- The paper finds through qualitative interviews with 19 U.S.-based text-to-image model creators that platform policies alone cannot support responsible development, because creators rely primarily on community norms and informal practices.
- Creators repurpose attribution and Responsible AI tools for self-protection, branding, liability management, and access control, revealing a gap between tools designed for transparency and how marketplaces actually use them.
- The study recommends concrete compliance guidance, broader harm taxonomies, supply-chain awareness tools, and community-informed oversight while warning that local norms may overlook wider societal harms and shift accountability to volunteers.
This paper presents a qualitative interview study of 19 individual creators of text-to-image (T2I) derivative models on open AI model marketplaces such as Hugging Face and CivitAI, examining how platform policies, Responsible AI (RAI) tools, and community norms shape their responsible development practices. The central argument is that formal platform governance alone is insufficient: creators' sense of responsibility is negotiated primarily within communities, and RAI tools are repurposed in ways that diverge from their transparency-oriented design intent.
Study design and method
The authors conducted semi-structured interviews (75–90 minutes each) with 19 U.S.-based individual model creators recruited from marketplaces, Discord communities, nomination lists, and snowball sampling. All participants built derivative models by fine-tuning pre-trained models; ten used node-based interfaces (ComfyUI), the rest code-based environments. Seventeen held other full-time jobs and created models as a hobby; two worked as freelancers. The protocol was developed iteratively across two phases: Phase 1 established baseline practices and policy perceptions, while Phase 2 added an "ideal platform design" session in which participants acted as policymakers choosing between governance options, plus design-critique probes of two tool concepts—a model attribution tool adapted from He et al.'s AI attribution toolkit and an attribution indicator drawing on the Content Authenticity Initiative's content credentials marker. Analysis followed reflexive thematic analysis with independent coding and consensus-based theme refinement.
The study's scope is deliberately narrow: it covers only image-generation creators, all U.S.-based, with a heavily male participant pool. These constraints bear directly on how far the findings generalize, as discussed below.
Three regulatory needs and compliance challenges
The first major finding identifies three regulatory needs arising from creators' workflows:
Reducing harmful depictions. Twelve of 19 participants endorsed strict platform prohibition of models capable of generating child sexual abuse material or realistic depictions of identifiable individuals, arguing that moderation ultimately broadens the user base by diversifying permissible creative concepts. Notably, harms propagate through the remixing chain itself—participants reported encountering harmful content embedded in upstream models they remixed, illustrating networked harassment circulating without creators' awareness. However, seven participants resisted strict regulation because they cannot verify that their models are safe: responsibility attribution across pre-trained bases and downstream users is ambiguous, models become "out of control" after release, and current safety practices are ad hoc heuristics rather than structured workflows. Participants also criticized vague policy language and asked for uniform, concrete guidelines applicable regardless of creative concept. One participant attributed continued production of offensive models to audience demand ("ethics are boring"), highlighting that incentive structures—not just rules—drive harmful output.
Incentivizing originality. Creators located originality in dataset curation choices, fine-tuning technique, and outputs unattainable by upstream models (e.g., culturally specific clothing absent from Stable Diffusion XL). Yet they frequently scraped unlicensed artwork, rationalizing this as transformative "fair use"—a conflation between folk notions of fair use and legal transformativeness that the authors flag explicitly. Nine participants worried about idea appropriation by other creators diverting audience attention. Critically, participants held divergent, unresolved definitions of what constitutes model originality, which undermines any straightforward platform mechanism for rewarding it.
Securing legal ownership. Creators wanted legally enforceable ownership to support future monetization, but expressed uncertainty about whether ownership claims survive given tangled upstream dependencies—models are typically combinations of multiple fine-tuned checkpoints, making full provenance tracking impractical. Some attempted to minimize dependencies entirely (e.g., collecting training images in person) to strengthen future copyright claims. Others warned that enforced crediting requirements could suppress open sharing and harm newcomers, echoing remix-culture tensions.
A second finding concerns attribution tools. Rather than using them to signal transparency to downstream actors, eight participants viewed the proposed attribution tool as a provenance-management intervention for self-protection: controlling downstream use, clarifying liability boundaries, and enabling access control after deployment. Participants also demanded traceability spanning platforms, governments, and AI corporations, proposing unified cross-platform licenses and embedded image metadata registries—motivated by the observation that no strong U.S. legal case has yet been brought by a creator side against misuse.
Views on visible attribution indicators split along a single underlying consideration: audience reception. Some saw watermarks as branding and discoverability assets; others feared privacy exposure through revealed provenance and wanted user-side control over watermark visibility. The implication is that attribution infrastructure designed for accountability can be co-opted into self-branding, potentially allowing creators to selectively omit or manipulate upstream information while distancing themselves from accountability—the authors connect this to ethics-washing risks documented in prior RAI literature.
Seventeen of 19 participants relied substantially on communities (14 primarily private Discord servers). Communities served three functions: iterative feedback for achieving high-originality outputs and gauging marketability; the source of responsibility norms—which prioritized crediting peers and preventing harm to members while largely ignoring explicit-harm prohibitions unless members were directly affected; and trust-gated distribution channels where models were soft-released before public deployment. Trust within small communities reduced the perceived burden of documentation and disclaimers that public marketplaces demand. Conversely, participants expressed distrust toward CivitAI specifically, citing its NSFW ecosystem and inadequate moderation of non-consensual content.
The implication is significant: community norms, not platform policy, are the effective regulatory force for most creators—but those norms are parochial, oriented toward member-protective harms rather than broader societal ones, and may not align with legal standards. Several participants also cautioned that heavy reliance on community feedback could dilute individual accountability, with creators positioning themselves as "learners" deferring to possibly uncredible experts.
Policy implications
The authors derive four recommendations: providing actionable guidance with concrete bad-case examples and audit checklists rather than vague prohibitions; expanding harm taxonomies beyond explicit/illegal content to representational harms shaped by skewed datasets and creator demographics; structuring multi-stakeholder deliberation into policymaking; and building supply-chain awareness tools that help creators understand their position relative to upstream data owners and downstream users. They also propose integrating communities into oversight—community-informed guideline drafting and distributed model auditing—while candidly noting the trade-offs: moderator burnout, liability-shifting by platforms onto volunteers, biased norms, and uncertain long-term engagement.
Limitations
The paper concedes several constraints plainly. Findings cover only T2I (and one I2I) creators; language-model creators may face different dynamics, particularly around originality. Participants may not represent all individual open-model builders—for example, those training from scratch—and the study did not distinguish which types of pre-trained bases participants used. Gender imbalance among participants limits generalization, though the authors note tentative evidence of distinct needs among female participants in male-dominated spaces. Finally, all participants were U.S.-based, so perspectives shaped by the EU AI Act or the UK Online Safety Act (which prompted CivitAI's UK withdrawal) are not captured.
Conclusion
Through interviews with 19 individual T2I model creators, this study shows that marketplace policies and RAI tools address creator realities only partially: creators need mechanisms for mitigating downstream harms, recognizing originality, and securing ownership; they repurpose attribution tools for self-protection and visibility; and their responsibility norms emerge from communities rather than formal governance. The paper's contribution is to reframe platform governance as requiring integration of policy, tooling, and community dynamics—an argument grounded in creators' own accounts, though bounded by its single-modality, single-country scope.