Source-linked AI summary
Collective creativity in hybrid societies
Mason Youngblood, Katie Mudd, Manuel Anglada-Tort, Cameron Jones, Elena Miu, Diana Omigie, Margaret Schedel
TL;DR
Generative AI raises whether creativity belongs to individuals or interacting populations, while novelty and diversity remain distinct, incompletely assessed properties. This paper frames creativity as an emergent property of human–algorithm collectives and finds that composition determines whether AI increases novelty while preserving or narrowing collective diversity.
Problem
Research lacks agreed population-level assessment of how embedding AI in creative collectives changes collective diversity and individual novelty.
Method
The paper synthesizes creativity, cognitive science, computational social science, and cultural-evolution perspectives to analyze hybrid human–algorithm collectives.
Results
Transmission-based studies find hybrid groups outperform human-only and AI-only groups while retaining more exploratory variation than AI-only groups.
Takeaways & Limitations
Creativity outcomes depend on collective composition, including which agents participate, their proportions, and how they connect.
Takeaways & Limitations
The paper notes that generative AI carries substantial environmental costs through energy use, carbon emissions, water consumption, and infrastructure.
Abstract
from arXiv · showhide
Generative AI is changing how cultural artifacts are created and circulated, and with it our understanding of creativity itself. Researchers disagree about whether these tools enrich or impoverish culture, and we argue that much of that disagreement comes from conflating two distinct components of creativity: novelty, a property of single artifacts, and diversity, a property of populations. We argue further that creativity in the context of generative AI is best understood as a property of hybrid collectives, or populations of interacting people and algorithms, rather than of individuals. AI-assisted ideation reliably raises the novelty of individual output while narrowing diversity in the aggregate, but this is not an inevitable consequence of putting machines in the loop. Because humans and models search in complementary ways, mixed groups can outperform and out-diversify groups of either kind alone, and machine-discovered solutions can enter human culture and persist there. What decides the outcome is composition: which agents are present, in what proportion, and how they are connected. The question is no longer whether AI helps or harms creativity, but which mixtures let individual gains accumulate without eroding collective diversity.
1 INTRODUCTION · 2 REFRAMING CREATIVITY · Beyond the lone genius to the creative collective
Generative AI reframes creativity as an emergent property of systems connecting people, while distinguishing novelty in individual artifacts from diversity in populations. This perspective shifts attention from lone creators to collective structures, composition, and interaction.
- 1 INTRODUCTION: Generative AI’s integration into writing, music, images, and code renews the question of whether creativity belongs to individuals or the systems connecting them.The paper argues that creativity in the age of generative AI is best understood as an emergent property.
- 2 REFRAMING CREATIVITY: Novelty measures how far a single artifact departs from a reference set, whereas diversity measures how much artifacts within a set differ from one another.Treating these measures as equivalent can obscure cases where a tool raises novelty while lowering diversity; population-level assessment remains unsettled.
- Beyond the lone genius to the creative collective: Creativity was traditionally located in individual brains through studies of personality, cognition, and neural correlates.This lone-genius view shaped much of twentieth-century psychology and historical accounts of invention and art.
- Beyond the lone genius to the creative collective: Collective-creativity research instead describes innovation as emerging through cultural learning, idea recombination, and distributed interaction across societies and groups [10–13].Independent near-simultaneous discoveries and the rise of teams as sources of high-impact work support this relational account.
- Beyond the lone genius to the creative collective: Solo performance predicts group performance only weakly, with correlations of r = 0.18–0.23 versus 0.38–0.54 between group conditions [14].Thus, assembling creative individuals does not guarantee a creative collective.
- Beyond the lone genius to the creative collective: Collective creativity depends on network size, topology, clustering, information efficiency, sociality, transmission fidelity, and cultural variance [10, 15–19].These features shape consensus, problem-solving, collective intelligence, and the aggregation of distributed knowledge, but existing evidence mainly concerns human groups.
Toward hybrid creativity: AI as a partner
Hybrid creativity depends on how AI agents alter collective dynamics, not merely individual behavior. Mixed systems can become more productive when beliefs about the partner and collaboration design support creative risk-taking and shared agency.
- Toward hybrid creativity: AI as a partner: AI agents can reshape group dynamics: centrally placed bots injecting small amounts of random noise helped human groups escape coordination-game dead ends, unlike noisier or more peripheral bots22.The effect depended on the bots’ network position and noise level.
- Toward hybrid creativity: AI as a partner: Beliefs about an AI partner change behavior: musicians expecting an AI collaborator take greater creative risks because they feel less judged, while users recast tools as friends, students, or managers24,25.The same tool can be interpreted through different social roles, changing how people behave in collaboration.
- Toward hybrid creativity: AI as a partner: Turn-taking design matters as much as belief, because meaning in human–AI collaboration accumulates across exchanges rather than from a single static prompt29.Rigid one-shot tools can foreclose the shared agency required for productive collaboration.
3 MECHANISMS OF HYBRID CREATIVITY
Hybrid creativity is organized around the cultural-evolutionary cycle of variation, selection, and retention, spanning the generation, evaluation, and transmission of creative variants. Novelty arises through combination, search within constraints, and transformation of the search space, which form a separate axis from the production stages.
- 3 MECHANISMS OF HYBRID CREATIVITY: The framework organizes hybrid-creativity mechanisms around cultural evolution’s cycle of variation, selection, and retention.Creative production proceeds by generating variants, evaluating them, and transmitting the survivors.
- 3 MECHANISMS OF HYBRID CREATIVITY: Creative production has three stages: generating variants, evaluating them, and transmitting the survivors.
- 3 MECHANISMS OF HYBRID CREATIVITY: Novelty arises through combination, search within a space of constraints, and transformation of the space itself, independently of those stages.Stages and mechanisms are separate axes rather than a nested hierarchy.
Idea generation
AI assistance raises the creativity of individual ideas while narrowing diversity across groups, because shared sampling concentrates outputs around common high-density concepts. However, transmission-based hybrid groups can outperform human-only and AI-only groups, showing that collective creativity depends on how people and models interact.
- Idea generation: When many people sample from the same generative distribution, shared sampling raises per-item quality while reducing between-item variance35.Individual participants’ idea diversity was similar with ChatGPT and a non-AI tool, locating the narrowing at the group level; mixing models, adding noise, or pairing generation with search can recover diversity.
- Idea generation: Transmission-based designs show that alternating humans and models can outperform human-only and AI-only groups while preserving more exploratory variation than AI-only groups27.This contrasts with studies that pool independently produced outputs and therefore measure individual work more directly than population dynamics.
Selection through evaluation
Selection shapes which generated ideas survive, operating both within creators’ iterative filtering and across creators as population-level selection pressure. AI may generate or evaluate options, while shared evaluation tools can correlate choices and potentially reduce novelty.
- Selection through evaluation: Selection operates at two levels: creators filter their own options across iterations, while independent choices across creators impose population-level selection pressure.These levels are often blurred in the literature.
- Selection through evaluation: Within creators, AI can generate many options for later selection, as in the AI Song Contest, or evaluate and rank options on the creator’s behalf [42].For students revising designs after rating GPT-3.5 proposals, expertise and baseline creative ability predicted revision quality, whereas accuracy in judging the AI’s ideas did not clearly do so.
- Selection through evaluation: Across creators, shared models make selection pressures correlated, while evaluation tools emphasize generation and non-experts may generalize from single successes or failures [35, 45, 46].The survey found 45% of publications focused on generation versus 18% on evaluation, and selection can work against novelty.
Transmission and collaboration · 4 HYBRID CREATIVITY IN PRACTICE
Collective creativity depends on how social information, network structure, and learning biases shape transmission between humans and machines. These mechanisms vary by problem difficulty and participant expertise, producing domain-specific effects in practice.
- Transmission and collaboration: Partially connected networks preserve variation for complex innovation, whereas fully connected networks converge quickly and suit easier tasks.Human creativity relies on social information, and network shape determines how widely solutions spread.
- 4 HYBRID CREATIVITY IN PRACTICE: Social learning strategies guide whom humans attend to, while machine-mediated recommendations can redirect that attention and alter collective innovation.These interventions operate through both network structure and the biases people use to interpret social information.
- Transmission and collaboration: Machines reshape creative transmission by altering network structure and making marginal views appear widespread through repeated algorithmic exposure.They can also promote content independently of its local prevalence, recruiting conformist bias on behalf of minority positions.
- Transmission and collaboration: Simple bots improve discovery on easy semantic landscapes by rebroadcasting neighbors’ guesses across distant network regions, but lose that advantage when search becomes difficult to track.The benefit depends on whether people can identify which regions of the problem space are worth searching.
- Transmission and collaboration: More capable tools broaden participation in creative domains, but professional artists still extract more from the same systems than laypeople.Lowering entry barriers changes who participates without fully equalizing the quality of what participants produce.
- 4 HYBRID CREATIVITY IN PRACTICE: The effects of these transmission mechanisms differ across domains, as illustrated by the paper’s three practical cases.The cases apply the preceding mechanisms to distinct creative settings rather than treating machine involvement as uniformly beneficial or harmful.
In science and technology · In art, music and literature
Across science, technology, and the arts, generative AI expands creative production but remains shaped by human judgment, prompting concerns about expertise, originality, quality, and diversity. Its collective effects depend on how people use, curate, and evaluate machine-generated outputs.
- In science and technology: AI research tools are usually compliant rather than provocative, tending to mirror users’ proposals and amplify their existing intentions. [58]Generative AI has entered research as an ideation partner whose outputs researchers interpret. [57]
- In science and technology: Models match or exceed humans on fluency and originality tests, but humans retain advantages in flexibility, human-rated creativity, and complex-skill transfer. [60,61]Systems designed to augment human judgment may therefore serve science better than systems designed to emulate it. [59]
- In science and technology: AI adoption is concentrated among PhD students and early-career academics, chiefly for writing, while non-experts often prompt unsystematically and anthropomorphize models. [62,46]Bypassing expertise-building tasks may leave trainees unable to recognize when AI is wrong.
- In art, music and literature: In visual art, originality depends heavily on human curation before and after generation, including training-set assembly and selecting which output to present. [63]Interactive machine-learning systems make this curatorial layer explicit through examples and iterative refinement. [24,64,6]
- In art, music and literature: Algorithmic music systems recombine human-created structural constraints, while contemporary creators similarly decompose and recombine AI-generated material. [66,42]This pattern makes human construction and selection central to the resulting compositions.
- In art, music and literature: AI-assisted books grew from near zero in 2022 to more than half of new releases in 2025 as Amazon releases roughly tripled and average quality fell.Higher volume nevertheless increased moderately good books enough to raise estimated consumer surplus by about 7%, an economic proxy rather than a literary-merit judgment. [69]
- In art, music and literature: The literary volume increase raised collective consumer value despite lower quality per item, while plot elements recur more in LLM stories and AI images can seem polished but “soulless.” [70,71]This illustrates a tension between individual-item quality and the value or diversity of the larger pool.
In games and other domains
Historical game records caution against inferring durable cultural change from short-term novelty: AlphaGo briefly disrupted professional openings, but produced only modest, short-lived change over four centuries of play. Opening diversity is highest at intermediate player-population sizes.
- Games: Over four centuries of recorded play, superhuman AI caused only a modest and short-lived disruption despite initially prompting professionals to depart from known openings more often and earlier [73].Expectations that standard openings would disappear after AlphaGo’s 2016 victory over Lee Sedol were not borne out [72].
- Games: Professional players initially departed from known opening sequences more often and earlier after AlphaGo defeated Lee Sedol in 2016 [38].This early behavioral shift did not amount to durable change across the longer historical record.
- Games: Opening diversity is highest at intermediate player-population sizes and falls outside that range.
5 CONSEQUENCES · Convergence and divergence
Individual gains in creativity do not determine population-level outcomes: collective composition, tools, and cultural context decide whether AI-assisted creation converges or diverges. Mixed human–AI systems can sustain or expand diversity, but repeated self-training and certain transmission structures can narrow it.
- 5 CONSEQUENCES: Population-level consequences are separate from individual-level gains and depend on how the collective is assembled.The composition of agents, tools, and cultural context largely decides whether outcomes expand or compress the creative space.
- Convergence and divergence: Mixed human–AI networks can overtake initially more diverse AI-only story chains in diversity over successive generations.AI-only chains converge as agents iteratively select and rewrite neighboring stories, while human continuity appears to offset that convergence.
- Convergence and divergence: AI can act as a blender whose effects depend on who is holding it, expanding or compressing the creative space through collective composition.Transmission studies therefore treat composition, tools, and cultural context as major determinants of outcomes.
- Convergence and divergence: Model collapse is a structural risk because repeated self-training on model-generated text and images can narrow outputs toward a few dominant traits [20].Convergence is partly a design choice: evolutionary search, explicit blending, and small curated training sets can sustain exploration or make model idiosyncrasies expressive signatures [75].
- Convergence and divergence: Hybrid human–AI groups outperformed AI-only groups in collective search while retaining more variation [27].Bots injecting small amounts of noise also pulled human groups out of coordination dead ends [22].
- Convergence and divergence: Machine agents can circulate strategies that human populations essentially never discover on their own [41].They also make tractable a scale of distributed search that neither humans nor machines could achieve alone.
Ownership, attribution and bias · Environmental costs
Hybrid creativity unsettles ownership and responsibility while reproducing biases and shifting creators’ sense of agency. Its material costs span labor, resources, infrastructure, and energy, but smaller task-specific models can reduce those costs.
- Ownership, attribution and bias: Training-data scraping, exploitative labor, and opaque corpora make it nearly impossible to trace whose work appears in outputs, leaving authorship and copyright unsettled76,77,79–81.Artists’ objections are also difficult to verify as effective opt-outs78.
- Ownership, attribution and bias: Model pipelines reproduce bias, with image generators misrepresenting or erasing demographic groups82–84.
- Ownership, attribution and bias: AI use can lower creators’ self-assessed abilities and responsibility for generated ideas, while inconsistent model behavior frustrates deliberate changes35,42,71,74,87.Computer scientists interpret malfunctions as failures, whereas new media artists may treat them as playful sources of ideas57.
- Environmental costs: Hybrid creativity depends on mined hardware materials, annotation and moderation labor, data-center land and water, and the electrical grid88.
- Environmental costs: Large-model training requires massive energy, while everyday inference dominates deployment costs and multipurpose models cost orders of magnitude more per query than task-specific models89–91.
- Environmental costs: Creative costs are embedded in continuously operating shared infrastructure and distributed unevenly across populations95,96.The burden is detached from the person performing the creative act.
- Environmental costs: Smaller, task-specific models can perform much of the same work using a fraction of the energy, showing that hybrid work need not have such high costs91.
6 FUTURE DIRECTIONS
AI changes the structure and dynamics of creative work, making the central open question which arrangements of people and machines preserve collective diversity. Future research should test these arrangements at population scale while accounting for the environmental costs of generative methods.
- Changing creative work: AI changes creative-work dynamics: novice music producers spend less time preparing and more time rapidly generating and selecting options [67].Individual assistance is well established, but its aggregate effects remain unresolved.
- Population-scale designs: Population-scale studies are needed because existing evidence mostly examines individuals or dyads using a single tool, not many humans and models interacting over multiple rounds.Such designs are necessary to determine whether local gains and global losses add up, although they are difficult to scale across structural, interactional, and individual axes [21,23,97].
- Resource constraints: Environmental footprint should shape the research agenda, including decisions about when resource-intensive generative methods are warranted.All three research priorities face the cost burden identified in Section 5.
- Reframing the research question: The field should ask which mixtures and coupling patterns let AI-assisted individual gains accumulate without eroding collective diversity.This reframes the debate from whether AI helps or harms creativity to which human–machine arrangements preserve diversity.