close the brainstorming process at once.
“Give me ten ideas and tell me which is best.”
That sounds efficient.
It also means the same system that generated the possibilities immediately narrows them.
I’d rather separate those jobs.
First, diverge.
Generate ideas from several frames. Force differences. Explore constraints. Ask for opposites. Combine things that don’t obviously belong together.
Then stop generating.
Look at what you have yourself.
Which idea keeps pulling your attention back?
Which two become more interesting when combined?
Which suggestion is based on an assumption you know is wrong?
Which “bad” idea contains one useful piece?
Add your own ideas to the pile.
Only after that would I bring AI back in to critique, compare or stress-test the shortlist.
That little gap gives you room to be more than the person approving AI output.
Don’t let the first page look finished
This may be the bigger lesson behind all of this.
Generative AI is exceptionally good at producing plausible first drafts.
That applies to writing.
Code.
Design directions.
Plans.
And ideas.
The danger of a plausible first draft is that it looks much closer to finished than the messy thing that would have existed without it.
So we stop earlier.
Not because the idea is excellent.
Because the screen no longer looks empty.
I think good brainstorming with AI requires resisting that feeling.
Treat the first ten ideas as raw material.
Ask what is missing.
Push toward places the model hasn’t naturally gone.
Inject things from your own life.
Change the constraints.
Challenge the assumptions.
Let one idea collide with another.
And occasionally close the chat and think without it.
AI can dramatically increase the number of directions available to you.
But more options are not automatically more originality.
The interesting part still happens when somebody makes an unusual connection, notices something others missed or refuses the obvious answer.
That somebody can still be you.
