The homogenization problem is real

Artificial intelligence is often discussed as if it has one inevitable effect on human creativity. It doesn’t.

The same AI can make one person more dependent on conventional answers while helping another person explore ideas that would have been difficult to reach alone.

This creates a more interesting hypothesis:

AI may homogenize predictive users while amplifying generative users.

That is different from saying that AI is simply “good” or “bad” for creativity.

A 2025 study by Kibum Moon, Adam E. Green and Kostadin Kushlev examined the effect of GPT-4 on collective creative diversity. Across three preregistered studies, the researchers analyzed 2,200 college-admission essays, comparing human-written essays with GPT-4-generated essays.

They developed a measure called diversity growth rate: rather than simply asking whether an individual essay was creative, they measured how much new diversity each additional essay contributed to the collective pool.

The result was striking: each additional human-written essay contributed substantially more new ideas than an additional GPT-4 essay. Across the studies, human essays increased collective semantic diversity by roughly 2–8 times more than GPT-4 essays.

But the viral interpretation needs an important correction. It does not mean that a student without ChatGPT is individually “8× more creative.” The 2–8× figure concerns collective diversity growth, not an individual’s creativity score.

The paradox: better individual outputs, narrower collective space

Imagine 1,000 people answering the same creative question without AI. Some answers will be excellent, some terrible, some strange, some boring, and some will contain an unexpected idea nobody else considered. That messiness is actually valuable.

Now give all 1,000 people access to the same powerful language model and ask it to help them. Individual answers may become better written, more coherent, more structured, more complete and more persuasive. But they may also begin moving toward the same answer-space.

The result is a paradox:

Individual quality can rise while collective diversity falls.

The researchers describe this concern at the collective level: LLMs may enhance individual creativity while reducing the diversity of creative outputs across groups.

AI is a mirror

The mistake is assuming that everyone uses AI in the same way.

They don’t.

1. The predictive user

The predictive user asks: “Give me the answer.”

AI supplies an answer. The user accepts it, perhaps edits it slightly and moves on. This is extremely efficient, but the process is fundamentally convergent. The user is asking the machine to determine what the likely good answer is.

AI becomes a prediction amplifier.

2. The productive user

The productive user approaches AI differently: “Here is my idea. Improve it. Attack it. Find alternatives. Test it against X. What am I missing?”

Now the human has supplied direction. AI becomes a tool for iteration. The user is no longer outsourcing the entire thinking process. They are using AI to increase the number of operations they can perform on an idea.

AI becomes a production amplifier.

3. The generative user

The generative user explores. They jump between domains, notice strange similarities, ask whether unrelated things might have the same underlying structure, and pursue an idea before knowing exactly where it is going.

Here AI can behave very differently. Instead of narrowing the search toward the most probable answer, the user deliberately uses the model to expand the search space.

AI becomes an idea multiplier.

The same AI can therefore produce opposite effects

Predictive use → convergence

Productive use → expansion through iteration

Generative use → exploration, recombination and divergence

The technology hasn’t changed. The user has.

And that may be one of the missing variables in the current debate about AI and creativity.

The gym analogy

Think about a gym. Give 100 people exactly the same gym. They won’t become equally fit.

One person walks on the treadmill for ten minutes. Another follows a structured training program. Another experiments with different exercises. Another becomes obsessed with understanding biomechanics and designs an entirely new training system.

Same equipment. Different extraction.

AI may work the same way. The important variable isn’t simply access to the machine. It is what the person does with the machine.

AI is a sea of patterns

AI is a sea.

It contains an enormous amount of information, patterns, associations and possible formulations. But access to the sea doesn’t determine what you obtain from it.

One person takes a cup. Another casts a net. Another goes deep underwater. Another builds a submarine.

The resource is the same.

Extraction is different.

That is why “AI makes people creative” is too simplistic. So is “AI destroys creativity.” AI provides an enormous computational and linguistic environment. What emerges depends partly on the questions, intentions, curiosity and cognitive habits of the person using it.

The outlier problem

If most people use AI primarily as a prediction engine, widespread AI adoption could increase homogenization.

But people who already think divergently might use exactly the same technology in the opposite direction. They might use AI to challenge assumptions, generate alternatives, cross boundaries between disciplines, search for hidden connections, test strange hypotheses, attack their own ideas, revisit abandoned ideas, and combine ideas that normally live in separate intellectual territories.

For such users, AI isn’t necessarily a substitute for creativity. It can become infrastructure for creativity.

The human provides the unusual direction. The machine provides the enormous capacity to explore it.

The real divide may not be human versus AI

Perhaps the future divide won’t be Humans vs AI.

It may be:

Predictive humans using AI predictively
versus
Generative humans using AI generatively.

The first group may increasingly converge toward the same machine-mediated answers. The second may use the same machine to increase the number and range of ideas they can investigate.

That produces a fascinating inversion:

The same technology that homogenizes the crowd may disproportionately amplify the outliers.

That is a hypothesis, not an established scientific conclusion. The existing research demonstrates the homogenization phenomenon in collective GPT-4 writing. It does not yet establish that highly generative individuals systematically benefit more from AI than predictive users. That part requires direct testing.

A better question for the AI era

Instead of asking, “Does AI homogenize creativity?”, we should perhaps ask:

Does AI homogenize people who use it predictively while disproportionately amplifying people who use it generatively?

That question changes the entire discussion.

It moves us away from treating AI as a single force acting upon passive humans. Humans are not passive. They bring different cognitive strategies to the same machine.

And if that is true, the most important AI skill of the future may not be prompting.

It may be knowing what to extract.

Because ultimately:

It’s not what the sea has.

It’s what you can extract from it.

Source: Moon, K., Green, A. E., & Kushlev, K. (2025), Homogenizing effect of large language models (LLMs) on creative diversity: An empirical comparison of human and ChatGPT writing, Computers in Human Behavior: Artificial Humans, 6, 100207.