Please flip your device.

AI Stories · The Truth is Out There

Defusing Hate Speech

At the time of this interview, Raashi Saxena was the global project coordinator of Hatebase at the Sentinel Project, working with a global community of contributors via the Citizen Linguist Lab. She is also a contributor to several other civic and internet governance communities, including as co-chair of The Internet Rights and Principles Coalition.

This is an extended cut of the interview from The Truth is Out There that has been edited for ease of reading.

What is the link between hate speech and disinformation, especially around elections?

There is an overlap. You can have hateful content in misinformation. And a lot of hate speech can also be seen as misinformation. So these two phenomena are very closely related. We have this analogy where we say, “Hate speech loads the gun, but misinformation pulls the trigger.” And what we mean by that, is that when you have a rumor that aligns with a pre-existing hostility, this makes violence more likely.

At the Sentinel Project, we’re particularly interested in understanding how online hate speech can lead to offline violence. We’re also interested in how offline atrocities can affect situations online. So we examine hate speech from a very contextual perspective. Hate speech in itself might not contribute to offline violence, but it sets a tone of background hostility towards a particular community. And then rumors in the form of malignant and harmful information circulate and can lead to election violence.

What is Hatebase?

Hatebase is a monitoring tool that can be used to analyze hate speech terms across the world. It’s a repository of hate speech in 95+ languages in 175 countries. We analyze publicly available conversations, a lot of which are on social media platforms. It was built to be collaborative, so that companies, government agencies, NGOs, communities and research organizations can use it to help moderate online conversations and potentially gain insights to predict regional or localized violence.

Getting more regional and linguistic inputs is a huge challenge as we’re a very small team. The environments we work in are usually high friction, with civil wars or conflicts. Our data supports a lot of research, and we have relationships with around 350 universities around the world. We have done misinformation management work to mitigate offline violence in Kenya, South Sudan, and Congo, and we’ve also done work in Sri Lanka and Myanmar.

How do you get input in so many languages? 

We have something called the Citizen Linguist Lab which is a form of crowdsourcing for anyone that wants to amplify and augment our database. You don’t have to be an expert to contribute. You can input terms with their contexts, categorizing them based on gender, sexual orientation, ethnicity, and class. You can also provide a citation or more information about the term’s origin. Contributors can also offer an assessment of the ‘offensiveness’ of a particular term. This lets us crowdsource sentiment analysis.

Terms change over time, or may only be offensive to a particular community or in one part of the world. The offensiveness rating helps us to understand. Communities are an integral part of any project to learn who is impacted by hate speech and misinformation. They’re also essential to help us direct research towards facets that matter the most by assisting with data collection, providing linguistic nuances, and highlighting different social and cultural considerations.

What are your thoughts on content moderation?

Big tech puts emphasis on stamping out content that violates their rules, but we’ve found this is mostly a losing proposition. Rather than doing reactive work, we believe there should be more effort placed on educating users on how to detect misinformation and providing impartial guidance on how to verify information. There are things that fall within a gray area that aren’t necessarily hate speech, or get lost in translation.

Social media companies use automation for moderation at a base level, but when it gets escalated to a secondary level, they might not have the local expertise necessary. One fix is to hire content moderators that understand local contexts. I think educating users on how to detect misinformation is better. Companies can delete content, but censorship seems to have a way of reinforcing the beliefs of people involved in spreading propaganda. You’re essentially removing the hateful term, but not the intent. We also see speech labeled as hate speech, when it’s just disliked by a particular group. So these systems can also be used to silence people.

We don’t believe in policing speech. With Hatebase, we monitor speech to understand it for research purposes. Data alone cannot reveal the full picture. We need local guidance. It’s a fruitful path that more technology companies and platforms should pursue. AI is not a magic wand that will automatically wash away hate speech.

What role is there for automation in Hatebase?

Working on this topic day in and day out, we understand the limitations of AI to identify hate speech, and we are pragmatic about it. One thing we do is collect the metadata of Hatebase terms we detect in public conversations. We call these “sightings” and there have been around a million so far. Soon, we expect to expand the natural language processing (NLP) engine of Hatebase — called HateBrain — so it also analyzes the textual context (ngrams) of terms, plus frequency and proximity patterns. These additional data points will increase Hatebase’s sensitivity to language, and help to expand the dataset.

Mozilla has taken reasonable steps to ensure the accuracy of the statements made during the interview, but the words and opinions presented here are ascribed entirely to the interviewee.

Portrait photo of Raashi Saxena is by Hannah Yoon (CC-BY) 2022