Two designers on the same team were asked to sketch a logo for a small bakery. They sat at opposite ends of the room, typed their own prompts into an AI image tool, and came back twenty minutes later with almost the same mark: a soft wheat stalk, a warm cream-to-amber gradient, the same rounded lowercase lettering. Neither had copied the other. Both had asked politely for something “warm and artisanal,” and the tool answered them the same way.
That small coincidence repeats at scale. The speed is real. So is the sameness that arrives with it.
The Output Paradox
Both designers made something good.
Photo by That’s the pattern researchers keep running into. In one study of AI-assisted storytelling, the stories were rated more original than unaided ones. At the same time, they were measurably more similar to each other [Pure].
In plain terms, each writer got better while the group got narrower.
The two effects run at once. Individual quality can rise. Collective variety can fall. A meta-analysis of 17 experiments found that generative AI’s edge over humans in raw idea generation was small and not statistically significant [Pure], which suggests the boost people feel is often about speed and finish rather than genuine novelty. A separate 2025 assessment found AI models outscoring humans on standard divergent and convergent thinking tests, the kind of tests that measure how many different ideas someone can generate versus how well they can narrow down to one correct answer [Nature]. The picture is mixed, not settled.
The draft in front of you may be sharper than what you’d write alone, and also closer to what your competitor is publishing this week.What The Model Is Actually Doing
Photo by The mechanism isn’t mysterious. A large language model, the AI system behind most text tools that predicts the next word based on patterns it learned from huge amounts of text, works by guessing the most likely next piece of text given everything before it. Image generators do something similar with pixels. They move a noisy canvas toward the most probable version of what you asked for.
Probable is the key word here. These systems are tuned to land on the expected answer, because the expected answer is usually the useful one. Ask for a bakery logo and you get the statistical center of every bakery logo the model has seen.
When the whole industry queries a handful of the same models, everyone pulls from the same center of gravity. Researchers analyzing text from GPT-3.5, GPT-4, and Llama 70B found the outputs clustering tightly by stylistic features [Nature]. Different companies, different prompts, same fingerprint.
The tool isn’t failing when it gives you something familiar. Familiar is what it was built to deliver.
The Same Pattern Across Industries
A cross-industry view helps here, because the effect shows up first where production cycles move fastest.
Photo by -
Marketing ships in days, so overlapping slogans and near-identical campaign visuals surface within a single season.
-
Product design teams describe AI suggestions defaulting to a small set of popular layouts and palettes, regardless of the brand’s own identity.
-
Software shows it structurally: AI-assisted code tends toward the same scaffolding patterns across companies solving unrelated problems.
Ecologists have a name for this shape. A monoculture, a field planted with a single crop, is efficient and high-yield but fragile, because one blight can take the whole harvest. Creative work is starting to look similar. When every team draws from the same generative source, the pool loses the odd variants that would have carried it through a shift in taste.
The risk isn’t bad work. It’s a market where good work stops being distinguishable.
It Can Be Measured
This stopped being a hunch.
Photo by Diversity in a batch of work can be scored by comparing how semantically close the pieces are to one another, and those scores can be tracked over time.
A field experiment with 99 knowledge workers found that generative AI significantly homogenized both the style and the content of their work [AISeL]. An MIT analysis frames the same tension directly:
A team can audit itself: line up five recent AI-assisted pieces and read them side by side, looking for the shared skeleton rather than judging each one alone.AI can enhance individual creativity while reducing the collective diversity of novel content. (MIT Sloan Management Review)
Working With The Tradeoff
None of this argues for unplugging. The practical move is to treat generated output as raw stock rather than finished goods.
- Generate, then break something on purpose. Change the structure, the tone, or the central metaphor before the work leaves your hands.
- Rotate inputs across the team, so two people aren’t feeding the same reference set into the same model.
- Run the same prompt three times and compare. If the three results look like siblings, your published version probably has cousins elsewhere.
The deviation has to be deliberate. Light editing tends to preserve the model’s shape, and the shape is the part everyone else also has.
The human contribution has moved. It sits less in producing the draft and more in choosing where to leave the well-trodden path.
The two designers with the matching wheat stalks didn’t do anything wrong. They asked a reasonable question and got the most reasonable answer available, twice. What they lost was the twenty minutes of stumbling that used to produce a crooked, specific idea nobody else would have reached. That stumbling was never efficient, and it was where the strange logos came from. Next time the tool hands you something that looks right immediately, try asking what it would look like if you got one thing deliberately wrong. Keep the speed. Spend some of it on the detour.
🔖
- Pure, The Roles of Idea Generation and Elaboration in Human-AI Collaborative Creativity
- Pure, Generative AI Has Not Yet Surpassed Humans in Creative Idea Generation
- AISeL, Towards Generative Monoculture, field experiment with 99 knowledge workers
- Nature Humanities and Social Sciences Communications, stylistic clustering of model-generated text
- MIT Sloan Management Review, The Hidden Cost of AI-Assisted Creativity
- Nature Scientific Reports, divergent and convergent thinking assessments, 2025
