The retrieval layer is now a strategy decision: your search tools are flattening your experts

Researchers writing in MIT Sloan Management Review report that the standard algorithms behind search, discovery, recommendation and large language models operate on what they call exploitation logic: they prioritise popular, relevant results and so reinforce what the user already knows ^[2]. Their experiments found that when several people use the same tools to attack the same problem, they are independently channelled toward the same information and independently produce the same ideas ^[1].

The authors call the result ideation bubbles, and their argument about why these persist is the part worth taking seriously: the bubble is invisible from inside, because people working separately assume their outputs are diverse when the shared retrieval layer has already steered everyone into one solution space ^[3]. Personalisation makes it tighter. The tools draw on each user's existing knowledge frameworks, including search and chat histories, when deciding what to show ^[4].

The test was a modification the researchers called XYZ, built with natural language processing on top of Google Search, which surfaced results from semantically distinct clusters of ideas instead of the most popular or relevant matches ^[5]. They ran a controlled laboratory experiment with 104 participants generating ideas for reducing resource overconsumption ^[6], and a global field experiment with 245 participants ranging from sustainability novices to seasoned experts, working on household food waste ^[7]. In both, output produced with standard Google Search was compared with output produced with XYZ, and ideas were scored by independent expert judges blind to the condition ^[8]. That is 349 participants in total across the two studies ^[13].

In the lab, ideas developed with XYZ were rated 14% more creative than those developed with standard Google Search ^[9]. The field result is the one that should concern anyone who pays for specialist headcount: using standard Google Search, domain experts showed no statistically significant advantage over novices at generating creative solutions, while under the exploration-based algorithm experts significantly outperformed novices and ideas were rated 11% more creative on average ^[10]. The researchers' reading is that the tools can silently narrow an organisation's creative potential by suppressing the value of expertise, and that the fault sits in the hidden architecture rather than with the experts ^[11].

Read plainly, that is a claim about defaults, not about talent. If the expertise premium only appears under one retrieval regime, then the choice of retrieval regime is a management decision that has been sitting with procurement and IT. The researchers describe the fix as surprisingly simple: algorithms built to surface diverse, uncommon information let experts get out of the bubble and work in new solution spaces ^[12].

Hold the enthusiasm at the size of the evidence. Two studies, both on sustainability prompts, both judged on rated creativity rather than on anything that shipped, and the reported gains are 11% and 14%, not a step change ^[9][10]. The source gives no cost of building or running an exploration layer, so "simple" is the authors' word and not a budget line.

What to watch: whether enterprise search and internal retrieval-augmented systems expose a diversity setting that a leader can actually turn, whether the expertise effect replicates on commercial problems where a wrong answer is expensive, and whether any vendor is willing to be measured on idea dispersion rather than click relevance. In the meantime, the cheap diagnostic is available to any executive: if three teams brief you with the same three options, look at what they all searched with.