identity, and technology intersect, and how digital misogyny and harm translates into physicalworld consequences. For CSOs, journalists and researchers, who are targeted by this AImodel, this understanding is a necessary foundation for data processing and analysis.
3. THE AI TOOL FOR DETECTION AND
RESEARCH
The increasing scale, speed, and complexity of online hate speech means that manual
observation and reporting, while valuable, cannot keep pace with the content generated
across digital platforms. In response to this challenge, AI–powered tools have emerged as
important resources for detecting, analysing and researching gendered online hate speech
and TFGBV. These tools offer new possibilities to systematically document digital harm,
generate evidence, improve protection mechanisms and support advocacy and policy
engagement.
CONCEPTUAL FRAMEWORK
Tadwein developed the conceptual framework of this AI model, in collaboration with KVINFO.
Drawing on feminist research, international normative frameworks and localized evidence,
ensuring that the conceptual foundation of the project was both theoretically sound, localised
and contextually relevant.
The term Technology-Facilitated Gender-Based Violence was selected, over more commonly
used but narrower terms such as “online violence,” “cyberbullying,” or “digital abuse”. TFGBV
is a gendered and intersectional form of violence, emphasizing how digital harm
disproportionately affect women, girls and sexual minorities, and how this harm is shaped by
intersecting factors such as age, class, sexuality, gender identity, ability, and social status.
This conceptual clarity was critical to ensuring that the AI model does not treat online abuse as
isolated or neutral behavior, but rather as part of broader systems of gender-based
oppression.
DEVELOPING THE TOOL
During the project, based on experiences from working with Jordanian Arabic, JOSA adopted
a machine learning lifecycle that consists of four phases – discovery, training, deployment, and
monitoring. In the discovery phase, the problem is defined, data is gathered, and goals are set.
The training phase involves building and training the model, using the collected and classified
data. Deployment integrates the model into the production environment. Monitoring
continuously assesses the model's performance and makes necessary adjustments.25
This AI-model has been developed based on large amounts of data (approximately 50.000
comments) humanly annotated twice, by Egyptian Arabic speaking annotators. The training
25 Democracy Reporting International (2024) Online Public Discourse in the MENA Region, Persistent
Tactics of Disinformation and an Increase in Online Gender-Based Violence, 65b21f4539b6e.pdf p.65-66
(NB most of above is a quote)
7