Experiments show: moderate your AI use when writing and creating — not too much, not too little
- How much AI should you actually use for creative work? A study from the University of British Columbia gives a concrete answer: there's a sweet spot — use too little or too much, and creativity scores drop either way.
- The main experiment recruited about 150 people to brainstorm a startup idea for a student with just $10 in seed money, split into three groups: one ChatGPT prompt, four to six prompts, or at least nine. Expert judges scored the ideas on novelty, practicality, and commercial value — the middle group scored highest on average. Repeating the experiment with 319 people produced the same result.
- Professional creatives show the same curve: fashion designers, visual artists, writers, animators, tech workers, and influencers filled out a 1-to-7 questionnaire, and their managers separately rated their creative performance — the people who self-rated 4 to 5 performed best.
- Huang's hypothesis is that AI supplies a perspective outside your own experience, the same reason teams tend to be more creative than individuals. The author adds a few caveats: large language models are fundamentally statistical tools that often produce an "average answer," and they can also erode your own sense of competence.
- The columnist who wrote this piece tried it himself, going back and forth with ChatGPT to build a plot more intricate than he could have written alone. But he says the process cost him the joy of synapses straining to find new connections, and he ultimately decided to write it himself.
He had ChatGPT draft a movie plot
Mischa's parents are throwing her engagement party at the family's country estate. As she gives her toast, the glass slips from her hand, and inside the stem is something hidden: a micro memory card. The find unravels a corruption scandal buried for decades, one that implicates her own family and forces her to choose between justice and loyalty.
No screenwriter wrote that. New Scientist columnist David Robson gave ChatGPT a creative-writing prompt, a single line: "Her face froze, eyes widening, as she watched the glass shatter on the floor." His goal was to sketch out a short plot outline in thirty minutes.
He wasn't angling for a career in Hollywood. He was testing a study's findings.
Why it's worth reading: three rounds within the same study, a roughly 150-person experiment, a 319-person replication, and a survey of professional creatives, all produced the same inverted-U result: moderate users performed best. The first two rounds were scored by human expert judges on novelty, practicality, and commercial value; the third measured creative performance through each participant's own manager.
Three rounds of testing, one consistent answer
The study is by Hsuan-Che Brad Huang, completed during his PhD at the University of British Columbia; he now teaches at the National University of Singapore Business School. The internet has argued for a long time about whether algorithms will replace human mental labor. Huang wanted to ask something different: can AI make us more creative ourselves?
The main experiment used a classic creativity test: help a student named Mike come up with an impromptu startup idea on a $10 budget. Participants had to generate multiple ideas, then write Mike an email spelling out the best one in detail. About 150 people took part.
The three groups differed in exactly one way: how much AI use they were told to aim for. One group used ChatGPT once, one used four to six prompts, and one used at least nine. For a fifteen-minute task, nine prompts is a heavy investment. All three groups used AI — the only difference was the suggested amount. The results went to human expert judges, who scored them on novelty, practicality, and commercial value.
The result matched Huang's prediction exactly, the middle group scored highest on average. To rule out chance, he ran a second experiment with 319 people and got the same result.
Switch to people who do this for a living, and the curve holds
What works in the lab doesn't always hold up on the job. So Huang did a third thing: he surveyed professional creatives, including fashion designers, visual artists, writers, animators, tech workers, and influencers.
The questionnaire had them rate statements like "I use AI to perform most of my job functions," from 1 (strongly disagree) to 7 (strongly agree). Creative performance wasn't self-reported, it came from each person's own manager. The result lined up with the first two rounds: the group that self-rated 4 to 5 showed the best creative performance.
Why Huang thinks moderation is the key
Huang's hypothesis rests on an observation about people: our creativity is often boxed in by our own experience. Everyone carries assumptions that push their thinking toward a fixed direction, which is why hearing a different way of seeing the world tends to help.
There's evidence this is exactly why teams tend to be more creative than individuals. AI can play the same role, offering a perspective outside your own experience.
The author then lists a few of the technology's shortcomings. It's on this basis that Huang concludes moderation is the key.
It supplies angles you wouldn't think of yourself, pulling you out of your fixed assumptions. That's also why a group brainstorm usually beats one person thinking alone.
- Large language models are fundamentally statistical tools that often produce an "average answer," lacking the personal quirks of a human mind
- They tend to absorb the biases and prejudices baked into their training data
- More troublingly, they can erode your own sense of competence, leaving you feeling like it doesn't really matter
The Goldilocks zone. The name comes from the fairy tale where a girl wanders into the bears' house and samples their porridge — the first bowl too hot, the second too cold, the third just right. Using AI works the same way: you're looking for that just-right middle setting.
He tried it himself, and decided to write it alone anyway
Back to that movie plot from the opening. Robson laid out his entire process, and after a few rounds back and forth, it's clear who was steering at each step.
His own assessment: at every step, the AI widened the range of events he was considering, which almost certainly let him land on a more intricate plot than he could have built alone in the time he had.
But he has three reservations. Each one bites at a different moment: one during the writing itself, one at the instant of handing it in, and one that won't show up for years.
The first point is covered in the graphic above. The second deserves another sentence: he's recently started writing short fiction, and what draws him in is exactly those motifs and themes that surface unexpectedly, somehow tied to his own experience and his mood in the moment. That channel was closed the one time he used AI.
Even Robson isn't sure about the third point. Grace Liu and colleagues at Carnegie Mellon University in Pennsylvania ran a study, its title translates roughly to "AI Assistance Reduces Persistence and Undermines Independent Performance," showing that people quickly grow dependent on AI, with persistence dropping once it's taken away. Submitted in April 2026, it remains a preprint and hasn't yet been formally published through peer review. Robson himself suspects this isn't necessarily a permanent change, but he doesn't like the idea that he might be setting himself up for lazier thinking down the road. For him, writing should be a learning curve that keeps climbing.
This isn't a judgment on anyone who chooses to create with AI. I believe plenty of people can do it without harming their own intellect. But if genius really is, as the saying goes, more sweat than inspiration, I'd rather sweat it myself.David Robson, closing the column
Writing with AI: from "more is better" to the middle tier is best
A University of British Columbia study, verified across three rounds, all pointing to the same sweet spot, laid out here in one page with charts.
↓ Read it in one page · includes an animated chart
The debate over using AI to help write and brainstorm has mostly been about whether it will replace people. Hsuan-Che Brad Huang, who ran this study, wanted to ask something else: can AI make people more creative themselves, and how much should they use?
Until now, the only answers to that question were anecdotal — nobody had actually tested it.
AI can hand you angles you wouldn't think of yourself
✘ What nobody could pin down
How many times you should actually use it per task
Both gut-feel extremes have believers: "use it sparingly so it doesn't take over," versus "use it as much as you can, it's faster anyway." No experimental data existed to say which was right.
The study measured a "sweet spot": creativity scores drop whether you use AI too little or too much, peaking in the middle tier. The researcher borrowed the fairy-tale term Goldilocks zone (a girl wanders into the bears' house and samples their porridge — one bowl too hot, one too cold, the third just right).
More use → better output
Either "less is safer" or "more is always better," either way, a straight line.
1 use ↓ 4-6 uses ↑ 9+ uses ↓
An inverted U. Both ends dip, the middle peaks.
Which raises the next question: why does using it more make things worse?
The main experiment recruited about 150 people to brainstorm a startup idea for a student named Mike, working with $10 in seed money and a fifteen-minute limit. All three groups did the exact same task — the only difference was how many times they were told to use ChatGPT. The results went to human expert judges, scored on novelty, practicality, and commercial value.
One experiment alone wouldn't prove much. This study ran three rounds, each closer to real working conditions than the last, and the result never changed.
David Robson, who wrote the column this is based on, tried it himself, five prompts with ChatGPT to build a movie plot. He admits the result was more intricate than what he'd have written alone, but once it was done, he decided to write it himself after all, saying the process cost him the joy of his brain straining to find new connections.
let AI help
don't get carried away
it's faster anyway
Perspective never widened.
Dropped again.
more make it worse?
by your own experience
like bringing in another person to brainstorm
That outside angle never arrives.
- × It speaks in statistics, often giving "the answer most people would give"
- × The biases in its training data come along for the ride
- × You stop feeling like the result is yours, and then you stop caring
then add a detail only I would know
in your own hands
for new connections — gone
Don't ask until you're no longer in the room.
Members can copy or download the full Markdown, and let their AI subscribe and ask.
Found this useful? Leave your email — every new explainer, the moment it's out. Unsubscribe anytime.