This month we were very happy to sit down with one of the brains behind Checkstep who is also a recognized talent among European academics. She is the co-head of research at Checkstep and also an associate professor at the University of Copenhagen.
She currently holds a prestigious DFF Sapere Aude Research Leader fellowship on ‘Learning to Explain Attitudes on Social Media’, and was also recently admitted to the Young Royal Danish Academy of Sciences and Letters.
In this interview she explains how she got involved in the Trust and Safety space and the important role researchers play in finding solutions to all manner of online harms.
- What made you get involved in the online Trust and Safety space?
First and foremost, online harms present a substantial societal problem — platforms are rife with abusive language, sexism, misinformation etc. Short of bringing about a cultural change, what one can realistically do to improve the status quo is to semi-automatically moderate harmful content. This, in itself, is very challenging from a technical point of view, which, as a researcher, I find exciting.
2. People tend to trust content that resonates with their beliefs or generally helps address some of their doubts. Thus, making it easy for misinformation propagators to easily target potential “victims”. Given your extensive work around fact-checking, do you think debunking some of these claims is the best way to keep society better informed?
Yes, prior work in psychology shows that it is very difficult to change peoples’ core beliefs and values. Thus, it is important to detect disinformation as early as possible before it spreads, and to provide automatic fact checks to content moderators for this purpose.
3. Better understanding the context of certain conversations helps us to put things into perspective, especially when it comes to social media. Could you tell us a little bit about your research project EXPANSE, that aims to explain attitudes on social media?
EXPANSE is a research leader fellowship I recently obtained (more information here: https://dff.dk/en/grants/copy_of_research-leaders-2020/researchleader-14).The project itself started on October 1, 2021, so there are unfortunately not so many results to share yet. Very briefly about the goals of the project though: the overarching aim is to be able to detect attitudes automatically (also called stance detection), but do so in a much more fine-grained and transparent way than is possible today. The key innovation is to imbue stance detection models with sociological knowledge, as I hypothesize they can shed a light on why people hold certain attitudes, and thus lead to more insightful automatically generated explanations.
4. Content moderation is a growing concern, given the recent infodemic. However, some criticize it as a means to suppress freedom of speech. How should content moderation companies position themselves? What are some of the areas they should focus on, to ensure online safety?
The concept of freedom of speech has existed since the 6th century BC, long before social media or even print media were invented. Before social media, it was much more challenging to spread and weaponize information — whereas now, everyone with access to the internet can do so, anonymously and with few repercussions. This format, by design, brings out the worst in people — things people would never feel comfortable saying to someone’s face, they feel comfortable writing in an anonymous online forum. The filter bubble effect means people additionally receive backing on their opinions from like-minded individuals. This means, in this day and age, one needs to very carefully weigh freedom of speech up against the real harms it can cause. One area I find particularly concerning is the negative impact of social media, especially image-sharing platforms, on depression and suicide among teenagers due to the distorted views of reality presented by many users on these platforms, including related to body image and lifestyle. I think a careful audit of such platforms is needed to address this problem more holistically, but content moderation can at least help to identify particularly harmful information, such as posts glorifying anorexia.
5. How can AI be applied to help with the problem? How does AI explainability help?
AI-based content moderation solutions can help identify harmful content before it even reaches users. Explainable AI can be useful in two ways: one, it can help continually improve content moderation models by identifying why they sometimes make mistakes, and two, especially for knowledge-intensive tasks such as automatic fact checking, they can provide content moderators with more information about why a model arrived at a prediction to make it easier to manually verify if the prediction is reliable.
6. The news is full of how bad actors propagate misinformation and also how platforms seem to exacerbate the problem. What role do you see for academics in addressing these problems?
Academics can provide crucial insights into why this phenomenon occurs, as well as potential solutions to the problem. For misinformation specifically, academics from many different disciplines have important and complementary research findings, which should be taken into account — e.g. from psychology, about the perception of misinformation; from computer science, about how to develop automatic content moderation solutions; from law, about how legislation applies to online platforms in different countries.
Academics can thus help inform online platform developers on how to make platforms safer for everyone, content moderation companies on how to automatically detect harmful content, and, perhaps most crucially, decision makers in governments on how to develop new legislation related to online harms.
PS. Something to look out for –
Isabelle’s higher doctoral dissertation defence, for earning the title of Doctor Scientiarum, on “Towards Explainable Fact Checking” on 6 December: https://di.ku.dk/begivenhedsmappe/begivenheder-2021/doctorate-defence-by-isabelle-augenstein/