Give a domain expert the right kind of AI search tool and their ideas spread across five distinct areas of thinking. Give the same expert ordinary search and they produce one, sometimes two.
Everyone else in the study produced one or two whatever tool they used.
That finding comes from research published in MIT Sloan Management Review on 20 August by Moran Lazar, Hila Lifshitz, Charles Ayoubi and Hen Emuna. It matters because it answers a question this audience keeps asking in different words: if the tool does the thinking, what is my experience worth now.
The answer this research gives is uncomfortable in one direction and reassuring in the other. The tool does very little on its own. The expert is the variable that makes it work.
What they did
Two studies, run separately.
The lab study put 104 participants through an idea-generation task. Some worked with what the researchers call exploration-based algorithms, which deliberately surface less popular and less obvious material. Others used standard Google Search. Ideas developed with the exploration-based tools were rated 14 per cent more creative.
The field study is the more interesting one. 245 participants, this time separating domain experts from non-experts. Experts using exploration-based tools produced ideas rated 11 per cent more creative on average.
With standard search, the expert advantage did not shrink. It disappeared. There was no significant difference between what an expert produced and what anybody else produced.
The number that carries the argument
Range is where the effect shows up most clearly.
Experts using exploration-based tools generated ideas spanning five distinct semantic clusters. Every other group in the study, experts on ordinary search included, managed one or two.
Semantic clusters are a way of measuring how far apart ideas sit from each other. Five clusters means five genuinely different directions rather than five variations on one thought. Anyone who has run a workshop knows the difference immediately: the session that produced a page of options that are all the same option, against the one that produced four you had not considered.
That is the finding worth carrying, because range is the thing managers actually notice going missing. Nobody says the team got less creative. They say the ideas all sound the same this year.
The ideation bubble
The researchers name the failure mode, and the name is good enough to steal.
An ideation bubble is what happens when the tool keeps steering you towards popular, familiar, well-trodden material. Standard search is built to do this. It is optimised to give you the answer most people wanted, which is exactly right when you have a question with an answer, and exactly wrong when you are trying to think of something nobody has thought of yet.
Their summary: “Standard search algorithms can trap teams in ideation bubbles. Exploration-based tools can help spark breakthrough innovation.”
The trap is quiet. A bubble does not feel like a constraint from the inside. It feels like getting good, fast, confident answers, which is precisely why it is worth knowing about.
The caveat that stops this being useful advice
Here is where I have to be careful, because the obvious next question is which tool to buy, and the answer is that there is not one.
An exploration-based algorithm, in this research, is an instrument built for the study. It is not a product you can subscribe to. Nothing on the market is the commercial version of it, and anyone telling you their tool is that thing is going beyond what this research supports.
What transfers is not the software. It is the instruction.
The tools most people use are tuned for the popular answer, and you can push against that deliberately. Ask what the unfashionable position is. Ask what somebody in a completely different industry would try. Ask for the answer that most people would disagree with.
None of that requires new software, and all of it moves you towards the behaviour the exploration-based tools were built to produce.
Two more honest limits. Both studies are small, at 104 and 245 participants. And they measure rated creativity of ideas, which is a reasonable proxy but not the same as an idea that worked.
Why this is the reassuring half
Yesterday I wrote about a widening gap between the organisations getting a lot from AI and the ones getting little, and about the discomfort of finding that early-career staff are further ahead than executives.
This is the other half of that.
The thing that closes the gap is not the licence, the budget or the vendor. In this research it is domain expertise, and expertise is the one input a mid-career professional already has in quantity. The tool did nothing much for people without it. It only became worth something in the hands of someone who knew the field well enough to recognise which of the unfamiliar options was worth pursuing.
That is a real reversal of the fear. The common worry is that AI makes accumulated judgement worth less. This research points the other way: the judgement is what turns the tool from a faster way to find the obvious answer into a way of finding a different one.
What to do with it tomorrow
Next time you use AI to think rather than to draft, ask it for the unfamiliar answer rather than the best one.
The best answer is what it is built to give you, and most of the time that is what you want. But when you are trying to get somewhere new, the best answer and the popular answer are the same thing, and that is the bubble.
Your expertise is what lets you look at five strange options and know which one is worth an afternoon. Without that, five options is just noise. With it, it is the most useful thing the tool can do for you.
Source: Algorithms Trap Us in the Familiar. Can They Also Spark Breakthroughs?, Moran Lazar, Hila Lifshitz, Charles Ayoubi and Hen Emuna, MIT Sloan Management Review, 20 August 2026.