When AI Ideas Start Sounding Like Everyone Else's
- networkingnerd
- 667 views
Models from different vendors keep converging on the same answers. Here's what that means for teams using AI to generate strategy and product ideas.

Picture a product strategy session where someone asks an AI model for an unusual positioning angle. The answer lands well. The room likes it. Then, a few months later, a competitor opens a keynote with almost the same framing. Leaked deck? Probably not. The more likely explanation is less flattering: you and your competitor asked similar questions of similar models and got similar answers back.

This is the frequency illusion at work, and it's not new. Linguist Arnold Zwicky described it on Language Log in 2005, noting that once you spot something, it seems to be everywhere. Usually, your attention changed and the world didn't. The difference now is that the repetition is real, and it's being generated on demand.
Same Well, Different Buckets
A Stanford team led by Chenglei Si, Diyi Yang and Tatsunori Hashimoto tested how many distinct ideas a model actually holds. They recruited more than 100 natural language processing researchers, who wrote their own ideas and blind-reviewed ideas from humans and an AI agent. Reviewers rated the AI ideas as significantly more novel than those from human experts.
The catch sits underneath the winners. The agent generated a large pool of ideas and ranked the best for reviewers. Out of 4,000 seed ideas on one topic, only about 200 were unique by the team's similarity measure. Reviewers saw the top picks. They never saw the near-duplicates stacked beneath them.
Does this repetition stay inside one person's chat window? A team from the University of Washington, Carnegie Mellon and the Allen Institute for AI looked at that question. They asked models each of 100 open-ended questions 50 times with randomness turned up. For 79% of questions, the answers still averaged above 0.8 on their similarity scale, where 1.0 means identical meaning and two random answers from the full pool land around 0.1 or 0.2. They then reworded the questions the way different people naturally would. The answers barely moved.
So you get one idea that feels new to you, and it might genuinely be good. What you can't see is everyone else who asked something similar and got something similar back. It's like finding a hidden-gem restaurant in a travel guide. The food is good. Every other table is full of people who found it the same way.
The mechanism is fairly mundane. These models train on heavily overlapping text. They're then tuned toward responses people rate highly, which tends to favour the likely answer over the odd one. Ask for something unconventional and you get the most probable version of unconventional. The model has no way to notice that contradiction.
Shoe Company Blues
In 2023, researchers from Harvard, MIT, Wharton and other institutions worked with Boston Consulting Group on a preregistered experiment involving 758 consultants, roughly 7% of BCG's individual contributor workforce. Some had GPT-4 and some didn't. In the creative task, 385 consultants imagined working for a footwear company and proposed new products. Those with AI produced better ideas, but the ideas also looked much more like each other. BCG's own write-up puts the drop in that group's diversity of thought at 41%.
The researchers spotted the competitive problem quickly. If many firms use the same tool, the people who still generate ideas unaided may be the ones who stand out. Anil Doshi and Oliver Hauser found a similar trade-off among short story writers and called it a social dilemma. Each individual gains, while the collective range of ideas shrinks.
Shopping for a Second Opinion
The obvious response is to switch tools. The Washington, Carnegie Mellon and AI2 team tested more than 70 models and examined 25 closely, including GPT-4o, Claude 3.5 Sonnet, Gemini 2.0, DeepSeek-V3 and OpenAI's o-series reasoning models. They named the result the Artificial Hivemind. Individual models repeat themselves, and models from different companies converge on one another. Answers across models scored between 0.71 and 0.82 on the same similarity scale.
In one test, they asked 25 models for a metaphor about time, 50 times each, and plotted the results. The responses formed two clusters: a large one around time as a river, and a smaller one around time as a weaver.
More randomness didn't fix it. Even with a sampling method designed for variety, about 61% of answer pairs from a single model stayed above 0.8. A separate study of 22 models found the same pattern on standard creativity tests, with AI answers far more similar to each other than human answers were to each other.
It's a bit like a second opinion from a doctor who trained at the same school, read the same textbook and worked under the same senior physician. You hear the same diagnosis in 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 the suspects. Either way, changing vendors doesn't guarantee a different perspective.
Iron Cage, New Lock
In 1983, Paul DiMaggio and Walter Powell published "The Iron Cage Revisited" in the American Sociological Review, asking why organisations in the same field end up looking alike. Their answer included a concept they called mimetic isomorphism: when you're unsure what to do, you copy whoever seems to be succeeding. They also pointed to consulting firms as a source of that pressure. Their study focused on organisational structures and practices, but strategy has felt the same pull for decades.
What's different now is the disguise. Mimetic copying used to be visible. You knew you were imitating the market leader, or that a framework came from a consulting deck, and you could discount it for that reason.
AI-assisted convergence arrives looking like your own idea. You wrote the prompt, picked the answer and argued for it in the meeting. It's a little like everyone painting their kitchen the same grey after seeing it on a TV makeover, except now they believe they chose it themselves. That's harder to spot, and harder to challenge.
Seeing Double
This is where the frequency illusion stops being a party trick and starts affecting budgets. Say your team lands on that unconventional pitch. A month later, a competitor announces something similar, and then an analyst note uses the same framing. The natural conclusion is that the market is moving your way and you got there first. Confidence rises, and spending follows.
The classic version is a perception error: you notice something more, so you assume it's more common. This version is worse because the repetition is real. The idea really is turning up everywhere. And DiMaggio and Powell's point about uncertainty applies directly. That's exactly when organisations copy one another, and a crowd of competitors arriving at your idea feels like the certainty you were missing. You end up mistaking an echo for a crowd.
Getting Off the Median
You might say ideas were always cheap and execution is what wins. I'd agree, but follow that through. If ideas converge, execution becomes the main thing left to compete on. If your team still treats the idea as the edge, you haven't planned for this. If you want AI to sharpen your position rather than sand it down, change how you use it.
- Generate before you prompt. Have people write ideas cold before anyone opens a chat window. The BCG researchers suggested unassisted ideas are the ones most likely to stand out. You can't stay unassisted indefinitely, but writing first buys some independence before the model pulls everyone toward the middle.
- Prompt for structure. Rewording a question barely changes the output. Research from Wharton's Lennart Meincke, Ethan Mollick and Christian Terwiesch found that chain-of-thought prompting, where the model reasons step by step through its options before answering, came close to the idea diversity of human groups. I'd also discard the first few ideas a model gives you, since the first answer is the one everyone else tends to get.
- Treat convergence as a question. When a competitor lands on your idea, ask where both of you got it before calling it validation. If the honest answer is the same tools and similar prompts, you're holding a commodity.
Tom's Take
The idea of a universal consciousness is a fun thing to chew on, but the research points somewhere more mundane. We've built a shared source. It's statistical, it has a training cutoff, and millions of people ask it for similar things at the same time. If your goal is to be different, the model will hand you the most popular version of different it knows. Do the unconventional thinking yourself first, then use the machine to refine it. Novel has become cheap. Genuinely different still costs extra.