Algorithms Trap Us in the Familiar. Can They Also Spark Breakthroughs?

Gary Waters/Ikon Images Algorithmic tools promise to democratize access to knowledge and thus spark creativity and innovation, but research we conducted revealed a hidden risk: Those tools may be silently narrowing organizations’ creative potential by suppressing the value of expertise. The fault lies not with the experts but with the hidden architecture of the tools they use. The invisible design choices embedded in algorithmic tools fundamentally shape creative output. Standard algorithms behind search, discovery, recommendations, or large language models (LLMs) are designed around exploitation logic: They prioritize popular, relevant results, which reinforces what users already know instead of challenging them to explore. When such tools are designed for efficiency rather than exploration, they channel users toward conventional information, creating what we call ideation bubbles: clusters of similar ideas that represent a dangerous homogeneity of thought. But our findings also revealed a solution: When we modified exploitation-based algorithms to surface diverse, uncommon information, experts who used them generated solutions that were significantly more creative, and they were able to break free from the convergent thinking patterns that can trap entire organizations. Most digital tools we use today are designed to prioritize efficient access to popular answers. They also draw on a user’s existing knowledge frameworks (such as their search or chat histories) when presenting information and rarely challenge them to explore new territory. While this approach excels at delivering quick, useful results, our research found that this kind of bias is detrimental to creativity and innovation. At the individual level, creators gravitate toward familiar solutions. The problem compounds at the organizational level. When multiple people use the same exploitation-based tools to brainstorm solutions to the same challenge, they are independently channeled toward the same information, and they independently generate similar ideas. As is the case with news bubbles, ideation bubbles are imperceptible to those inside them: Individuals believe that they are generating diverse ideas because they are working independently, but the shared algorithmic infrastructure steers everyone toward the same solution space. This convergence risk is especially dangerous for strategic challenges requiring breakthrough thinking. Surfacing a Wider Spectrum of Ideas To test our thinking about exploration versus exploitation, we designed an algorithmic modification we called XYZ that uses natural language processing. Built on top of Google Search, it surfaced results from semantically distinct clusters of ideas rather than the most popular or relevant matches — prioritizing exploration over exploitation. We then conducted two complementary studies: a controlled laboratory experiment with 104 participants, who were asked to generate creative ideas for reducing resource overconsumption; and a global field experiment, in which 245 participants, ranging from sustainability novices to seasoned sustainability experts, participated in an ideation challenge to reduce food waste in households. In both cases, we compared their creative output when using either Google Search or XYZ. The ideas were evaluated by independent expert judges blind to the experimental conditions. In the laboratory study, ideas developed with XYZ were rated 14% more creative than those developed with standard Google Search, evidence that exploration can lift creativity even without deep domain knowledge. The more striking finding came in the field study: When using standard Google Search, domain experts showed no statistically significant advantage over novices at generating creative solutions. When using exploration-based algorithms, experts significantly outperformed novices, and ideas were rated 11% more creative on average. The algorithmic design, operating imperceptibly beneath the surface, shaped whether experts could make their strongest contributions. What explains the gap between experts and novices? The key lies in what we call recombinant innovation : synthesizing diverse information elements into novel combinations. Exploitation-based algorithms surface familiar information that echoes existing mental models, but exploration-based algorithms expose users to insights from fields that may be unfamiliar to them, such as, in our study, behavioral economics or supply chain optimization. Experts have the knowledge foundation to harness this diversity effectively. A novice encountering the same diverse information lacks that scaffolding. Experts can recognize which unfamiliar ideas are relevant, which are dead ends, and how to integrate them into a workable solution. For instance, one expert in our study combined insights about community food-sharing platforms with smart-home technology and behavior change techniques to propose a neighborhood-based “food rescue network” with automated inventory matching. That recombination required a deep understanding of food waste patterns. Helping Domain Experts Break Out of Ideation Bubbles While the effects of exploration-based algorithms are profound for individuals, the organizational implications may be even more significant. Using natural language processing to semantically cluster all ideas from our field study, we found two distinct effects and a critical interaction between them. First, exploration-based algorithms increased idea diversity for all participants. Novices using Google Search (exploitation) produced ideas that fell into just one semantic cluster; novices using XYZ (exploration) produced ideas spanning two clusters. Second, expertise alone had a similarly modest effect: Experts using Google Search generated ideas across two clusters compared with novices’ one. But the interaction between expertise and exploration-based algorithms was dramatic: Experts who used XYZ generated ideas across five distinct clusters, compared with one or two from every other group. These experts did not just contribute more ideas within existing solution spaces; they generated entirely new ones, breaking dominant ideation bubbles and creating unconventional clusters of thinking. These findings have immediate implications for how organizations structure their innovation processes. First, they should treat algorithm type as a design input, not a default. Most tools that organizations use are optimized for efficiency, not exploration, which can suppress the value of what experts can contribute. When tackling strategic challenges that require breakthrough ideas, companies should consider approaches that surface diverse and uncommon information instead of the most popular or obviously relevant results. Try the Tool Try our ideation bubble tool to discover how ideas from your ideation process cluster into distinct bubbles and how you can burst them to generate unique ideas. Second, organizations should match the algorithm to the task. Exploitation-based algorithms remain valuable for accuracy and efficiency — identifying best practices or answering well-defined questions. Exploration-based approaches are best reserved for early-stage ideation, when divergent thinking is most valuable. This applies directly to how organizations use LLMs: Prompting for exploration rather than exploitation — for instance, asking for approaches that draw from unrelated industries or explicitly challenging dominant assumptions — can dramatically affect creative output. When reviewing AI-generated material, experts should particularly attend to unexpected or unfamiliar elements and avoid gravitating toward expected, recognizable patterns. Novel results are where the raw material for recombinant innovation often lies. Third, organizations should audit their idea portfolios for ideation bubbles. If a team’s ideas cluster around a narrow set of solutions, the problem may be the tools. Semantic clustering can help