Facebook, Twitter Struggling in Fight against Balkan Content Violations

Interesting investigation reveals serious holes in content policing by Facebook and Twitter in the Balkans.

The authors have quoted  Sanjana Hattotuwa, Special Advisor, ICT4Peace Foundation, in this article. The full Interview with him can be found below:

1. How would you assess the effectiveness of control of the disputed content on FB and TW in the languages of the people from non-English language areas, based on your research and experience?

My research focussed on Sri Lanka. I set up the first Facebook & Twitter accounts for civic media in the country and South Asia. I continued to curate both accounts daily till mid-2019, and have maintained a very light touch since. I have close ties with news feed, platform, electoral integrity teams as well as those working on disinformation in both companies. At the Foundation, I have worked in South Thailand, Myanmar, Maldives, the Balkans and Afghanistan, again using both platforms extensively. Since 2007, but more intentionally and intensely since early 2018 (around the time of Mark’s testimony to Congress), the promise of content oversight in non-English / local languages is a far cry from promise. Much has improved by way of AI and ML recognition as well as contextual awareness of content, but overall and put simply, the scale of the problem is greater and more complex than the availability and grounding of the tools in either company to address adequately. Furthermore, the pace of hate innovation is greater than the development of tools to address them, which means that there is often a disconnect between what is reported on globally and available in English (or any of the world’s major languages, for example, spoken at the UN) and those that in Burma/Myanmar, Sri Lanka and in countries like India (with tribal languages) only spoken in those countries or regions.

2. Facebook notes that for certain content such as nudity, it is not necessary for the one who evaluates the content to speak the language, but to what extent it differs when it comes to content such as hate speech, threats or harassment? How important is the hybrid – AI + human assessment model?

Context is key. A nipple can be male and fine, or female and taken down even if it is was for political art, or for example, around breast cancer awareness. A nipple is thus, not a nipple. No AI and ML I am aware of even in English language contexts can accurately identify the meaning behind an image – though I have been following this field and much has changed in just the past 2-3 years, albeit with the kind of bias you can expect – works better with white skin, in Western contexts. With regards to content inciting hate, hurt and harm, it is even more of a challenge. I have noted in public the degree to which I have, pro bono, helped Facebook on tooling and response. Twitter is not Facebook, but also face significant problems even though, beyond the scope of this email and response, their approach to the challenges are fundamentally different to and better than Facebook’s. But this is a shifting field, and in general, non-English language markets with non-Romanic (i.e. not English letter based) scripts are that much harder to design AI/ML solutions around. And in many cases, these markets are out of sight and out of mind, unless the violence, abuse or platforms harms are so significant they hit the New York Times front-page. Humans are necessary for evaluations, but as you know, there are serious emotional / PTSD issues related to the oversight of violent content, that companies like Facebook have been sued for (and lost, having to pay damages). In sum, the complexity of context as well as the volume of toxicity demand two competing solutions. The first requires AI/ML tooling that needs to be trained on cultural, contextual, historical, identity, religious and political nuances that are often contested, complicated and utterly confusing, even for those who are from the regions concerned. The second is a better fit for AI/ML simply because the everyday tsunami of content is impossible for humans to oversee. A combination of both is required, but in this mix, non-English language markets suffer, and will continue to suffer, since there is no easy or quick fix to this.

3. In your opinion, what needs to be done to improve the whole process?

The introduction of ethics, algorithmic transparency and human rights norms into the core business operations of Silicon Valley companies.

4. Do you think that the categories in which violations are divided, are clear enough? For non-english speakers, how difficult is it to choose the appropriate category to which a particular abuse belongs, bearing in mind that something can violate several rules, but not be assessed as an abuse due to the lack of context? The same applies to the proactive removal of content, as evidenced by examples in which satirical content has been removed or content that deals with the research of some tricky topics, such as war crimes and the like.

The in-app or web-based reporting process is confusing, and moreover, is often in English even though the rest of the UI/UX could be in the local language. Furthermore, the laborious selection of categories is, for a victim, not easy – especially under duress. The rest you have already answered in your question, or I have already alluded to.