pith. sign in

arxiv: 2605.17141 · v1 · pith:ELV4PJQ6new · submitted 2026-05-16 · 💻 cs.AI

Dynamics of collective creativity in AI art competitions

Mason Youngblood , Jeff Nusz , Joel Simon This is my paper

Pith reviewed 2026-05-20 14:31 UTC · model grok-4.3

classification 💻 cs.AI
keywords collective creativityAI artcultural evolutionnoveltycomplexityremix dynamicsthematic attractorsgroup size
0
0 comments X

The pith

In AI art remix parties, images simplify and converge on common themes as novelty and group size shape creative output.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The study examines how groups collaboratively create art through iterative remixing with AI assistance. Data from numerous remix parties shows that images tend to become simpler over time and cluster around shared thematic ideas. Novel starting images lead to more original and intricate follow-ups that gain popularity, but participants tend to build upon more conventional ones instead. Larger groups foster greater originality at the expense of detail and complexity. A sympathetic reader would care because this sheds light on the processes driving cultural change in emerging human-AI creative networks, which differ from traditional one-way transmission chains.

Core claim

Analysis of 130,882 images across 368 remix parties reveals that images become simpler and converge toward common thematic attractors. More novel parent images generate more novel and complex children that attract more likes, yet users prefer to remix less novel and complex images. Larger remix parties produce more novelty but with lower complexity.

What carries the argument

Branching lineages of iteratively remixed images in collective AI art competitions, tracked through novelty and complexity metrics.

If this is right

  • Images in collective creative processes simplify and align with popular themes over successive remixes.
  • Novel parent images foster greater novelty and complexity in subsequent generations, boosting user engagement through likes.
  • Users show a preference for remixing less novel and less complex images despite the appeal of novelty.
  • Larger groups in remix activities increase overall novelty produced while decreasing the complexity of outputs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • These patterns of simplification and attraction could influence design choices in other collaborative AI creative tools.
  • Balancing group size might help maintain both novelty and complexity in cultural production systems.
  • Further studies could test if similar attractor dynamics appear in non-art domains like music or story remixing.
  • The user preference for simpler images may create feedback loops that stabilize certain cultural forms.

Load-bearing premise

The 130,882 images and 368 remix parties provide an unbiased view of creativity dynamics where the chosen metrics for novelty and complexity accurately reflect the key properties without major biases from user choices or platform effects.

What would settle it

If new observations from similar remix activities show images becoming more complex rather than simpler or failing to converge on common themes, that would challenge the central findings.

read the original abstract

Creativity is a fundamental aspect of how culture evolves, yet the mechanisms by which groups produce novelty are notoriously difficult to infer from the historical record. Iterated learning experiments have shown that cultural transmission reliably distorts artifacts toward the inductive biases of learners, but most of this work uses linear chains between human participants, leaving open how these dynamics play out in the networked, human-AI systems that increasingly shape cultural production. In this study, we leverage one such system, Artbreeder, which hosts daily "remix parties" where users iteratively build on each other's work from a single seed image, producing branching lineages of human-AI co-created images. We analyze a dataset of 130,882 images from 368 remix parties over 13 months and find that images become simpler and converge toward common thematic "attractors" (e.g., steampunk scenes, alien architecture). We also find that while more novel "parent" images produce more novel and complex "children" that attract more likes, users paradoxically prefer to remix images that are less novel and complex. Finally, larger remix parties produce more novelty at the cost of lower complexity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript analyzes a dataset of 130,882 images from 368 remix parties on the Artbreeder platform over 13 months. It claims that images in these branching lineages become simpler and converge toward common thematic attractors, that more novel parent images produce more novel and complex children which receive more likes, yet users prefer to remix less novel and complex images, and that larger remix parties generate more novelty at the expense of lower complexity.

Significance. If the novelty and complexity metrics prove robust and non-circular with the underlying generative model, the study offers a significant empirical contribution to understanding collective creativity in human-AI systems at scale. The use of a large, real-world dataset from networked remixing extends traditional iterated learning experiments and provides falsifiable observations on simplification, convergence, and trade-offs in novelty and complexity.

major comments (3)
  1. The quantification of novelty and complexity is not specified with sufficient detail (e.g., no equations or algorithmic description provided in the abstract or methods overview), which is load-bearing for the central claims about simplification, convergence, parent-child relationships, and party-size effects.
  2. The identification of 'thematic attractors' (e.g., steampunk scenes) appears post-hoc without a clear, pre-specified method for detecting convergence or controlling for platform algorithms and user selection biases that may shape the observed lineages.
  3. Potential circularity between the novelty/complexity metrics and Artbreeder's generative model is not addressed, raising the risk that reported convergence and attractor dynamics are artifacts of the model's inductive biases rather than emergent from collective remixing.
minor comments (1)
  1. The abstract reports directional findings but omits any mention of statistical methods, error bars, or controls, which should be summarized for clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments highlight important areas for improving methodological transparency and addressing potential limitations in our analysis of collective creativity on Artbreeder. We respond to each major comment below, indicating where revisions have been made to the manuscript.

read point-by-point responses

  1. Referee: The quantification of novelty and complexity is not specified with sufficient detail (e.g., no equations or algorithmic description provided in the abstract or methods overview), which is load-bearing for the central claims about simplification, convergence, parent-child relationships, and party-size effects.

    Authors: We agree that the abstract and methods overview would benefit from explicit quantification details to support the central claims. The full algorithmic specifications, including the embedding-based novelty measure and complexity proxy, appear in the Methods section. We have revised the manuscript to include a concise description of these metrics along with the primary equations in both the abstract and an expanded methods overview subsection. This change directly addresses the load-bearing nature of the metrics for the reported findings on simplification, convergence, and party-size effects. revision: yes

  2. Referee: The identification of 'thematic attractors' (e.g., steampunk scenes) appears post-hoc without a clear, pre-specified method for detecting convergence or controlling for platform algorithms and user selection biases that may shape the observed lineages.

    Authors: The referee is correct that the thematic attractors were initially identified via qualitative review of prominent lineages. To strengthen this, we have added a pre-specified quantitative procedure using embedding-space clustering to detect convergence across parties. We have also expanded the discussion to explicitly address platform recommendation algorithms and user selection biases as potential confounds, including sensitivity checks where feasible. These revisions make the attractor analysis more systematic while retaining the original observations. revision: yes

  3. Referee: Potential circularity between the novelty/complexity metrics and Artbreeder's generative model is not addressed, raising the risk that reported convergence and attractor dynamics are artifacts of the model's inductive biases rather than emergent from collective remixing.

    Authors: We have added a new paragraph in the Discussion section that directly engages this concern. The metrics rely on general-purpose visual embeddings rather than Artbreeder-specific generative parameters, and we present supplementary checks showing that simplification and convergence patterns persist under alternative feature representations. We acknowledge, however, that complete separation of model inductive biases from collective human-AI dynamics remains difficult with observational data and have noted this explicitly as a limitation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical patterns derived from independent measurements

full rationale

The paper reports statistical patterns observed in a dataset of 130,882 images across 368 remix parties, including simplification, convergence to thematic attractors, and relationships between parent novelty/complexity and child outcomes or likes. These are presented as data-driven findings rather than predictions derived from equations or parameters fitted to the same quantities. No self-definitional reductions, fitted inputs renamed as predictions, or load-bearing self-citations appear in the abstract or summary that would make the reported dynamics equivalent to the inputs by construction. The analysis treats novelty and complexity as measured properties applied to the generated images, with claims resting on observed correlations rather than tautological mappings.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claims rest on unstated definitions of novelty and complexity, assumptions that remix parties are comparable across time, and that platform data capture genuine collective dynamics without heavy algorithmic filtering.

axioms (2)
  • domain assumption Novelty and complexity can be reliably quantified from image features in a way that reflects human perception of creativity.
    Required to interpret convergence and parent-child effects; location implied in abstract description of metrics.
  • domain assumption Remix parties form independent lineages without significant cross-party influence or platform-wide trends.
    Needed to attribute changes to within-party dynamics over 13 months.
invented entities (1)
  • thematic attractors no independent evidence
    purpose: To describe convergence points such as steampunk scenes or alien architecture.
    Postulated to explain observed simplification and clustering; no independent falsifiable prediction given in abstract.

pith-pipeline@v0.9.0 · 5725 in / 1521 out tokens · 35807 ms · 2026-05-20T14:31:11.306993+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 3 internal anchors

  1. [1]

    Fogarty, L., Creanza, N., and Feldman, M.W. (2015). Cultural evolu­ tionary perspectives on creativity and human innovation. Trends in Ecology & Evolution. https://doi.org/10.1016/j.tree.2015.10. 004

    work page doi:10.1016/j.tree.2015.10 2015

  2. [2]

    Perry, S., Carter, A., Smolla, M., Akçay, E., Nöbel, S., Foster, J.G., and Healy, S.D. (2021). Not by transmission alone: the role of invention in cultural evolution. Philosophical Transactions of the Royal Society B. https://doi.org/10.1098/rstb.2020.0049

    work page doi:10.1098/rstb.2020.0049 2021

  3. [3]

    Boden, M.A. (1998). Creativity and artificial intelligence. Artificial intelligence. https://doi.org/10.1016/S0004­3702(98)00055­1

    work page doi:10.1016/s0004 1998

  4. [4]

    de (2023)

    Frascaroli, J., Leder, H., Brattico, E., and Cruys, S.V. de (2023). Aesthetics and predictive processing: grounds and prospects of a fruitful encounter. Philosophical Transactions of the Royal Society B. https://doi.org/10.1098/rstb.2022.0410

    work page doi:10.1098/rstb.2022.0410 2023

  5. [5]

    Falandays, J.B., Kaaronen, R.O., Moser, C., Rorot, W., Tan, J., Varma, V., Williams, T., and Youngblood, M. (2023). All intelligence is collective intelligence. Journal of Multiscale Neuroscience. https://doi.org/10.56280/1564736810

    work page doi:10.56280/1564736810 2023

  6. [6]

    Buskell, A., Enquist, M., and Jansson, F. (2019). A systems ap ­ proach to cultural evolution. Palgrave Communications. https:// doi.org/10.1057/s41599­019­0343­5

    work page doi:10.1057/s41599 2019

  7. [7]

    Kirby, S., Griffiths, T., and Smith, K. (2014). Iterated Learning and the Evolution of Language. Current Opinion in Neurobiology. https://doi.org/10.1016/j.conb.2014.07.014

    work page doi:10.1016/j.conb.2014.07.014 2014

  8. [8]

    Kelly, P., Winters, J., Miton, H., and Morin, O. (2021). The pre ­ dictable evolution of letter shapes: an emergent script of West Africa recapitulates historical change in writing systems. Current Anthropology. https://doi.org/10.1086/717779

    work page doi:10.1086/717779 2021

  9. [9]

    Anglada­Tort, M., Harrison, P.M.C., Lee, H., and Jacoby, N. (2023). Large­scale iterated singing experiments reveal oral transmis ­ sion mechanisms underlying music evolution. Current Biology. https://doi.org/10.1016/j.cub.2023.02.070

    work page doi:10.1016/j.cub.2023.02.070 2023

  10. [10]

    Marjieh, R., Anglada ­Tort, M., Griffiths, T., and Jacoby, N. (2025). Characterizing the interaction of cultural evolution mechanisms in experimental social networks. arXiv. https://doi.org/10.48550/ arXiv.2502.12847

    work page arXiv 2025

  11. [11]

    Shiiku, S., Marjieh, R., Anglada­Tort, M., and Jacoby, N. (2025). The dynamics of collective creativity in human ­AI hybrid societies. arXiv. https://doi.org/10.48550/arXiv.2502.17962

    work page doi:10.48550/arxiv.2502.17962 2025

  12. [12]

    Acerbi, A., and Stubbersfield, J. (2023). Large language models show human ­like content biases in transmission chain exper ­ iments. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2313790120

    work page doi:10.1073/pnas.2313790120 2023

  13. [13]

    Brock, A., Donahue, J., and Simonyan, K. (2019). Large scale GAN training for high fidelity natural image synthesis. Interna ­ tional Conference on Learning Representations. https://doi.org/ 10.48550/arXiv.1809.11096

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1809.11096 2019

  14. [14]

    Karras, T., Laine, S., and Aila, T. (2020). A style ­based generator architecture for generative adversarial networks. IEEE Transac ­ tions on Pattern Analysis and Machine Intelligence. https://doi. org/10.1109/TPAMI.2020.2970919

    work page doi:10.1109/tpami.2020.2970919 2020

  15. [15]

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Muller, J., Penna, J., and Rombach, R. (2023). SDXL: improving la­ tent diffusion models for high­resolution image synthesis. arXiv. https://doi.org/10.48550/arXiv.2307.01952

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2307.01952 2023

  16. [16]

    Black Forest Labs, Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., et 6 al. (2025). FLUX.1 Kontext: Flow Matching for In ­Context Image Generation and Editing in Latent Space. arXiv. https://doi.org/10. 48550/arXiv.2506.15742

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Zeilinger, M. (2021). Generative adversarial copy ma ­ chines. Culture Machine. https://culturemachine.net/vol­20­ machine­intelligences/generative­adversarial­copy­machines­ martin­zeilinger/

    work page 2021

  18. [18]

    Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. (2023). Re­ producible scaling laws for contrastive language­image learning. IEEE/CVF Conference on Computer Vision and Pattern Recogni ­ tion. https://doi.org/10.1109/CVPR52729.2023.00276

    work page doi:10.1109/cvpr52729.2023.00276 2023

  19. [19]

    ­Y., et al

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W. ­Y., et al. (2023). Segment anything. arXiv. https://doi.org/10.48550/arXiv.2304. 02643

    work page doi:10.48550/arxiv.2304 2023

  20. [20]

    Lo, R.K.L. (2023). denmarf: a Python package for density estima­ tion using masked autoregressive flow. arXiv. https://doi.org/10. 48550/arXiv.2305.14379

    work page arXiv 2023

  21. [21]

    Bürkner, P.­C. (2017). brms: An R Package for Bayesian Multilevel Models Using Stan. Journal of Statistical Software. https://doi. org/10.18637/jss.v080.i01

    work page doi:10.18637/jss.v080.i01 2017

  22. [22]

    Shen, T., Nath, S.S., Brielmann, A., and Dayan, P. (2024). Simplicity in complexity: explaining visual complexity using deep segmen ­ tation models. ICLR Workshop on Representational Alignment. https://doi.org/10.48550/arXiv.2403.03134

    work page doi:10.48550/arxiv.2403.03134 2024

  23. [23]

    Youngblood, M. (2019). Conformity bias in the cultural transmis ­ sion of music sampling traditions. Royal Society Open Science. https://doi.org/10.1098/rsos.191149

    work page doi:10.1098/rsos.191149 2019

  24. [24]

    Deffner, D., Kandler, A., and Fogarty, L. (2022). Effective popu ­ lation size for culturally evolving traits. PLOS Computational Biology. https://doi.org/10.1371/journal.pcbi.1009430

    work page doi:10.1371/journal.pcbi.1009430 2022

  25. [25]

    Youngblood, M., and Passmore, S. (2026). Simulation ­based in ­ ference with deep learning suggests speed climbers combine innovation and copying to improve performance. Proceedings of the Royal Society B. https://doi.org/10.1098/rspb.2025.1433

    work page doi:10.1098/rspb.2025.1433 2026

  26. [26]

    Kline, M., and Boyd, R. (2010). Population size predicts technol ­ ogy complexity in Oceania. Proceedings of the Royal Society B. https://doi.org/10.1098/rspb.2010.0452

    work page doi:10.1098/rspb.2010.0452 2010

  27. [27]

    Fay, N., De Kleine, N., Walker, B., and Caldwell, C. (2019). Increas­ ing population size can inhibit cumulative cultural evolution. Proceedings of the National Academy of Sciences. https://doi. org/10.1073/pnas.1811413116. 7

    work page doi:10.1073/pnas.1811413116 2019