Govern the Model, Not Only the Data: Storage, Circulation, and Learning in Creative AI
Phoenix Perry, George Simms, Elizabeth Wilson, Yasmine Boudiaf, Nick Bryan-Kinns, Tega Brain, R. Luke DuBois, Alix Rule, et al.
cs.AI · cs.HC
Quick verdict
A position paper arguing federated learning alone doesn't fix extractive AI: governance must extend to models. Compelling framework, thin evidence.
Summary
This position paper argues that federated learning, though framed as privacy-preserving, is not by itself a remedy for extractive AI: it distributes computation while the resulting model stays with whoever convened training. The authors distinguish this from the federated social web (e.g., ActivityPub), which distributes publication together with governance, showing how the shared vocabulary of 'federation' lets federated learning inherit unearned political standing. They propose three governance layers for creative communities — storage, circulation, and learning — and argue, through cases such as the TRANSFER Data Trust, Serpentine's Choral Data Trust, Creative Commons' CC Signals, and Local Contexts' Indigenous data labels, that artist governance is established at storage and circulation but stops at learning: contributors can consent to training yet have no say over the resulting model. They derive four design principles for a creative data commons (govern the model, not only the corpus; make terms legible at the moment of contribution; treat refusal as a first-class state; decide stewardship in the open and account for it) and map open problems: small-client aggregation, provenance under model merging, post-training withdrawal, and feasibility without institutional compute. Limitations: the argument is entirely conceptual, with no interviews or systematic case survey; the central 'governance stops at learning' claim rests on a small illustrative case set; the principles are not yet operationalized; some load-bearing citations are blogs or vendor pages.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Federated learning borrows the vocabulary of the federated social web while inverting its logic: it distributes computation, and the resulting model stays with whoever convened the training.
- Federation is a protocol, not a politics — whether it builds or erodes a creative community's agency depends entirely on who governs the data and the resulting model.
- Across the artist-governed initiatives examined (TRANSFER Data Trust, Serpentine's Choral Data Trust, consent infrastructures), governance is established at storage and circulation but stops at learning: contributors can consent to training yet have no say over the resulting model.
- Four design principles — govern the model, not only the corpus; legible terms at the moment of contribution; refusal as a first-class state; stewardship decided in the open and accounted for — can guide a creative data commons that governs models and their federation, not only datasets.
Strengths
Sharp disambiguation of two conflated 'federations'
Sections 1-2 cleanly separate the FLOSS federated web (ActivityPub, where publication and governance are both distributed) from federated learning (McMahan et al. [2], Kairouz et al. [6]), where computation is distributed but the aggregated model and ownership stay with the convening organization, illustrated by Gboard [7,8] and Apple's differential privacy [9,10]. This distinction has real analytic and policy value.
Usable three-layer governance vocabulary
Section 4 and Figure 1 define storage (who holds files and access), circulation (attribution, licensing, payment, exit), and learning (who may train and who holds the model), and — importantly — separate consent-to-training from governance of the resulting model, a distinction GDPR-era debates routinely blur.
Grounded in real, checkable governance experiments
Research agenda pairs technical and human questions
Honest treatment of power and labour asymmetries
Weaknesses
Entirely conceptual; central claim outruns evidence
The decisive case is asserted, not shown
Attribution conflation at a load-bearing example
Principles are directional but unoperationalized
Weak sourcing for technical privacy claims
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a position paper (9 pages, 45 references; funded by AHRC BRAID, grant UKRI3355) at the intersection of HCI, creative practice, and AI governance. Its three stated contributions (Section 1) are: (1) disentangling two conflated meanings of 'federation' — the FLOSS federated web (ActivityPub [3]), which distributes publication together with governance, and federated learning (McMahan et al. [2]; Kairouz et al. [6]), which distributes computation while the aggregated model and its ownership stay with the convening organization, with Google's Gboard [7,8] and Apple's differential-privacy infrastructure [9,10] as the emblematic cases; (2) a three-layer governance framework — storage, circulation, learning (Section 4, Figure 1) — carrying the central claim that artist-governed initiatives have established governance at the first two layers but not the third: consent to training exists (Serpentine's Choral Data Trust, 15 UK choirs [27,28]), governance of the trained model does not; (3) four design principles for a creative data commons (Section 7): govern the model, not only the corpus; legible terms at contribution; refusal as a first-class state; open, accountable stewardship.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What are the three layers of data governance in creative AI according to this paper?
The paper proposes storage (who holds the files and who can reach them), circulation (attribution, licensing, payment, and the right to withdraw), and learning (who may train a model and who holds the resulting model). It argues artist communities have built governance for the first two layers — e.g., the TRANSFER Data Trust — but not the third: contributors can consent to training yet have no say over the trained model.
Is federated learning actually privacy-preserving for artists?
Only partially, per this paper. Personal data do stay on-device, which aligns with GDPR-style individual rights. But the authors argue privacy is not agency: the aggregated model ends up owned by whoever convened the training, so contributors gain no control over what was made from their work, and the term 'federation' misleadingly imports the governance reputation of the federated social web.
What is the TRANSFER Data Trust?
A member-owned cooperative formed in 2023 when the TRANSFER gallery dissolved and returned its assets to its artists. Each studio keeps its own archive, linked across peers for redundancy (federated rather than pooled storage); a commission on sales is redistributed annually, and stewardship is settled by quarterly member votes.
What are the paper's four design principles for a creative data commons?
(1) Govern the model, not only the corpus; (2) make the terms legible at the moment of contribution; (3) design for refusal as a first-class state; (4) decide stewardship in the open and account for it. Each is addressed to specific constituencies: systems builders, interface designers, creators, communities, funders, and policy.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1's labels and caption are interleaved into the Section 4.1 body text by PDF extraction (visual layout lost); all of its textual content appears present, so no figure content is treated as missing in this review.
- The provided metadata lists 8 authors, but the paper's byline lists 13 (adding Rachel Meade Smith, Kelani Nichole, Atharva Pravin Pawar, and Rebecca Fiebrink) — a metadata mismatch, not a text gap.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.