CONCERNS
The negative environmental and climate aspect of AI is possibly less talked about than ethical
considerations. Fact remains that training an AI consumes large amounts of water (for cooling
in data centres) as well as electricity, and electronic hardware. In the case of this AI-model, the
day-to-day use will be less resource demanding than large-scale AI language models, due to
the much lower number of daily users.
Finally, it is important to not place too much faith in AI. This model will be able to help analyse
hate speech at a larger scale than manual screening, but it cannot address the underlying
cultural, political, and structural drivers that fuel misogyny. It is one aspect of a the complex,
interconnected approaches needed to improve gender justice. It is important to use the data to
advance our knowledge and understanding of gendered hate speech online and of TFGBV,
and that this knowledge is used in advocacy, policy making and practices, to create the
changes we aim for.
6. RECOMMENDATIONS
Automated systems can rapidly process large volumes of content and identify overt patterns,
but they lack the cultural, social, and contextual understanding necessary to fully interpret
nuanced or coded forms of abuse. Therefore, core recommendations include to:
approach AI as a complementary tool to human expertise. Ideally, researchers, CSOs,
activists and others using the developed model should adopt a hybrid workflow that
combines the AI-detection with human review. By integrating current knowledge and
expertise, lived experiences of specifically exposed online users and contextual
understanding, findings will be both accurate and meaningful for advocacy, prevention and
protection.
build and develop one’s own capacity, both on how to operate the tool, and on an
understanding of the model, together with critical analysis. The intention is that the model
should be useful for strategic monitoring, including identifying patterns, and online trends,
for some time to come. Using this AI-tool has already been facilitated by several trainings
for researchers, journalists, CSO representatives and activists.
Promote the tool to public institutions in Egypt, who could also benefit from using it.
consider how to keep the data model updated, to ensure accurate classification and
processing. Possibly there is interest at a university or similar to “inherit” the model,
provided it will continue to be open source, for the purpose of keeping it fully up to date,
seeing as online language, slang, memes and expressions change rapidly.
for anyone working with large scale data harvesting to be required to develop internal
protocols for data management, including secure handling of sensitive content.
AI-generated evidence can also be used as a foundation for advocacy aimed at improving
platform accountability. It could potentially be used to pressure social media companies to
strengthen moderation policies, respond more quickly to harmful content, and other actions
that are known to be conducive to more constructive online interactions. As mentioned earlier,
12