Abstract
Black Twitter is an informal online network of Black users who leverage Twitter to share perspectives, build community, and mobilize around cultural and social justice issues. Black Twitter is a unique socio-cultural space where collective identity, shared experience, and cultural production converge. In this empirical study, we discuss Black Twitter as a virtual community, an online field site shaped by social interaction and platform dynamics, drawing on digital ethnographic methods like participant observation to understand its dynamics and the socio-technical systems that shape it. While existing frameworks and checklists fall short in providing methods for specifically studying virtual communities, we introduce an ethics-centered checklist for studying Black Twitter and more generally, marginalized virtual communities by addressing risks such as data misuse, data ownership, and misrepresentation. Applying this checklist, we conduct case studies to: 1) analyze how automatic moderation systems impact Black Twitter users’ experiences and 2) explore in-group agreement on ownership and usage of reclaimed language. We find that moderation systems often misinterpret culturally-specific language and norms around reclaimed terms. Ultimately, our checklist offers an actionable framework for researchers to ethically engage with marginalized virtual communities emphasizing cultural nuance, accountability, and self-awareness.
Introduction
In Machine Learning (ML) and AI research, marginalized digital communities are frequently treated as "mere data sources" rather than living, breathing communities. This leads to "extractive" research practices where data is harvested without regard for the potential harms of misrepresentation, generalization, or the exposure of cultural secrets. A specific pain point is automated content moderation. Standard toxicity classifiers often flag culturally specific language—such as African American Vernacular English (AAVE) or reclaimed slurs—as "hateful," even when used in a communal, positive, or resistant context. This forces users to either accept suppression or perform "linguistic gymnastics" to stay visible. The authors argue that existing ethical frameworks are too broad and fail to provide actionable guidance for navigating these socio-technical complexities.
Methodology
The authors employ a mixed-methods approach rooted in digital ethnography. First, they propose a 7-point actionable framework for researchers, covering aspects like positionality, data anonymization, and community sovereignty. Second, they conducted a qualitative survey ($N=12$) exploring in-group vs. out-group views on the ownership and usage of reclaimed language. The demographic breakdown of this participant group is provided below.
Finally, the researchers performed a quantitative toxicity audit. They collected 1,695 tweets containing both censored (e.g., using asterisks like n*gga) and uncensored versions of a reclaimed slur. These were processed through Detoxify, a popular open-source toxicity classifier, to measure how "identity attack" scores changed based on linguistic perturbations.
Results: Toxicity and Self-Censorship
The quantitative analysis found that character-level perturbations (self-censorship) reduced the "identity attack" score by an average of 30.7%. Uncensored reclaimed terms received an average toxicity score of 0.462, while their censored counterparts dropped to 0.204, even though the underlying sentiment and context remained identical.
As shown in the distribution above, the model prioritizes surface-level keywords over contextual understanding. This forces users into "mimetic language play"—intentionally perturbing text to circumvent flawed algorithms—as a necessary strategy to preserve in-group identity while avoiding shadow-banning or suspension.
Community Perspectives and Ethical Implications
The survey revealed a "drastic variance" in how members define the community. Some link reclaimed language to the broader African Diaspora, while others feel it belongs strictly to the descendants of enslaved Black Americans. This highlights the danger of researchers making "majority-vote" assumptions in data aggregation. Furthermore, 80% of survey respondents stated that out-group individuals should never use the reclaimed slur, reinforcing the idea that "ownership" of language is a critical boundary for virtual communities.
The paper also critiques platform capitalism, noting how platforms extract value from Black cultural production (which drives trends and engagement) while simultaneously suppressing those same voices through over-moderation.
Conclusion
The authors acknowledge limitations, including a small survey sample size and the difficulty of perfectly identifying "Black Twitter" via hashtags. However, the study underscores the need for "community-centered" AI design where marginalized users have a say in how moderation systems are built and governed. The proposed ethical checklist offers an actionable framework for researchers to move beyond simple "one-size-fits-all" toxicity filters toward more culturally aware engagement.
Summary updated