with human rights principles but also limits the scope of analysis when abusive content occurs
in semi-private or closed online spaces.
SAFETY
This also means that using the AI-tool will not entail processing of sensitive or personal
information. If such information should already be made public, it could end up in the tool, in
which case it is important to determine practices for legal action, and the responsibility of the
researcher or activist that harvests the data if such sensitive or leaked information is
discovered.
Handling large amounts of material, including material with threats, harassment and violence
is inevitable when using the model. This requires careful data protection to avoid exposing
victims/survivors of online abuse to additional risks. There is a need for secure infrastructure,
clear data management policies, as well as robust anonymisation.
Data protection is also related to the need, or encouragement, of securing personal support
systems for researchers and activists who will work with the Egyptian Arabic AI-model. Some,
if not a vast deal, of the material that they will process while working on this topic will be
violent, and research shows that hateful information affects us, even though the messages are
not directed towards us.27 Therefore, it is important both to limit the number of persons viewing
the material, and to support the ones who do.
BIAS, IRONY AND DIALECTS
One of the most persistent technical limitations encountered during the data annotation phase
was the difficulty in accurately interpreting the complexity of human language. Online
communication is rarely straightforward; it involves sarcasm, ironic and coded language,
humour, and contextual references that automated systems need additional efforts to be
trained on, in order to categorise correctly. The lessons learnt from the pilot project and model
with Jordanian Arab dialect was crucial is this case. Also the annotation guidelines were
important, which were developed by the external expert, technical experts, annotators, and
JOSA members, active throughout the process.
Despite these challenges, the Egyptian Arabic AI-tool reached an impressive accuracy during
the different phases of training. It remains important however, to maintain human oversight of
any analysis. The methodology used in is published on the Nuha website, for transparency.
A related challenge concerns the bias within AI models. Machine learning models are only as
reliable as the data on which they are trained, and as we all know, we live in a society of bias.
This is also why gender-based hate speech has been known to be more difficult to detect than
for example racist hate speech.28 Gender biases are still internalised in our societies, meaning
that they are oftentimes unconscious. During this project, biases were countered by the
extensive use of human annotators. This is also why having a feminist awareness was crucial
throughout the development process – both among the technical staff and annotators.
27 See for example: Vicarious trauma after viewing media - Wikipedia
Analyse & Tal, KVINFO and TrygFonden (2025) Digital hverdagssexisme: Hvordan forebygger og
håndterer man den hårde tone i onlinedebatter, https://kvinfo.dk/wp-content/uploads/2025/04/Digitalhverdagssexisme_20250430.pdf
28
11