PaperlayerView exact revision on arXiv

Source-linked AI summary

"Okay, I've Actually Softened My Take on This": How People in Decentralized Social Media Reason about the Appropriateness of Generative AI

Romina Mahinpei, Manoel Horta Ribeiro, Andrés Monroy-Hernández, Sohyeon Hwang

arXiv:2609.18709v1cs.HC

TL;DR

Broad pro- or anti-AI labels offer little insight into what people find appropriate in DSM and why, where decisions are distributed across ecosystem actors. Through interviews structured around seven GenAI scenarios, the paper finds that participants draw conditional boundaries based on technology, integration, and use, with boundaries varying across people and scenarios.

  • Problem

    Broad pro- or anti-AI characterizations provide little insight into what people consider appropriate in DSM and why, despite the conditional nature of technology judgments.

    Source: §1 · pp. 1–2

  • Method

    The study used semi-structured interviews with people from Mastodon and Bluesky, discussing seven scenarios involving GenAI in DSM.

    Source: §1 · p. 2, §3.2 · p. 7

  • Results

    Participants drew conditional boundaries around GenAI configurations through considerations about technology, integration, and use, and those boundaries varied across and within participants.

    Source: §1 · p. 2, §5 · p. 15

  • Takeaways & Limitations

    Making boundaries explicit may support more grounded design, policy, and collective deliberation about GenAI governance in DSM.

    Source: §7 · p. 19

    The findings capture scenario-based interview boundaries, and the scenarios primarily varied GenAI use, which may have contributed to the relative prevalence of use-related considerations.

    Source: §6 · p. 19

Abstract

from arXiv · show

Generative AI (GenAI) is increasingly integrated into social media, raising questions about whether, where, and how it belongs. In decentralized social media (DSM), these decisions are distributed across users, developers, moderators, and administrators, making GenAI a collective governance challenge. At the same time, public discourse often flattens arguments to broad pro- or anti-AI positions that offer little insight into what people actually find (in)appropriate and why. Through 20 semi-structured interviews with people from Mastodon and Bluesky, structured around seven GenAI scenarios, we examine how people reason about GenAI's appropriateness in DSM. We find that participants drew conditional boundaries around particular GenAI configurations through distinct, salient, and weighted considerations spanning technology, integration, and use. We conceptualize this as boundary drawing and show how making such boundaries visible can support more grounded design, policy, and collective deliberation around GenAI in DSM.

1 Introduction

As GenAI enters decentralized social media (DSM), distributed decision-making makes it important to understand when particular applications belong, beyond broad pro- or anti-AI labels. Interviews show that people draw conditional boundaries using considerations about the technology, its integration, and its use.

1 Introduction

  • DSM distributes decisions about GenAI across users, developers, moderators, and administrators, who may hold different interests, values, and responsibilities.Unlike centralized platforms, DSM does not leave these decisions solely to the companies running the platforms.Source: §1 · p. 1
  • Public discussion often reduces reactions to pro- or anti-AI positions, although judgments about technology are often situated and conditional.Understanding how judgments vary across configurations can inform governance and decision-making in DSM.Source: §1 · p. 2
  • 20 interviews found that participants drew conditional boundaries around GenAI configurations using considerations about technology, integration, and use.The boundaries varied between participants and within the same participant across scenarios.Source: §1 · p. 2
  • Participants’ considerations varied in salience and weight across people and across scenarios for the same person.Different combinations of considerations produced boundaries that varied within and between participants.Source: §1 · p. 2
  • Making heterogeneous and partly articulated boundaries explicit may support collective deliberation about GenAI governance in DSM.The authors identify boundary articulation as a useful step toward that deliberation.Source: §1 · p. 2

2 Background

The background situates GenAI governance across social media, decentralized governance, and technology appropriateness. It identifies a gap in understanding how conditional judgments are structured and negotiated in DSM.

2 Background

  • The paper builds on research about GenAI in social media, DSM governance, and technology appropriateness to examine how heterogeneous judgments are structured.These three strands motivate the study.Source: §2 · p. 3

2.1 Emergence of Generative AI in Social Media

  • GenAI can support content creation, curation, and moderation, while raising concerns about privacy, authenticity, manipulation, bias, and synthetic-content provenance.Social-media governance has visibly addressed AI-generated content, but appropriateness questions for applications such as summarizing or filtering remain less investigated.Source: §2.1 · p. 3

2.2 Governance in Decentralized Social Media

  • Mastodon and Bluesky distribute control across actors, creating opportunities to shape spaces around local needs and values.Prior work examined particular governance activities, while this paper studies how people across DSM roles reason about a broader range of GenAI applications and construct boundaries.Source: §2.2 · p. 4

2.3 Appropriateness of Technology and Generative AI

  • Prior research treats technology judgments as context-dependent, but offers less insight into how considerations shift and combine to produce different boundaries around the same application.The paper examines which considerations people draw on, how their salience changes across configurations, and how they shape judgments in DSM.Source: §2.3 · pp. 4–5

3 Methods

The study used semi-structured interviews with 20 people from Mastodon and Bluesky to examine reasoning about GenAI appropriateness. Researchers presented seven scenarios and analyzed the interviews thematically.

3 Methods

  • The researchers conducted 20 semi-structured interviews with Mastodon and Bluesky participants, each lasting 45–60 minutes over Zoom.Participants provided informed consent, and the study received Institutional Review Board approval.Source: §3 · p. 5

3.1 Participant Recruitment

  • Recruitment sought people familiar with DSM across platform and ecosystem roles, without aiming to compare role groups or platforms systematically.The pool included Mastodon, Bluesky, and dual-platform users, with varied technical experience and involvement.Source: §3.1 · pp. 5–6, Table 1 · p. 6

3.2 Interview Protocol

  • The interview protocol presented seven scenarios spanning GenAI-generated or edited posts, feed summaries and curation, and eliciting feed preferences.Participants discussed their reactions and reasoning and what changes might alter their views, then reflected on GenAI’s future role in DSM.Source: §3.2 · pp. 6–7

3.3 Thematic Analysis

  • Thematic analysis began with inductive open coding, developed a shared codebook, and synthesized related codes into higher-level themes and patterns.The first and last authors examined relationships across participants and scenarios, with themes iteratively revised by the research team.Source: §3.3 · p. 7, Figure 1 · p. 7, Figure 2 · p. 8

4 Findings

Participants judged GenAI through conditional boundaries, weighing technology, integration, and use differently across people and scenarios. Their judgments also reflected values of decentralization, agency, and authentic interaction, but those values did not produce a shared position on where GenAI belonged.

4 Findings

  • Participants’ boundaries were conditional and multifaceted: they weighed considerations across technology, integration, and use, with different considerations carrying different weight across people and scenarios.They did not necessarily invoke every dimension, and recognizing a concern did not mean it determined the boundary.Source: §4 · p. 8, §4.4 · pp. 13–14

4.1.1 Participants considered the broader societal costs and impacts of GenAI. Participants frequently refer-

  • Societal costs—including environmental and resource impacts, labor, copyright, and unequal distribution of benefits—could matter, but changing development conditions or the relative weight of other concerns could shift a boundary.Participants cited local models, reduced environmental consumption, and consent or compensation for training data as potentially relevant conditions.Source: §4.1.1 · p. 9

4.1 Technology: What GenAI Is and What It Entails

  • Some participants viewed centralized GenAI infrastructure as reproducing concentrated power and privacy concerns, while favoring local, user-owned, or community-controlled alternatives.For local models, participants valued constraints on where models operated and who could access data.Source: §4.1.2 · p. 9, building centralized social media.” · pp. 9–10
  • Reliability and bias concerns mattered differently by task: participants were more cautious about truthful content or image generation than short summaries or edits they could inspect.Participants cited hallucinations, biased outputs, and difficulty recognizing bias in opaque systems.Source: §4.1.3 · p. 10, intentionally or inadvertently… · p. 10

4.2.1 Participants considered whether GenAI aligned with community norms and values. Participants

  • Participants assessed integration against community needs and values, finding assistance more aligned when it supported accessibility or moderation without displacing valued human participation.Novelty or technical capability alone did not establish that a feature belonged in a shared space.Source: §4.2.1 · pp. 10–11

4.2 Integration: How GenAI Is Introduced into Shared Spaces

  • Control and transparency shaped integration boundaries: participants valued opting out and retaining agency, while expecting disclosure for generated content and greater visibility into systems shaping feeds or experiences.The needed transparency varied with the feature, from labeling generated content to explaining how recommendations shaped what users encountered.Source: §4.2.2 · p. 11, §4.2.3 · pp. 11–12

4.3 Use: What GenAI Is Used For and How

  • Participants distinguished uses by how far GenAI transformed their input and whether people retained a meaningful role; transformation could raise concerns about voice while enabling expression otherwise withheld.Shortening or grammar correction was seen as less likely to depart from intent than generating a post or image, but assistance and user control could make more transformative uses acceptable.Source: §4.3 · p. 12, §4.3.1 · p. 12, §4.3.2 · p. 12

4.3.3 Participants considered whether GenAI was suitable and added meaningful value. GenAI’s ability

  • Whether GenAI added suitable value depended on the task: participants questioned uses with simpler alternatives, but saw value in accessibility support and moderation assistance that kept a human involved.Examples included translation, alt-text, and helping moderators sort or flag content for human review.Source: §4.3.3 · p. 13

5 Discussion

The discussion frames GenAI appropriateness as boundary drawing: conditions people use to distinguish configurations across technology, integration, and use, rather than fixed pro- or anti-AI positions. It argues that making those conditions visible can inform individual reflection and collective deliberation, while noting that the findings are grounded in Mastodon and Bluesky.

5.1 Understanding and Responding to GenAI Appropriateness Through Boundaries

  • Boundary drawing captures conditions delimiting which GenAI configurations people consider appropriate across technology, integration, and use, which binary labels can obscure.Similar concerns may lead to different boundaries, and similar outward reactions may reflect different underlying conditions.Source: §5.1.1 · p. 15, §5 · p. 15, §5.1.3 · p. 16

5.1.2 Characterizing boundaries across applications. Our technology, integration, and use dimensions provide

  • The dimensions make heterogeneous judgments more comparable, but support for one GenAI configuration should not be treated as support for introducing it elsewhere in the same ecosystem.People may accept a technology yet object to a use, or support a use only under specific integration conditions.Source: §5.1.2 · p. 16

5.2 Supporting Individual Boundary Articulation

  • Reasoning through concrete scenarios can help people articulate or reconsider boundaries, so elicitation may require comparing configurations and reflecting on which conditions matter.Participatory activities could vary technical properties, integration, and use, using structured prompts such as proposed boundary cards.Source: §5.2.1 · pp. 16–17, §5.2.2 · p. 17

5.3 From Individual Boundaries to Collective Deliberation

  • Collective deliberation can make disagreements more concrete by eliciting conditions across the three dimensions and identifying points of convergence and disagreement.The discussion describes adapting structured deliberation approaches to compare members’ reasoning about specific configurations.Source: §5.3.1 · p. 17
  • Deliberation need not produce one shared boundary: communities may use individual controls, shared minimum conditions, or prohibitions, while surfacing boundaries alone does not resolve disagreement or guarantee consensus.Different conditions may call for action by model providers, application developers, administrators, or communities, and those mappings can overlap.Source: §5.3.2 · pp. 17–18, §5.3.3 · p. 18

5.4 Boundary Drawing Beyond Decentralized Social Media

  • Because the findings come from Mastodon and Bluesky, their salient concerns reflect a setting where decentralization, autonomy, and agency matter; the study does not establish that the same dimensions structure judgments elsewhere.The authors present boundary drawing beyond DSM as a direction for future work, not as an established generalization.Source: §5.4 · p. 18

6 Limitations

The findings are constrained by the small, qualitative sample and by scenario-based interviews that elicited rather than observed boundaries. Because both GenAI and DSM are evolving, the considerations people weigh may change over time.

6 Limitations

  • The 20 participants represent a relatively small subset of Mastodon and Bluesky users, so the findings are not a definitive overview of what matters across DSM.The analysis is interpretive and shaped by participants’ subjectivity and self-presentation.Source: §6 · p. 18
  • Changes in GenAI and DSM may shift which considerations matter and where people draw boundaries, motivating longitudinal study of boundary articulation as an ongoing governance process.The passages do not establish how quickly or in what direction those shifts will occur.Source: §6 · pp. 18–19
  • Scenario-based interviews elicited boundaries rather than capturing those established in practice, and the scenarios may have shaped which considerations surfaced.They varied uses more than technology or integration, potentially contributing to the relative prevalence of use-related considerations; frequency should not be read as relative importance.Source: §6 · p. 19

7 Conclusion

The conclusion frames GenAI appropriateness in DSM as conditional boundary drawing, not a simple pro- or anti-AI stance. The disclosure describes the authors’ use of GenAI tools in research and manuscript preparation.

7 Conclusion

  • Participants’ boundaries reflected considerations about technology, integration, and use, varying across people and within the same person across scenarios.Making boundaries explicit may support deliberation across agreement and disagreement in DSM’s distributed governance.Source: §7 · p. 19

Generative AI Disclosure

  • The authors used Claude, ChatGPT, and Elicit to find related research, ChatGPT to refine writing, and Claude for formatting tasks.They manually verified and edited the refined text.Source: Generative AI Disclosure · p. 19

A Interview Protocol

The appendix introduces the semi-structured interview protocol and its opening questions about participants’ backgrounds and experiences with decentralized social media and GenAI.

A Interview Protocol

– How have you liked it so far?

The protocol asks about participants’ prior GenAI experiences, work on DSM projects, and understandings of AI before presenting scenarios. It also invites broader reflections on GenAI’s future in decentralized social media.

– How have you liked it so far?

– If not, why not?

– If so, how was GenAI used?

– Why was GenAI used / not used?

  • The protocol presented seven GenAI scenarios and asked participants about their reactions and the factors shaping them.For each scenario, it also asked about personal use, others’ use, prior experience, and what might change their view.Source: – Why was GenAI used / not used? · pp. 23–24
  • Participants were asked to reflect on GenAI’s future role in DSM, possibilities they welcomed or wanted avoided, advice for ecosystem actors, and whether their views had changed.They could also raise anything they felt had not been discussed.Source: – Why was GenAI used / not used? · p. 24

B Codebook

Table 3 records the shared codebook used in thematic analysis, distinguishing codes revised after initial development from codes added later. Revisions broadened three existing codes, while two additions captured concerns about opacity and bias and expectations that features prioritize users’ and communities’ interests.

B Codebook

  • Two later-added codes captured concerns about GenAI opacity and bias and expectations that features keep users’ and communities’ interests at the forefront.Source: B Codebook · p. 24
  • Three existing codes were broadened to cover user ownership of outputs, task appropriateness and effectiveness, and pre- and post-deployment evaluation.Source: B Codebook · p. 24
  • Table 3 marks revised codes in gray and later additions in gray with asterisks.Source: Table 3 · p. 25

C Annotated Excerpts for the Image-Generation Scenario

The annotated excerpts show how participants reasoned about image generation through technology, integration, and use considerations. P4 weighs alternative image sources and disclosure, while the excerpt annotations identify the dimensions and considerations in participants’ reasoning.

C Annotated Excerpts for the Image-Generation Scenario

Transparency around GenAI].”

Participants connected image-generation judgments to human involvement, resource use, transparency, and training-data practices. P9 distinguishes augmentation from replacement, while P12 describes concerns about visual artists and avoids using GenAI for that purpose.

Transparency around GenAI].”

Related papers

Loading 2609.18709v1…