Does generative AI enhance creativity while homogenizing ideas?

Creativity generally includes not only novelty but also usefulness for the purpose. When evaluating an individual's proposal, novelty and usefulness are important, but in product development or strategy formulation, it is also crucial how many different hypotheses the multiple proposals explore. Even if there are ten proposals with high average scores, if they all depend on the same customer profile and solution, there is only one option for responding to environmental changes.

Therefore, we divide the creative effect of generative AI into two questions. First, does AI improve each person's proposal? Second, when collecting proposals from people using AI, is the breadth of meaning, structure, and expression maintained? Research has reported results where these two move in opposite directions. [1][2]

• Individual Quality: Novelty, usefulness, feasibility, clarity
• Collective Diversity: Semantic distance between proposals, overlapping clusters, number of independent premises
• Exploration Upper Limit: Whether the best proposal beyond the existing range was generated, rather than the average proposal

Holzner et al. integrated 28 studies and 8,214 participants, reporting no significant difference between generative AI alone and human creative performance, while human+AI slightly exceeded humans alone. The integrated effect size was Hedges' g of 0.273. [1] This is not a result that generative AI surpasses human creativity overall, but rather that using AI as an aid slightly raises the evaluation of average proposals.

In the same meta-analysis, the diversity of proposals generated by human+AI was g=−0.863. [1] However, this paper is an unreviewed preprint, the integration of diversity is based on a small number of effect sizes, and the heterogeneity between studies is high at I²=93.70%. The direction should be taken seriously, but it cannot be said that homogenization of the same magnitude occurs across all models, tasks, and dialogue designs.

Doshi and Hauser randomly assigned 293 writers to conditions with no AI proposals, one AI proposal, or up to five AI proposals, and had another 600 people evaluate short stories. For writers with originally low creativity, access to five AI proposals increased novelty by 10.7% and usefulness by 11.5%. [2] On the other hand, works in the AI-assisted condition were more similar to each other than works by humans alone. This is a structure where rational choices for individuals narrow the exploration range for the collective.

Moon et al. compared 2,200 college admission essays by humans and GPT-4 in three pre-registered studies. When adding one work, human-written essays added more new ideas, and the diversity gap widened as the number of works increased. The difference did not disappear even when prompts or generation parameters were changed. [3] However, this is a comparison between AI-only generation and human writing, and does not directly measure interactive co-creation.

Wenger and Kenett in 2026 compared 102 people and 22 LLMs on three standard creativity tasks, reporting that LLM responses were more semantically similar to each other than human responses. [5] The direction did not change even in analyses focusing on one model series at a time. However, this is a study measuring AI-only output of models provided via public APIs, and does not directly show the diversity of co-creation where people add their own experiences or practical deliverables.

In a large-scale observational study of digital art, analyzing over 4 million works and over 53,000 artists, results showed that while productivity and evaluation increased after AI adoption, the novelty of average content and visual expression decreased. Maximum content novelty increased, indicating that while some people expanded the exploration upper limit with AI, there is asymmetry where many works cluster in known areas. [4]

Even if final proposals are similar, whether only the same proposals were generated in the generation stage, whether evaluators selected only similar proposals, or whether a few proposals were eliminated in feasibility review are separate issues. The recommended diagnosis saves independent initial proposals, all proposals after AI expansion, primary selection, and final selection, comparing the number of premises and overlapping clusters at each stage. This is not new empirical evidence.

For dissection, one can use methods such as not informing some evaluators of the AI usage condition, comparing the same initial proposals with and without AI expansion, and comparing forcing a few proposals to pass to the next stage. If results narrow immediately after generation, input and generation conditions are candidates; if narrowing occurs during selection, evaluation criteria are candidates; if narrowing occurs during verification, resource constraints are candidates. Do not determine the cause with one comparison, but seek reproducibility.

This is a hypothetical example. If initial proposals include five customer hypotheses, and five are maintained even after AI expansion, but concentrated into one during the evaluation stage, the model alone cannot be treated as the cause of homogenization. The possibility that evaluators equated clarity with usefulness remains. By saving proposal sets by process, one can consider not only prompt changes but also correction candidates for selection criteria and verification budgets.

If weights are assigned to quality and diversity to form one score, the choice of weights will make minority proposals invisible. In the recommended selection, separate proposals that have quality superior to others without being inferior in distance, and proposals that cover unique premises even if quality is slightly lower, leaving them as a non-dominated candidate set. This is not a statistical optimal solution, but a decision-making tool to avoid confusing exploration and selection.

This is a hypothetical example. Suppose proposal A has high usefulness but is close to existing proposals, while proposal B has slightly lower evaluation but targets different customers and technologies. Before dropping B due to average score, check if B's weaknesses can be improved with small-scale verification. The condition for survival is not to protect proposals that are different but useless, but that unique premises are verifiable and failure does not break the budget of other proposals.

In the recommended meeting, before deciding on the final adopted proposal, list the quality axis, distance from existing proposals, unique premises, and next required evidence for each candidate. Check not the number of proposals, but how many mutually non-substitutable hypotheses remain. Minority proposals that do not improve in quality after verification may be dropped, but the reasons for selection and un-covered areas are carried over to the next exploration.

• Candidate ID: Connect from initial proposal through AI expansion, selection, and verification in the same lineage
• Quality Axis: Keep novelty, usefulness, and feasibility as separate evaluations
• Distance Basis: Specifically describe different customers, causal hypotheses, and solution principles
• Next Evidence: Set deadlines for keeping minority proposals, verification costs, and rejection conditions

The practical implication is not to exclude AI from creative work. Since consulting the same AI from the beginning creates a common anchor, first have people independently define problems and create initial proposals, then use AI for falsification, expansion, and recombination. Separate generation and selection, and have evaluators record not only proposal quality but also reasons for differences from existing proposals.

Diversity cannot be measured by the number of prompts. Cluster proposals using semantic vectors or manual classification, and check the concentration rate into the largest cluster, overlap rate, number of independent premises, and the customers, technologies, and time axes covered by minority proposals. Research shows that AI use does not necessarily harm creativity, but optimizing only quality improvement may lead to loss of collective exploration. [1][3]

The exploration record keeps the model, prompt, and pre-AI independent ideas as identifiers, while separately setting minima for quality and independent premises and maxima for largest-cluster concentration and overlap. If quality or independent premises miss a minimum, AI-assisted selection stops and returns to independent human ideation; if concentration or overlap exceeds a maximum, further generation with the shared model stops. Portfolio reviewers may restart only after the next review confirms recovery of both quality and exploration breadth with added premises.

• Divide before AI: Each person independently defines problems and creates initial hypotheses
• Divide inputs: Have AI explore from different materials, roles, and counter-hypotheses
• Separate generation and selection: Do not immediately adopt the first plausible proposal
• Measure quality and distance: View average evaluation and overlap between proposals simultaneously
• Protect minority proposals: Keep proposals with different premises in the verification framework even if average scores are low