Ctrl-F-Resist. Practices, Challenges, and Technical Needs of Civil Society Organizations Monitoring the Far-Right Online

AI summary

Overview

Research area: Human-Computer Interaction / CSCW (Computer-Supported Cooperative Work), at the intersection of civil society studies, online monitoring, human-centered AI, and civic technology.

Technical level: Intermediate. The paper is a qualitative interview study rather than a methods-heavy technical paper, but it engages with AI concepts (transcription, OCR, automated classification, semantic search) and describes an open-source monitoring prototype, so some familiarity with social media research and CSCW vocabulary helps.

Scope (one sentence): A qualitative study of 15 practitioners from 12 Germany-based civil society organizations that monitor far-right and antidemocratic content online, examining their workflows, challenges, and technical needs, and translating those into design recommendations and an open-source Telegram monitoring prototype.

What This Paper Is About

Civil society organizations (CSOs) monitor far-right and antidemocratic activity online over the long term, in local and often underreported contexts, but they work under chronic underfunding, legal uncertainty, limited platform access, and little technical capacity. Existing research and tool development has mostly targeted fact-checkers and content moderators, leaving CSOs' specific workflows and constraints underexplored. This paper asks what CSOs actually do when monitoring (RQ1), what technical, ethical, and legal challenges they face (RQ2), and what technological support they need and expect (RQ3).

Key Contributions

1. An in-depth empirical account of CSO monitoring practices under complex organizational and political conditions, based on 15 practitioners from 12 Germany-based organizations — a stakeholder group the authors describe as key yet overlooked in digital governance research.
2. A conceptual monitoring workflow plus its implementation in an open-source Telegram monitoring prototype, built through eight new components designed to flexibly support diverse monitoring goals and to be configurable via lightweight default settings for organizations with limited technical capacity.
3. Concrete design and policy guidance: ten recommendations for the design and development of monitoring tools, three recommendations for policymakers, and six directions for future CSCW research.
4. Two new conceptual contributions: the manual labor trap, an empirically grounded concept explaining why monitoring CSOs remain locked into labor-intensive, low-capacity arrangements, and a proposed CSCW-grounded theory of collaborative sociotechnical scaling to understand collaboration as a pathway beyond fragmented, labor-intensive monitoring work.

Main Findings

• Monitoring is largely manual. Participants describe routinely triggering full-text searches by hand — "trigger the full-text search manually every time" (P4) — using the Telegram desktop client or mobile app, with automation via the Telegram API rarely used. Documentation is done by hand through screenshots and notes, making the process "very time-consuming and ineffective" (P11).

• Low-tech documentation dominates. Screenshots, manual media downloads, and written notes are the norm; scrollshots are used for long threads, producing "a hundred photos" (P6). Screenshots are treated as essential evidence where "the incitement to violence and the account must be identifiable" (P9). Only one participant raises concerns about evidence manipulation, using cryptographic hashing and independent web archiving (e.g., the Internet Archive) to establish a "chain of evidence."

• Weak technical infrastructure. Most storage is local files; few organizations maintain databases (typically manual and text-only); only two participants report using in-house scraping scripts.

• Platform relevance shifts over time. Mainstream platforms (Instagram, TikTok, Facebook, YouTube, X) and Telegram are most relevant, with some monitoring of VKontakte, right-wing forums, imageboards, and media comment sections. Participants note Facebook "in the past" and Telegram "still relevant to a certain extent, but there is not so much going on publicly anymore." Instagram and TikTok are described as particularly important for right-wing extremism among younger audiences, while "stalking takes place on Instagram, and hate campaigns take place on X" (P4). Cross-platform analysis is emphasized.

• Content is increasingly multimodal. Text remains the primary focus as it yields substantial insight while being comparatively efficient and reliable to analyze (P14). Visual content is resource-intensive to store and process but important (sharepics, videos from demonstrations, emojis as coded signals). Voice messages have declined in relevance, while podcasts and long-form livestreams remain valuable for identifying actors and documenting statements.

• Enhanced search capabilities are the most pressing technical need, across a broad range of content types and modalities including audio, images, and video.

• Openness to AI is narrow and task-specific. AI-based automation is seen as potentially useful for media processing such as transcription or optical character recognition (OCR). Participants express substantial skepticism toward automated content classification.

• Trust is central to AI skepticism. AI is often viewed not only as unreliable but as a potential threat to legal validity, professional credibility, and the value of human expertise.

• Ethical and regulatory challenges hinder collaboration across organizations — both in data sharing and in coordination of content-related efforts — which in turn limits resource-efficient, collaborative technical innovation.

• Participants' organizational profiles (n=15): organizational focus included far-right (8), conspiracy theories (3), various phenomena (3), antisemitism (2), racism (1), and human rights violations (1); main activity included monitoring (7), consulting (5), archiving (2), analysis (2), education (1), and software development (1); 3 had technical education and 12 did not.

Methodology in Plain English

The study was part of a research-to-practice project with a partner CSO, run in parallel with the development of an open-source prototype whose technical focus is Telegram.

• Participants: 15 practitioners from 12 Germany-based CSOs, recruited through non-probability sampling that combined purposive (expert-based) and snowball methods, using contacts from civil society including the project's partner CSO. No financial incentives were offered. Organizations' activities spanned archiving, monitoring, and counseling/advisory work.
• Interviews: 13 remote video-conference interviews conducted between October 2024 and January 2025, lasting 30 minutes to one hour, using a semi-structured guide. Two interviews had two participants from the same organization. Interviews were audio-recorded when possible; in two cases where participants declined recording, a second researcher attended and wrote a detailed protocol. Interviewers sometimes introduced examples such as topic modeling, clustering, network analysis, or video summarization to spark discussion with less technical participants.
• Ethics and anonymization: The researchers drew on Internet Research Ethics guidelines and research on online extremism. Informed consent was obtained, audio-recording consent was separate and not required, and the authors deliberately refrained from describing participants or their organizations in detail because even indirect identifiers (region, monitored groups, founding date) could reveal identities. Transcripts were auto-generated locally with Whisper AI (2022), manually proofread, then audio and video files were securely deleted; identifiable details were removed and anonymized transcripts encrypted with password protection. Participants were not financially compensated.
• Researcher positionality: Access was mediated through the partner CSO, which fostered trust. Participants knew the researchers were involved in developing AI-based tools for automated hateful-content detection, which may have shaped framing — though participants frequently voiced critical, skeptical views of automated detection.
• Analysis: Thematic Analysis (Braun and Clarke, 2006), with data-driven inductive analysis. Familiarization was done by the first author; open coding stayed close to participants' wording. Two interviews were independently coded by both authors to align interpretations. The preliminary thematic map comprised 714 codes in 25 themes, revised to 21 themes and 497 codes, and the analysis ultimately consolidated into four meta-themes, 21 themes, 80 subthemes, and 329 codes translated into English. Themes were mapped onto research questions only after the thematic structure was established. A full code list is available via Zenodo.
• Workshops: Two one-day participatory workshops combined presentation of interview findings and the emerging workflow, hands-on use of an early prototype centered on participants' own thematic interests, and moderated group discussions. Participants submitted relevant Telegram channels in advance for pre-integration. In hands-on sessions they registered accounts, tested lexical and semantic search, inspected automatically generated audio transcripts, and explored a "conspiracy score" on message level from an automated classifier.

The four meta-themes structuring the findings: Current Practices and Workflows (4.1), Technical, Ethical and Legal Challenges (4.2), Needs and Expectations Regarding Technological Support (4.3), and Perspectives on the Role of AI and Human-AI Collaboration (4.4).

Why This Matters

Impact on research. The paper foregrounds civil society actors as stakeholders in the governance of digital spaces, complementing CSCW work on invisible work and articulation work, HCI research on human-centered AI, and sociotechnical infrastructuring. It argues that existing tools and system designs are often built on an incomplete account of early-stage monitoring activity, which is far less examined in research than later stages such as claim verification or dissemination of insights.

Real-world applications:

• Civil society organizations monitoring far-right activity can use the conceptual workflow and the open-source Telegram prototype components as a starting point for their own tooling.
• Tool developers and civic-tech groups can apply the ten design recommendations to build monitoring software that fits resource-constrained, legally cautious, and ethically sensitive settings.
• Policymakers can act on the three policy recommendations to address the funding, platform access, and regulatory conditions that keep CSOs locked into manual work.
• Funders and networks can use the "manual labor trap" concept to reason about how funding structures and competition among CSOs inhibit shared infrastructure.

Industry relevance. Platform companies, analytics vendors, and AI developers get a clear signal that automation demand is not uniform: users in this domain want assistive media processing (transcription, OCR) and better search far more than automated content classification, where distrust about reliability, legal usability, and professional credibility is high. It also highlights the structural vulnerability of platform-dependent data access, given the termination of affordable platform access and discontinuation of social media analytics tools.

Future Directions

1. Develop the proposed CSCW-grounded theory of collaborative sociotechnical scaling to understand both the promise and the challenges of inter-organizational collaboration as a path beyond fragmented, labor-intensive monitoring.
2. Pursue the six directions the paper identifies for future CSCW research on deepening engagement with civil society needs, responsible technology integration, and more inclusive digital governance.
3. Extend the prototype and workflow beyond Telegram to other platforms, services, and communication modalities — the authors state this broader transferability is an explicit aim of the study.
4. Bridge the gap between automated classification research and CSO practice: how to make AI support trustworthy, legally usable, and credible to practitioners, given that general-purpose hate speech classifiers are unlikely to capture locally embedded and evolving phenomena.

Target Audience

CSCW and HCI researchers studying civil society, online harms, and human-centered AI; designers and developers of monitoring, OSINT, and content-analysis tools; civil society organizations and their funders; and policymakers concerned with platform governance, digital regulation, and support for organizations defending democracy online. Journalists and fact-checkers working on far-right online activity will also find the workflow findings directly applicable.

Authors’ abstract

As far-right actors increasingly exploit online platforms to disseminate ideology and mobilize supporters, civil society organizations (CSOs) play a vital yet underrecognized role in monitoring antidemocratic dynamics online. Unlike fact-checkers or content moderators, CSOs engage in long-term, contextualized analysis, often in resource-constrained settings and under precarious conditions. Despite their critical societal role, CSOs face significant barriers to adopting or co-developing technical solutions, including legal uncertainty, limited platform access, and chronic underfunding. Existing research and tool development efforts have largely overlooked these actors in favor of more institutionally embedded stakeholders. This paper addresses this gap through a qualitative study with 15 practitioners from 12 Germany-based CSOs engaged in online monitoring, positioning them as key yet overlooked stakeholders in the governance of digital spaces. We explore their current practices, challenges, and expectations regarding technological support. Our findings show that monitoring remains largely manual due to the lack of tailored tools, with enhanced search capabilities emerging as the most pressing technical need. While participants express openness to AI-supported features such as media processing and content discovery, many remain skeptical of automated classification, citing concerns around trust, legal usability, and professional credibility. Grounded in these findings, we introduce a conceptual monitoring workflow and describe its implementation in an open-source Telegram monitoring prototype designed to flexibly support diverse monitoring goals. We outline concrete design, policy, and research recommendatios, and introduce the manual labor trap as an empirically grounded concept that explains why monitoring CSOs tend to remain locked into labor-intensive, low-capacity arrangements.