continued until a satisfactory accuracy level (F1 score of 82.0% for the main categories
(binary) and 77.8% for sub-categories (non-binary)) was reached. It is important to highlight
the work of JOSA’s annotators’ and external expert’s efforts throughout the project, as they
played a crucial role; helping in collecting account lists; reviewing and enhancing methodology;
participating in developing annotation guidelines; developing and training annotators;
reviewing annotation and providing quality assurance for the model; developing mitigation
strategies for challenges faced regarding annotations; annotating the collected data and
completing quality checks.
Tadwein, as an organisation strongly rooted in Egyptian feminist discourse, provided
conceptual guidance on how to distinguish harmful, gendered violence from legitimate speech,
helping to mitigate risks of overreach, misclassification, or political misuse of the AI tool. This
distinction is essential both ethically and technically, particularly given the restrictive civic
space and the history of state surveillance and prosecution of marginalised groups. Tadwein
further strengthened the knowledge base by embedding Egyptian socio-political and legal
context into the conceptual framework.
Practically, Tadwein monitored 75 individuals, initiatives, and pages across social media
platforms, including Facebook, Instagram and TikTok, to identify potential hate speech and
TFGBV-related content. This monitoring was conducted in alignment with the criteria outlined
in the project methodology. For each selected account, Tadwein provided analytical
comments, direct URLs to the accounts, and links to relevant posts identified for further review
and analysis. The sample included additionally over 75 accounts, monitored by JOSA and
their data annotators. The two organisations worked closely, to ensure that the selected
accounts and content were consistent with the Egyptian context and fully aligned with the
agreed methodology.
Initially, Tadwein reviewed the methodology section and systematically flagged unclear or
ambiguous points that could potentially mislead the monitoring and data annotation processes.
These observations were shared to support methodological clarity, consistency, and accuracy
throughout the implementation of the monitoring and annotation phases. The main work and
responsibility for the methodology remained with JOSA.
RELEVANT CONTEXTS
The AI-based tools are valuable in a research, media and civil society contexts. They enable
organisations to monitor large volumes of public online content in real time or retrospectively,
identify spikes in abusive language, and uncover recurring narratives or coordinated
harassment campaigns. AI-tools can also support comparative analysis across platforms, time
periods, or geographic areas, making it possible to document trends and patterns that would
otherwise remain invisible. For advocacy purposes, the data generated through these tools
can strengthen evidence-based engagement with policymakers, social media platforms, and
international human rights mechanisms.
A MORE RELIABLE AND ACCESSIBLE TOOL
While many commercially available AI tools offer limited insight into how annotation-decisions
are made or how models are trained, this model gives users insight into which data has been
analysed, while knowing that the training and development process has followed ethical
8