Automated censorship in Arabic
Machine moderation and the speech it silences
This event is part of a demonstration of what Lera offers for running knowledge events. Its content is written with care and draws on real sources, but the conference has not taken place, and apart from the convener its participants are fictional.
In May 2021, during the escalation in Jerusalem and Gaza, Palestinian and Arab users reported mass removals of their posts on Instagram and Facebook. Meta commissioned an independent review from BSR, whose report was published in September 2022. Its finding was clear: Arabic content had been over-enforced and Hebrew content under-enforced, and part of this imbalance came from the fact that Meta had an automated classifier for hostile speech in Arabic and none for Hebrew.
Two years later, in December 2023, Human Rights Watch reviewed more than a thousand cases from over sixty countries and called the censorship of Palestine-related content "systemic and global". The report traced the problem to flawed policies, inconsistent enforcement, over-reliance on automated tools and undue government influence over removals. In March 2024 Meta's Oversight Board also recommended ending the blanket removal of posts containing the word "shaheed".
Layla Haddad, who has worked for years in Beirut publishing, argues that the problem is not only political; it is also linguistic. Arabic is a family of dialects, sometimes written in the Latin alphabet, and full of religious, poetic and ironic expressions. A classifier trained on limited, decontextualised data reads a verse of the Qur'an or a word of mourning as incitement, and takes irony at face value. The machine's error here is not random; it comes from a place where another language is the centre.
The consequences for knowledge are deep. Journalists, archivists and publishers censor themselves before publishing. The documentation of events disappears, and what a period leaves in memory is decided by a tool that thinks in another language. Sarah T. Roberts has shown that content moderation is hidden, outsourced work whose workers go unseen; so do its rules.
Haddad makes four proposals: transparency, broken down by language, on removal rates and the outcome of appeals; human review by people who know the dialect and the context; preservation of removed content for research and for justice, instead of its destruction; and for publishers and archives to have their own infrastructure rather than entrusting their memory to platforms whose rules are written elsewhere.
Questions for discussion
- Should transparency broken down by language be a legal obligation for large platforms?
- Where, and under whose supervision, should removed content that is a historical record be kept?
- How can a knowledge infrastructure stay safe from automated censorship without itself becoming a closed platform?
References
BSR. 2022. Human Rights Due Diligence of Meta's Impacts in Israel and Palestine in May 2021. San Francisco: BSR, 22 September 2022.
Human Rights Watch. 2023. Meta's Broken Promises: Systemic Censorship of Palestine Content on Instagram and Facebook. New York: Human Rights Watch, 21 December 2023.
Oversight Board. 2024. Policy Advisory Opinion on the treatment of the term "shaheed" (PAO 2023-01). 26 March 2024.
Roberts, Sarah T. 2019. Behind the Screen: Content Moderation in the Shadows of Social Media. New Haven: Yale University Press.