Does generative AI improve creative quality at the expense of collective diversity?

Experimental findings, measurement validity and the limits of transfer to marketing

This retrospective review considers source versions available by 18 August 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: generative AI; creative cognition; collective diversity; semantic similarity; human–AI collaboration; marketing

1. Introduction and research questions

Discussion of AI-supported creative work frequently combines claims about faster production, better individual outputs and a wider range of ideas. These claims concern different properties of a process. An output can receive a favourable evaluation without establishing that a collection of outputs has become more varied. Conversely, variation across a collection does not establish that any particular item is useful. Quality and diversity therefore need separate assessment.

The central question is whether improved creative quality necessarily comes at the expense of collective diversity. Three subsidiary questions guide the analysis. First, what is actually measured when an experiment reports “creativity”? Second, whose diversity is being assessed: one person's idea set, the outputs of multiple people, or a model's sampled generations? Third, does an observed benefit remain when assistance is removed? These distinctions determine which results can be compared.

This review synthesises existing findings without proposing a new causal hypothesis or theoretical model.

2. Review method and evidential boundaries

2.1 Search and selection

Targeted searches were conducted on 21 October 2026 using combinations of “generative AI”, “individual creativity”, “collective diversity”, “homogenization”, “creative ideation”, “divergent thinking” and “measurement validity”, together with the titles of relevant studies. Sources were located through publisher sites, conference proceedings, university repositories and author manuscripts. Foundational measurement references were included where necessary to interpret the experimental outcomes. The publication date of an underlying study is distinct from the preparation date of this review.

Selection was purposive. Priority was given to reports that distinguish individual and collective outcomes, compare assistance arrangements, assess performance after assistance, or clarify a measurement used by the creative-task studies. Adjacent productivity studies are included only to distinguish professional task quality from creative diversity. There was no comprehensive subscription-database search, preregistered review protocol, duplicate independent screening or formal quantitative risk-of-bias instrument. Eighteen references therefore describe this review's selected evidence base, not an exhaustive inventory of the field.

2.2 Extraction and comparison

For each central report, the review records the task, intervention, comparator, outcome and principal interpretive constraint. Sample sizes are reported only where confirmed in the consulted material. Participant counts are distinguished from counts of texts, ratings and ideas. Numerical effects are not combined across tasks or converted into a common effect size. Percentage changes, percentage-point differences and standardised effects are reported on their original scales.

Publication versions also matter. An early working paper and its later journal article may contain different samples or reporting conventions. Where only an abstract or publication record was accessible, the discussion is restricted to what that material establishes. It does not imply examination of the statistical model, supplementary materials, exclusions or underlying data. The source-access register identifies the material consulted for each reference.

2.3 Synthesis approach

The narrative comparison considers the design and outcome behind each result; a count of statistically significant findings would obscure those differences. A null result is not interpreted as equivalence; a positive result is not automatically a commercially important result. A confidence interval from one task cannot define the uncertainty of another task. No meta-analysis is attempted because the interventions, units of analysis and operational definitions differ materially. The conclusions distinguish direct findings from the limitations of transferring those findings.

3. Defining the outcomes before comparing results

3.1 Originality and appropriateness

Runco and Jaeger's account of the standard definition makes originality and effectiveness joint requirements for creativity [1]. In this context, effectiveness concerns suitability or value within the relevant domain; it is not synonymous with advertising return on investment. An unusual response can fail the suitability requirement. The definition supplies a conceptual distinction, not an experimental estimate of AI's impact.

A separate issue is whether a test captures the domain of interest. Erwin, Tran and Koutstaal studied 99 design students using several divergent-thinking measures and a product-ideation task. Their analyses showed that related measures did not assess identical constructs, and that task-related and individual factors contributed to performance [2]. Their study is not an AI intervention. Its relevance is methodological: substituting one creativity score for another requires justification rather than an assumption that all such measures are interchangeable.

3.2 Associative distance and its interpretation

Olson and colleagues introduced a brief task in which participants produce ten words as unrelated as possible, using semantic distance to score their responses. The research involved 8,914 participants and linked this Divergent Association Task to other creativity measures [3]. The validation concerns a human assessment. Applying the score to models, professional expertise or commercial communication requires further validation.

Sentence-BERT provides one technical route for representing sentence meaning as vectors that can be compared using cosine similarity [4]. Such representations make large-scale text comparison tractable. However, a similarity score has a specified representation and comparison rule: it is not a direct observation of audience interpretation. “Less semantically similar”, “more original”, “more suitable” and “more persuasive” should therefore remain separate labels unless an empirical validation connects them in the relevant setting.

3.3 Within-person and between-person diversity

A collection of ideas can be diverse relative to other ideas from the same participant, yet overlap with the collections produced by other participants. Anderson, Shah and Kreminski explicitly separate those levels in their creative-ideation study [5]. The distinction prevents a common inference error: a person experiencing a wider range of suggestions does not necessarily imply greater variety across a population of people receiving suggestions.

For reporting purposes, the unit should be named wherever “diversity” appears. A per-person collection, a condition-wide collection, and a stream evolving through successive participants are different objects of analysis. They may also use different aggregation rules. Each unit answers a different question. The same principle applies when comparing repeated generations from one model with human outputs collected from many individuals.

4. Direct evidence on assisted creative outputs

4.1 Short fiction: gains in rated quality with convergence across stories

Doshi and Hauser analysed 293 eight-sentence stories, with 600 evaluators providing 3,519 evaluations. Writers were randomly assigned to no AI access, access to one idea, or access to up to five GPT-4 ideas. AI access improved assessed novelty and usefulness, with larger gains among writers scoring lower on a baseline divergent-association measure. Stories also became more similar to other stories within the same assistance condition. Writers could not customise the prompt or repeatedly converse with the model. The result establishes a conditional effect in constrained fiction, not a universal property of all creative collaboration [6].

4.2 Argumentative essays: the model comparison changes the finding

Padmakumar and He compared argumentative writing without assistance, with a base GPT-3 model, and with feedback-tuned InstructGPT. Their conference abstract reports reduced lexical and content diversity with InstructGPT, but not the base model. They attribute the observed reduction mainly to model-contributed text rather than changes in the human-contributed portions [7]. Even within this writing task, the two models produced different diversity outcomes. The abstract alone does not establish which training component causes every difference.

4.3 Alternative creativity support: quantity and detail do not settle collective variety

In a 36-participant within-subject comparison, Anderson and colleagues examined ChatGPT against the non-AI Oblique Strategies creativity tool. Participants produced 1,271 ideas across four ideation tasks. ChatGPT was associated with more numerous and detailed ideas and greater group-level semantic similarity; individual-level homogenisation did not show the same difference. Participants also reported less responsibility for their ideas [5]. Because both conditions offered creativity support, the comparison estimates the difference between those tools. The small participant sample limits broad population claims.

4.4 Dynamic exposure: collective diversity can increase

Ashkinaze and colleagues used a dynamic experiment involving over 800 participants across more than 40 countries. Participants saw examples originating from AI or earlier participants before proposing an unusual use for an object. High AI exposure increased collective diversity without improving individual creativity scores; lower exposure did not produce the same effect. The design measures evolving example exposure rather than unrestricted dialogue with an assistant [8]. Its findings show that greater collective difference can occur without better-rated individual ideas, but cannot identify every reason it differs from the fiction and essay experiments.

4.5 Persona-conditioned inputs: evidence of a possible boundary condition

Wan and Kalman's revised author abstract describes generating 300 story plots using ten diverse AI personas, followed by a human writing comparison. It reports that the varied inputs preserved diversity relative to a human-only baseline, with some evidence of enhancement in the one-plot condition. A journal publication is linked from the repository [9]. Because full-text appraisal remains outstanding, no numerical effect, sample-size inference or robustness claim is extracted here. The result suggests that convergence is avoidable under some conditions, although its transfer to other domains remains untested.

4.6 Human ideation versus AI refinement

Dong and Yakura's July 2026 preprint abstract describes a preregistered metaphor experiment with native and non-native English writers. It reports compressed collective diversity under AI ideation, but preservation under refinement of human ideas. Simulated writer pools did not reproduce the diversity of the human pools [10]. The preprint distinguishes AI-generated starting ideas from AI editing of human ideas. Independent verification and tests in other settings are needed before generalising either result.

5. Performance during assistance and performance afterwards

5.1 Randomised exposure followed by unaided tests

Kumar and colleagues conducted two preregistered experiments involving 1,100 participants, using divergent and convergent tasks. Conditions provided direct LLM answers, coach-like guidance or no assistance, followed by an unassisted test. Assisted performance did not translate into consistent improvements in subsequent independent performance; some comparisons favoured the unassisted control. The experiment measures the effects of that exposure and test sequence; it provides no evidence of permanent damage to creativity [11]. The later test assesses independent performance, which the assisted score alone cannot establish.

5.2 A smaller longitudinal report

Liu and colleagues' preprint abstract reports a seven-day experiment and a thirty-day follow-up involving 61 students, 3,302 ideas and 427 solutions. It reports that creative-performance gains reverted when ChatGPT was unavailable, while homogeneity persisted [12]. The large output count reflects repeated observations from a relatively small participant sample. Full-text assessment and peer-reviewed publication status have not been established here. The report is relevant to persistence, but its short horizon and student sample cannot support a general conclusion about lifelong creative ability.

5.3 What follows from separating these endpoints

An assisted performance gain can have practical value even when unaided performance does not improve. Deliverable quality and learning are separate evaluation criteria. A study intended to assess both needs observations during assistance and after its removal; measuring only one leaves the other unresolved.

Follow-up duration also limits the interpretation. An immediate test captures a different interval from a thirty-day follow-up. Claims about learning, dependence or deskilling must specify the observed period; a brief intervention cannot establish how performance develops over a working career.

6. Machine benchmarks and measurement validity

6.1 Comparing distributions rather than one human average

Bellemare-Pepin and colleagues benchmarked language models against a dataset of 100,000 humans and assessed creative writing as well as divergent association. Their 2026 journal article reports that models can exceed average human performance on the association task while remaining below the more creative segment of human participants. Prompt strategies and temperature also altered semantic divergence [13]. This is a benchmark comparison, not a trial of AI-assisted workers. Human-machine ranking on a selected test cannot establish a treatment effect on human creative work.

6.2 High distance scores can fail a contextual test

Nakajima, Zuiderveld and Pezzelle examine the validity of applying divergent association tests to models. Their manuscript reports that non-creative baselines can outperform models on the original task and proposes a conditional task that adds contextual appropriateness. The repository identifies acceptance to Findings of EACL 2026 [14]. A distance metric therefore needs a validity check: can it distinguish useful novelty from irrelevant variation?

6.3 Predictive validity is construct-specific

Schapiro and colleagues' 2026 preprint compares creativity tests against creative writing, divergent thinking and scientific ideation benchmarks. It reports differing predictive performance across target constructs and no single test performing well across all of them. Its proposed combined test is an author-reported contribution, not an independently validated standard adopted here [15]. The score must be interpreted against the outcome it predicts. The preprint's findings cannot justify treating one metric as a universal measure of creative achievement.

6.4 A technical caution about comparing diversity scores

Cosine similarity compares vector directions; sentence embeddings make that comparison possible for textual material [4]. The meaning of the resulting statistic still depends on the embedding model, the text being represented and the aggregation rule. A distance to a centroid and an average pairwise distance do not describe exactly the same quantity. Both calculations can be useful, but reproducing an analysis requires the embedding model and aggregation rule. The label “semantic diversity” supplies neither.

For this reason, no value from one paper is translated into an equivalent percentage of “creative variety” in another. Such conversion would require a shared scale and a defensible mapping that this review has not established. Larger numerical distances also cannot be equated with more meaningful strategic differences. Assessing strategic differences requires criteria beyond the embedding calculation.

7. Adjacent evidence: productivity, professional quality and fixation

7.1 Professional writing productivity

Noy and Zhang's published study reports a preregistered experiment with 453 college-educated professionals, a 40% reduction in task time and an 18% increase in rated output quality. The earlier March 2023 working paper consulted here reports 444 participants and standardised effects instead [16]. The published and working-paper estimates concern different reported samples and scales. Neither result is a direct estimate of collective creative diversity. The study establishes improvements in the specified professional writing tasks, not a universal productivity multiplier or proof of improved marketing outcomes.

7.2 Uneven effects within knowledge work

Dell'Acqua and colleagues' published report describes a preregistered experiment with 758 consultants. On a set of eighteen tasks within the assessed capability frontier, AI users completed more tasks and worked more quickly, with improved quality. Performance worsened on a selected task outside that frontier [17]. The relevant comparison is task-specific. “Knowledge work” does not identify a uniform treatment response, and apparent similarity in task difficulty is not sufficient to infer equal benefit. This is adjacent evidence, not a direct test of campaign creativity.

7.3 Fixation predates generative AI

Jansson and Smith's 1991 research investigated design fixation: adherence to ideas or concepts that constrains conceptual design [18]. The study places the influence of examples on design in a longer research history. It predates contemporary language models, so it supplies conceptual context rather than direct evidence of the mechanisms in the AI experiments.

8. Comparative synthesis

The studies cannot be reconciled by replacing their separate outcomes with a single claim that AI is either beneficial or harmful to creativity. Some assess a product; others assess the spread of products or performance after the tool is removed. Some compare support tools; others compare a tool with no assistance. The task, comparator and outcome must accompany each reported effect.

Equally, finding differences between study designs does not demonstrate which design feature explains the difference in results. Models, tasks, samples, interfaces and assessments may all differ simultaneously. Establishing a moderator requires an appropriate comparison within a study or a sufficiently supported synthesis across studies. The present evidence does not isolate a workflow that eliminates the documented risks.

9. Boundaries of application to marketing

Three inferential steps must remain visible when applying this literature to marketing. The first moves from a research task to a professional assignment. The second moves from the professional output to a target audience's response. The third moves from that response to a commercial outcome. Evidence at the first step cannot be silently treated as evidence for the complete chain. Each step changes either the population, the material or the outcome requiring evaluation.

For example, a favourable rating of a fictional story is a judgement under the study's assessment conditions. A diversity statistic describes the relation among texts under its representation rule. Neither directly measures whether a consumer identifies an advertiser, comprehends an offer, forms a preference or purchases incrementally because of exposure. Marketing success or failure remains unmeasured by those outcomes.

The same caution applies to asset quantity. A larger number of executions is a production count. It does not itself specify the number of distinct concepts, the variety of consumer situations addressed or the degree of brand consistency. Those properties require operational definitions. Whether greater diversity is desirable depends on the assignment. Asset counts alone answer neither the quality question nor the diversity question.

10. Limitations and conclusion

This is a selective narrative review; it has not undergone independent extraction or external peer review. Several sources were accessible only at abstract or publication-record level. It does not reproduce analyses, audit datasets, assess every model release, estimate publication bias or establish how common each finding is in the complete literature. The inclusion of preliminary manuscripts expands coverage but does not give those reports the same evidential status as a completed, independently appraised body of replications.

The resulting answer is conditional. Selected experiments show that AI assistance can improve specified outputs while collective similarity increases; other arrangements report preserved or greater diversity. Assisted success does not establish improved subsequent unaided capability. Differences in measurement further prevent a single “creativity” score from resolving all questions. The literature supports precisely bounded conclusions about tasks and interventions. It does not establish an invariant quality–diversity trade-off, a universal creative benefit, or an advertising-effectiveness claim.

Source access and appraisal register

“Full text consulted” means the manuscript was available for examination; it does not mean independent replication or a completed formal risk-of-bias review. “Abstract/record” means that claims were restricted to the consulted author or publisher summary. Where an author version differs from the version of record, the distinction remains explicit.

References