Imagine your product strategy team gets stuck on a positioning problem. Somebody opens an AI chat window and asks for an unconventional angle. The model comes back with a concept nobody in the room has thought about in years. Maybe it’s the Lindy effect or Chesterton’s fence or something borrowed from evolutionary biology. The room loves it. It feels fresh and a little clever. Three months later a competitor walks onto a keynote stage with the same framing. How did that happen? Did someone leak the deck?

You’ve probably felt a smaller version of this. You learn a new word and it shows up in three articles that same week. The linguist Arnold Zwicky named this on Language Log in 2005. He called it the frequency illusion, though you may know it by its folk name, the Baader-Meinhof phenomenon. Once you notice something, you think it’s everywhere. Most of the time it isn’t. Your attention changed. The world didn’t. When the coincidences really start piling up, some people reach for something grander. Great minds think alike. Ideas are in the air. Maybe there’s a universal consciousness out there that we’re all dipping into. Or maybe it’s as bland as a plate of shrimp.

When teams ask AI for differentiation, they get the statistical center of everyone else’s thinking.

When a competitor lands in the same place, the frequency illusion kicks in. The convergence starts to look like market validation instead of a shared source.

Same Well, Different Buckets

A team at Stanford ran one of the best tests of how many ideas a model really has. Chenglei Si, Diyi Yang and Tatsunori Hashimoto recruited more than 100 researchers from natural language processing, the field that builds these models. The researchers wrote their own ideas and blind reviewed ideas from both humans and an AI agent. The reviewers judged the AI ideas as significantly more novel than the ones from human experts.

The agent worked by generating a huge pile of ideas and ranking the best ones for the reviewers. When the researchers looked at the pile, duplicates were everywhere. Out of 4,000 seed ideas generated on a topic, only about 200 were unique by their similarity measure. The reviewers saw the winners. They never saw the near repeats stacked up underneath.

That’s one model repeating itself to one person. Does it repeat itself to everyone else too? A team from the University of Washington, Carnegie Mellon and the Allen Institute for AI ran that test too. They asked models each of 100 open-ended questions 50 times with the randomness turned up. For 79% of those questions the answers still averaged above 0.8 on their similarity scale. For a sense of scale, 1.0 means the same meaning, and two random answers pulled from the whole pool land around 0.1 or 0.2. Then they reworded the questions the way different people naturally would. The answers barely changed.

Think about what that looks like from inside one chat window. You get one idea and it’s new to you. It might even be genuinely good. What you don’t see is everyone else who asked something similar and got something similar back. It’s like finding the hidden gem restaurant in a travel guide. The food is great. Every other table is full of people that found it the same way you did.

Why does this happen? These models are trained on heavily overlapping piles of text. Then they’re tuned to give the kinds of answers people rate highly, and that tends to favor the likely response over the odd one. Ask for something unconventional and you get the most likely version of unconventional. Which is a contradiction the model has no way to notice. Even though it feels like your chats are leaking between users it’s more benign and systemic.

Shoe Company Blues

In 2023 researchers from Harvard, MIT, Wharton and others worked with Boston Consulting Group on a preregistered experiment with 758 consultants. That’s about 7% of BCG’s individual contributor workforce. Some of them got GPT-4 and some didn’t. In the creative half of the experiment, 385 consultants imagined working for a footwear company and came up with new products. The consultants with AI produced better ideas. They also produced ideas that looked a lot more like each other. BCG’s own write-up puts the drop in that group’s diversity of thought at 41%.

The researchers saw the competitive problem right away. If lots of companies use the same AI, the ones that generate ideas without it might be the ones that stand out. Anil Doshi and Oliver Hauser found the same trade with short story writers and gave it a good name. They called it a social dilemma. Everyone is better off individually, and the collective range of ideas shrinks.

Shopping for a Second Opinion

The obvious fix is to use a different model. If everyone’s on the same tool, switch tools, right? That team from Washington, Carnegie Mellon and AI2 tested more than 70 models and looked closely at 25 of them. The list includes GPT-4o, Claude 3.5 Sonnet, Gemini 2.0, DeepSeek-V3 and OpenAI’s o-series reasoning models. They call the result the Artificial Hivemind. Individual models repeat themselves, and models from different companies converge on each other. Answers from different models scored between 0.71 and 0.82 on that same similarity scale.

They asked 25 models for a metaphor about time, 50 times each, and plotted the answers. The responses formed two clumps. A big one gathered around time as a river and a smaller one around time as a weaver.

Turning up the randomness didn’t help. Even with a sampling method designed for variety, about 61% of the answer pairs from a single model stayed above 0.8. A separate study of 22 models found the same thing in standard creativity tests. AI answers were much more similar to each other than human answers were to each other.

It’s like getting a second opinion from a doctor that went to the same school, read the same textbook and trained under the same senior physician. You’ll hear the same diagnosis with slightly different words. The Hivemind authors don’t claim to know exactly why the models converge. Shared data pipelines and synthetic training data are among their suspects. Either way, swapping vendors doesn’t buy you a different brain.

Iron Cage, New Lock

In 1983 Paul DiMaggio and Walter Powell published “The Iron Cage Revisited” in the American Sociological Review. They wanted to know why organizations in the same field end up looking so alike. Rational people trying to change their organizations end up making them more similar. Some of that pressure comes from regulators and from professions that train everyone the same way. The part that matters here is what they called mimetic isomorphism. When you aren’t sure what to do, you copy whoever seems to be winning. They even called out consulting firms as a source of that pressure. Their work was about organizational structures and practices, but I’d argue strategy feels the same pull and has for decades.

What’s different now? Mimetic isomorphism used to have a face. You knew you were copying the market leader. You knew the framework came out of a consulting deck. You could discount it because you could see where it came from.

AI-assisted convergence arrives as your own idea. You typed the prompt. You picked the answer. You argued for it in the meeting. Think about everyone painting their kitchen the same shade of gray because a designer did it on TV. Now think about everyone painting their kitchen that gray while believing they picked it themselves. The second one is a lot harder to spot. It’s harder to push back on too.

Seeing Double

This is where the frequency illusion stops being something for trivia night and starts costing money. Let’s say our team lands on that unconventional pitch. We’re primed now. A month later a competitor announces something similar. Then an analyst note uses the same framing. What do we think? We don’t think we all asked the same oracle. We think the market is moving our way and we got there first. Confidence goes up. Budget follows.

The classic frequency illusion is a perception error. You notice something more, so you assume it happens more. This version is worse because the frequency is grounded in reality. The idea really is showing up everywhere. And remember what DiMaggio and Powell said about uncertainty. That’s exactly when organizations start copying each other. A crowd of competitors arriving at your idea feels like the certainty you were missing. Which means we mistake an echo for a crowd.

Getting Off the Median

You’re probably thinking, “Well ideas were always cheap. Execution is where companies win.” I agree. But follow that all the way through. If everyone’s ideas converge, execution becomes the only thing left to compete on. If your team still thinks the idea is the competitive edge, you haven’t planned for that. If you want AI to help you find an edge instead of sanding yours off, change how you use it.

• Generate before you prompt. Have people write ideas cold before anyone opens a chat window. The BCG researchers suggested that unassisted ideas are the ones likely to stand out. You can’t stay unassisted forever, but writing first buys you some independence before the model pulls you toward the middle.
• Prompt for structure. Rewording a question barely changes what comes back. What helped in research from Wharton’s Lennart Meincke, Ethan Mollick and Christian Terwiesch was chain-of-thought prompting, where you make the model reason step by step through its options before it answers. That got close to the idea diversity of human groups. I’d also throw away the first few ideas it gives you. The first answer is the one everyone else got.
• Treat convergence as a question. When a competitor lands on your idea, ask where both of you got it before you call it validation. If the honest answer is the same tools and the same prompt, what you’re holding is a commodity.

Tom’s Take

I started thinking about this because of the idea of a universal consciousness. The “plate of shrimp” to quote Repo Man. Then I looked at the research and realized we built one. It’s statistical. It has a training cutoff. And millions of people are asking it for the same thing at the same time. But if the goal is to be different, the model will hand you the most popular kind of different it knows. Do the weird thinking yourself first. Use the machine to sharpen it. Novel is cheap now. Different still costs extra.