Democracy Technologies: Can you explain the basic idea behind TrollWall and the motivations to start it?
Tomáš Halász: I was working as a community manager, managing online discussions on Facebook for a variety of pages. After the Russian invasion of Ukraine, I saw a growth in the amount of hateful and toxic comments. At the time, I was manually checking and removing vast amounts of comments. It was very time-consuming and also had a bad effect on my psychological well-being and that of other moderators I worked with. I tried to find another tool that would hide those comments automatically. But there was nothing available that met our requirements.
At the same time, we also noticed that the amount of online hate was pushing minorities, women, and generally decent people out of the discussion. When people see the discussion is full of hate and toxic language, they don’t join. So I called Filip first and then another I.T. specialist to make the first version of TrollWall. Once we tested the beta version and it was up to our standards, we approached NGOs here in Slovakia. We gave them the tool for testing for free and support to our civil society. And then we contacted the communications team of the Slovak President Zuzana Čaputová, as well as the team of Petr Pavel – at the time a candidate in the Czech presidential elections, which he went on to win.
DT: You mentioned political campaigns – does TrollWall also work to tackle disinformation?
Halász: That’s a good question. We have a detector for disinformation websites. We have a list of NGOs who deal with disinformation, and who provide us with the official list of disinformation websites. So if our client wants, they can forbid anybody to link to these sites. But when it comes to spreading false narratives, to be honest, that is harder. We have found that disinformation usually goes together with hate speech. For example, if somebody is attacking Ukrainians, they usually don’t use decent language.
Filip Strýčko: The problem is finding out if something in a sentence or a piece of text is disinformation, malinformation, or misinformation. It’s hard both for computers and humans. Right now, it’s a huge research topic, and it’s something that is technically very hard to do.
Halász: But from the data we gathered and from our experience, we see that once you start moderating hate speech, the amount of disinformation goes down, too, because the narratives are usually pushed through in that form.
DT: Can you tell us how TrollWall’s moderation works?
Halász: Once connected to our TrollWall platform, clients link their social media profiles (Facebook, Instagram, YouTube, TikTok) via the official API. Comments are sent to our AI for assessment, which determines whether they should be hidden based on its training. The process, including real-time monitoring and client control, occurs within seconds. We’ve successfully moderated live discussions with thousands of comments instantaneously. For example, we moderated a discussion between Zuzana Čaputová, who was being attacked because she’s a woman, and Czech President Petr Pavel, who was also being attacked, and the moderator said he didn’t have to do anything because TrollWall instantly deleted the hateful comments.
Strýčko: We believe that our technology and our processes are superior to tools like ChatGPT, both in terms of speed and cost. ChatGPT and such large models are very expensive to run and they’re very slow. It’s impossible to use them for a live discussion, like a presidential discussion.
We also believe TrollWall is better in terms of accuracy because of how we collect and focus on linguistic differences. This is especially important for languages that get overlooked by larger firms like Microsoft or OpenAI, who mainly focus on English. Also, if I may add: Each language has its own swear words and hate phrases, which are very culturally specific. That’s why we ensure accuracy by using real words and comments from the language in question, annotated by native linguistic experts. Our data is independently labelled by at least three native speakers, allowing us to quickly adapt to new developments with updates released four times a month.
DT: You mentioned the presidential campaign. What impact did TrollWall have on online discourse and public engagement during the campaign?
Halász: We collaborated with Petr Pavel’s campaign team during his presidential run, utilising TrollWall to manage their social media profiles, which were being inundated with hateful comments. Initially handling around 10,000 comments weekly, it became unmanageable, peaking at 100,000 comments per week by election day.
TrollWall automatically filtered out hateful and vulgar comments, fostering a more positive and substantive exchange of ideas. This was the main aim of the campaign team. In these days of social media campaigning, it is unfortunately a common misconception that hate brings political reach. Some political candidates don’t want to moderate because they think this will decrease their reach. However, our data from different tests and campaigns shows that moderation of hateful comments doesn’t affect reach. It actually brings decent people back to the discussion. Other politicians also think that they should not moderate because of freedom of speech. But given our recent history in Europe, I think we can agree that one has a right to free speech, but not the right to hate.
DT: What were some of the key challenges in setting up this project? How did you overcome them?
Halász: As Filip mentioned, language was one of the main challenges. Understanding the local context of hate speech was also crucial, as AI comprehends languages but lacks contextual understanding. However, thanks to the extensive data we’ve fed the AI, the project is now scalable, enabling us to onboard new clients regardless of language or country.
Another major challenge is convincing people of the seriousness of online hate speech and their responsibility to moderate it. Administrators have the duty to moderate discussions on their platforms — it’s their domain, their responsibility. Before AI, this was not easy to achieve. However, now there’s no excuse for neglecting moderation. Our data shows that if the 100 largest pages in each country commit to moderation, it will significantly impact societal discourse.
