AI Gives Correct Answers, But Rarely Makes Weird Inventions

I’ve been using AI constantly lately.

Writing code with it, talking through designs, bouncing blog ideas off it.

Honestly, it’s incredibly useful.

But the more I use it, the more I think:

AI is amazing at “honor student answers,” but not that good at “weird inventions.”

AI Answers Score Around 80

For example, when you ask AI to “design a user management feature,”

it returns a textbook-perfect answer: layered architecture, DI containers, repository pattern…

It’s correct. It’s fast. It won’t break.

But you almost never get back:

“Wait — does this product even need user management? Couldn’t we just use Google login?”

That kind of reframing rarely comes out.

AI scores 80 within a given frame — beautifully.

But it’s weak at questioning the frame itself.

AI Is Good at Extensions, Bad at Leaps

There’s a classic story from the birth of the automobile:

“If you asked customers, they’d want a faster horse.”

The point being, customers want better versions of what they already know.

AI is similar.

“Faster” “Cleaner” “More efficient”

It’s good at these.

But weak at leaps like “forget the horse — let’s build a car.”

In reality, the interesting products in the world usually emerged from weird paths:

• Slack: a failed game company’s internal tool became the product itself
• Twitter: spun off from a podcast company’s internal chat
• Airbnb: founders couldn’t pay rent, so they put air mattresses in their apartment
• YouTube: originally a video dating site

None of these are “best practices.”

They came from failures, detours, and weird obsessions.

Try to get AI to produce things like this — it just won’t.

Adding Noise Is the Human’s Job

What’s a bit scary lately is that when people adopt AI’s answers as-is, everyone’s designs start to look the same.

Not bad.

But everything gets averaged out.

And when everyone uses the same AI, the answers converge even more.

So in the end, someone needs to ask:

“Should we deliberately break this average?”

That part has to be human.

Use AI to generate the foundation fast, and humans inject the weird noise at the end.

Concretely:

• Don’t throw away the “something feels off” gut reaction to AI’s output
• Keep at least one weird option in the discarded pile
• Bring your own quirks, hobbies, and pointless detours into the design

This, I think, is what becomes the human’s job from here.

In the end, the role of humans in the AI era is:

bringing in “irrational discomfort” and “unintelligible ideas.”

Everything else can be handed over to AI.