Most companies now treat AI search and chat tools as a shortcut to good ideas. A new study suggests that shortcut has a cost nobody notices until it is too late: it makes everyone's ideas look the same.
Researchers built a modified version of Google Search that, instead of surfacing the most popular and relevant results, deliberately surfaced unusual and unrelated information. They tested it two ways. In a lab test with 104 people, ideas produced with the modified tool were rated 14 percent more creative than ideas produced with ordinary Google Search. In a bigger, real-world test with 245 people tackling household food waste, the effect got more interesting.
With ordinary search, domain experts were no better at generating creative solutions than complete novices. Give both groups the standard tool, and their ideas landed in the same one or two clusters of concepts. But hand the modified, exploration-based tool to experts, and something changed: their ideas spread across five distinct concept clusters, while everyone else stayed stuck at one or two. Expertise only paid off once the tool stopped funneling people toward the obvious answer.
This is not an isolated finding. A separate study out of Wharton, published in Nature Human Behaviour, tested ChatGPT directly on a brainstorming task and found that 94 percent of the resulting ideas shared overlapping concepts, with nine participants independently landing on the identical product name for a toy. Only 6 percent of AI-assisted ideas were judged unique, against 100 percent for people working without the tool. Two separate research teams, two different tools, the same conclusion: efficient AI systems make groups converge.
The pattern shows up outside brainstorming too. Researchers studying hiring algorithms have documented what they call algorithmic monoculture: when many employers use the same few vendors, the same candidates get rejected everywhere, at a rate higher than chance would predict. Popular tools built for efficiency do not just save time. They quietly compress the range of outcomes for everyone using them.
For a business, the takeaway is not to distrust expertise or throw out AI tools. It is to stop assuming that any AI tool automatically produces variety. A tool tuned for quick, popular, "good enough" answers will always pull a room full of smart people toward the same three ideas, no matter how experienced they are.
The fix is not complicated. When the task is generating options, not confirming a known best practice, ask your AI tools for unfamiliar comparisons, unrelated industries, or arguments against the obvious answer. Save the standard, efficient search for jobs where you already know roughly what the right answer looks like. Save the deliberately weird prompts for the moments when you actually need a new one.