Researchers gave twenty-five different AI models the same prompt — write a metaphor about time — and collected fifty responses from each. Over a thousand generations came back as essentially two ideas. Different companies, different continents, same sentences: one measured pair of models overlapped 81 percent of the time, down to identical phrasing. This episode […] Below is the full script of this episode as an article.
Here’s an experiment some researchers ran. Take twenty-five different AI models… different companies, different sizes, built on different continents… and give every one of them the same simple prompt. Write a metaphor about time. Fifty responses each. That’s over a thousand chances to surprise you.
Want to guess how many ideas came back?
Two.
The answers piled up into two dominant clusters. Time is a river. Time is a weaver. Over a thousand generations from twenty-five supposedly different systems, and almost everything landed on one of those two concepts. The wording shifted around the edges. The ideas didn’t move.
The Artificial Hivemind
That’s from a study out of the University of Washington, Carnegie Mellon, and the Allen Institute for AI. The researchers gave the phenomenon a name that’s been rattling around in my head ever since I came across it: the Artificial Hivemind.
And it gets stranger. Inside a single model, nearly four out of five responses to the same open-ended prompt were so similar they were barely distinguishable. But the part that got me is what happens across models. When they asked for a product description for a phone case, two models built by different companies on different continents produced the same phrases. Word for word. “Elevate your iPhone with our.” The measured similarity between those two models averaged eighty-one percent.
Eighty-one percent. Different companies. Different training runs. Same sentences.
Now, the honest caveat. The researchers don’t fully know why this happens. Maybe overlapping training data. Maybe models learning from each other’s synthetic output. Maybe similar alignment choices. They’re careful to say the causal analysis hasn’t been done yet. But the convergence itself isn’t speculation. It’s measured, and it’s not subtle.
So why am I telling you this?
Because it lands directly on the heart of chapter five of the book. That chapter is called “I Am the Scarce Input,” and the argument goes like this. The AI brings processing capacity, broad knowledge, tireless availability. Those are commodity inputs. Everyone has them. Your competitor has them for the same twenty bucks a month you do. What determines whether the output is generic or actually valuable is the one input that isn’t commodity… your judgment, your context, your sense of what matters.
When I wrote that, “commodity” was doing economic work. I meant widely available, undifferentiated, cheap. What this study shows is that it’s more literal than I knew. The models aren’t just each available to everyone. They’re converging on each other. The default output of the entire category is collapsing toward the same average. Which means the only differentiating input left in the whole stack… is you.
Scope
Let me be careful with scope here, because this is where the argument gets flattened into something dumb.
For routine work, the average answer is fine. Reformatting a document. Summarizing a meeting. Converting a file. You don’t need a distinctive voice in your file conversion, right? Take the default and move on. That’s what routine means.
The trap lives in the work that’s supposed to differentiate you. Your positioning. Your marketing. Your proposals. Your strategy. The work where sounding like everyone else is the failure condition. And that’s exactly where most people type a one-line request and take the first draft.
The book calls this the deference trap, and it’s a loop. When you believe the AI is the capable one, you stop contributing. When you stop contributing, the output goes generic… because now it’s just the model’s default. The average of everything. And when the output is generic, that quietly confirms the belief that you didn’t have much to add. So next time you contribute even less. Round and round. Each pass through the loop, less of you makes it in.
The hivemind research tells you where that loop bottoms out. It bottoms out at time is a river. At “elevate your iPhone with our.” At the exact same paragraph your competitor got this morning, because the machine that wrote yours agrees with the machine that wrote theirs eighty-one percent of the time.
The second piece of research
There’s a second piece of research worth setting next to that one. A review published this spring in Trends in Cognitive Sciences pulled together more than a hundred and thirty studies on how these tools shape human thinking. The finding that jumped out at me: individuals using AI produce more. More content, more ideas, faster. But groups relying on the same tools produce fewer distinct ideas overall. Less diversity across the whole pool.
Sit with that for a second. Everybody gets more productive, and everybody gets more alike. Both at once. The productivity is real… and so is the sameness.
And there’s a related result that keeps showing up in this literature. People who use AI to generate ideas report feeling less responsible for them. Less ownership. That one’s psychology, and it’s the deference trap wearing a lab coat. The idea came out of the machine, so it doesn’t feel like mine. And if it doesn’t feel like mine, I don’t push on it the way I’d push on my own thinking. I just… pass it along.
I get it. I’ve felt that pull myself.
In my actual work
Here’s where this shows up in my actual work. I run a content system at Auspicious, my fractional marketing practice, where AI handles a lot of production. Drafts, derivatives, the mechanical layer. And the single most important document in that whole system isn’t a clever prompt. It’s a style guide. Pages of rules about what the default machine voice does, so we can catch it and strip it out.
That document is basically a field guide to the hivemind. The words every model reaches for. The “it’s not this, it’s that” sentence pattern that shows up in machine prose everywhere. The punctuation habits that cluster three to a paragraph. I didn’t need a study to tell me the models converge. I keep a checklist because they converge. Every rule in that guide exists because the default output, from any model you pick, kept making the same choices.
But the style guide is only the defensive half. The offensive half is what chapter five is actually about. What goes in. The drafts that work aren’t the ones where the prompt was clever. They’re the ones where the input was full of things no model could know. The client’s real numbers. The thing the owner actually said in the room. The reason the obvious angle is wrong for this particular audience. That material can’t come out of the machine, because it was never in the machine. It’s scarce in the literal economic sense. One source. You.
What to do Monday morning
So what does this mean for Monday morning?
Try one test. Pull up the last thing you shipped that AI helped you write, and ask a single question. What’s in here that only we could have said? Not phrased better… said at all. A number from your own operation. A pattern from your own customers. A position your competitor would be scared to take. If the honest answer is nothing, you shipped the average. And the average is now a measured quantity. It’s two metaphors deep.
Notice how it felt to make, too. If it was fast and easy and the first draft looked great, that’s not automatically bad. But it deserves suspicion. Ease can mean mastery. It can also mean you were never really in the room. Here’s the uncomfortable part… from the inside, those two feel identical. The difference only shows up in the output. One version has your fingerprints on it. The other could have been anyone’s, and increasingly, it literally is everyone’s.
The strange comfort in all this research is that the machines converging makes the human contribution more valuable, not less. Everything the model brings, everyone gets. What you bring, nobody else has. That was the book’s argument before this data existed, and the data keeps arriving on its side… with the caveats I gave you, and I’ll keep watching how the causal story shakes out.
The commodity is spoken for. Bring the scarce part.
—
This show extends the book AI Empowered. Read the first chapter free at https://www.aiempoweredbook.com, or get the book on Amazon.
Aaron Douglas runs Auspicious, a fractional marketing practice.
Blog: https://auspicious.llc/insights/
X: https://x.com/aaron_douglas
LinkedIn: https://www.linkedin.com/in/aaronddouglas/
