Introduction
In many regions across the Americas, official government data regarding feminicide—the gender-related killing of women and girls—is either non-existent, incomplete, or obfuscated. To fill these critical information gaps, activist groups engage in the production of "counterdata." This process involves the manual monitoring of news reports and social media to document cases, providing a necessary alternative to state narratives. However, this labor is resource-intensive and carries a significant emotional toll, as activists must consume graphic reports of violence daily.
This paper explores a case study in participatory machine learning (ML), detailing the co-design and implementation of ML-driven tools intended to support feminicide data activists. By automating the detection of relevant news articles, the system aims to streamline the creation of counterdata while addressing the mental health burden associated with manual monitoring.
Methodology
Participatory Action Research (PAR)
The researchers adopted a Participatory Action Research framework, ensuring that the community affected by the technology was involved throughout the development lifecycle. The process began with qualitative contextualization, involving interviews with ten activist groups to map their workflows and identify pain points. This was followed by a deep co-design phase with two specific partner groups to brainstorm technical interventions.
The Toolset
Two primary tools were developed to assist activists:
- Browser Extension: A utility designed to highlight key entities—such as names, dates, and locations—within news articles to accelerate the manual extraction of data points.
- Email Alert System: An ML-powered notification system that scans the web for news articles and suggests those with a high probability of being related to feminicide.
Model Development
The project initially utilized a multinomial Naive Bayes approach but transitioned to a logistic regression model. This model predicts the probability of an article being about feminicide based on its textual content. The training data was curated from positive and negative news articles identified and labeled by the co-design partners.
Pilot Study
The system underwent an eight-week real-world pilot with seven activist groups—four based in the United States and three in Latin America. The pilot included weekly surveys, focus groups, and debriefing sessions to gather qualitative and quantitative feedback on the system's utility and accuracy.
Key Results and Performance
Technical Metrics
On a held-out evaluation set of 74 articles, the logistic regression model achieved:
- Accuracy: 81%
- Recall: 93%
In this specific domain, high recall is prioritized over precision. Activists expressed a preference for receiving occasional "false positives" (irrelevant articles) rather than missing a single actual case of feminicide.
User Satisfaction and Efficiency
Groups tracking "general" feminicide reported high levels of satisfaction, with an average ease-of-use score of 3.86/5. These users noted that the system successfully identified cases they would have otherwise missed and significantly reduced the time spent on manual searches.
Discussion: The Intersectional Challenge
Intersectional Failure and Media Bias
A critical finding of the study was the disparity in performance for groups tracking racialized feminicide (e.g., cases involving Black or Indigenous women). These groups reported much lower satisfaction scores (averaging 2.55/5). The ML models frequently failed to find relevant articles for these specific populations.
This failure was not merely technical but reflective of systemic media bias. If mainstream media outlets do not report on the deaths of marginalized women, or if they use language that differs from the "standard" reporting of feminicide, the ML model lacks the necessary signal to function. This highlights a fundamental constraint: ML tools are limited by the biases inherent in the data sources they monitor.
ML as a Mental Health Intervention
Beyond technical efficiency, the study identified the primary value of ML as a tool for mental health. By filtering out irrelevant content, the system limits the volume of traumatic material activists must process. In this context, the ML model acts as a protective layer, allowing activists to focus their energy on relevant cases while reducing unnecessary exposure to unrelated violent content.
Conclusion and Future Work
The project demonstrates that participatory ML can effectively support grassroots activism, but it also underscores the limitations of "one-size-fits-all" models. Future work will focus on:
- Specialized Modeling: Developing niche models tailored to specific intersectional contexts, such as Black women killed in officer-involved shootings.
- Social Media Integration: Incorporating data from social media platforms to bypass the limitations of traditional news media, which often ignores cases involving marginalized victims.
- Continuous Iteration: Maintaining a long-term commitment to retraining models as the language of reporting and the nature of violence evolve.
Ultimately, this research frames counterdata collection as a form of resistance, where ML serves not as a replacement for human judgment, but as a supportive infrastructure for community-led justice.
Summary updated