HATEDEMICS is a research project with a focus on addressing online fake news and hateful content, with teams based in Italy, Malta, Poland and Spain. The project aims to use AI to help detect and counter any disinformation it finds on social media platforms and the web.
The name derives from ‘hate speech and infodemics’, defined in full as “Hampering hate speech and disinformation through AI-based technologies to prevent and combat polarisation and the spread of racist, xenophobic, and intolerant speech and conspiracy theories”.
HATEDEMICS, which began last April, hopes to empower non-governmental organisations (NGOs), public authorities, young activists and fact checkers to effectively combat polarisation online.
It has 14 consortium partners, including ALDA (France), Stowarzyszenie Demagog (Poland), and Malta’s Ministry for Home Affairs, Security, Reforms and Employment.
HATEDEMICS: A human-machine collaboration
HATEDEMICS is starting from the point of view of practitioners, NGO operators and fact checkers, who have guidelines on how to respond to hate speech and fake news.
The project focuses on counter speech put forward by these NGOs or volunteers, who directly answer hate content they find online. When this can’t be done manually, AI can help in answering hate speech to generate counter speech, as part of human-machine collaboration.
HATEDEMIC’s primary objective is to tackle online hate as well as the interrelated issues between hate speech and disinformation, which can sometimes be overlooked.
There is also the aim to raise awareness and improve critical thinking. This will be done by advancing AI-based technologies which will record hate speech and estimate the ‘HATEDEMICS’ risk, which is the potential exposure to misinformative messages.
Dialogue-based counter-narratives will be produced, which support professionals and activists by implementing an effective human-in-the-loop approach, where human operators will always be given the final say on posted content.
Fighting hate speech and misinformation together
Marco Guerini is scientific coordinator of HATEDEMICS and head of the Language and Dialogue Technologies group at Fondazione Bruno Kessler (FBK), one of the project partners.
“We always want to have humans in the loop. We want to help civil society citizens and operators in answering speech,” Guerini says.
“What’s missing is a project, an idea, an effort to address that grey area that stands between misinformation and hate speech. Why? Because in many cases the two go together. In many cases you find hate speech which is supported by misinformation, so you cannot just fight hate speech or misinformation.
“You have to combat the combination of the two. So the idea of this project is to put together practitioners in order to create guidelines, educational materials and training data to develop an AI platform that can empower civil society in fighting hate speech and misinformation together.”
Using AI to identify hate speech
Guerini, who works in natural language processing (NLP), a branch of artificial intelligence, shared how hate speech has been tackled within a framework of identification and sanction strategies for many years.
Hate speech is usually identified online by users of social media platforms: they report it, and then there are sanctions usually put forward by the administrator of the platform. However, due to more people being online and the use of automatic tools, Marco said there is a growing amount of hate and disinformation produced on a daily basis which poses a problem for platforms.
“That’s where NLP and AI came into play and started to produce tools, usually based on neural networks, tasked with outputting whether a text is hateful or not.
“This neural network gives back a label saying whether this is hateful or not, and the sanction is up to the moderators or the owner of the platform,” he said.
Using AI to generate counter speech
HATEDEMICS also uses AI models to generate counter speech for any hate detected online. The project team fine-tune these models to produce more sophisticated responses, which Guerini suggests current tools are lacking.
“If you take a commercial or even open-source tool, what you get most of the time are superficial responses. When you ask them to respond, you start realising that the way they respond is in a very stereotypical way, it’s always the same kind of arguments.
“Whilst what NGO operators or fact checkers do is a much more deep and meaningful way of responding.”
HATEDEMICS asked experts including operators and fact checkers to write various examples of counteracting contributions to an online discussion involving hate speech. This would then simulate dialogue to train the AI language model to address a situation as an expert would.
Students will test the tools in a simulated environment, where counter speech will be generated based on real-life situations.
Guerini says: “In principle, the tools can be used by anybody. We want to empower the civil society at large. In this particular project, the main final users will be students testing our tools.”
The AI researchers at HATEDEMICS did not choose to focus on a particular group in society, which they have done in a past project on anti-Muslim online hatred, but rather aimed to counter hate speech against various groups.
Preserving the freedom of expression
At the same time as working towards countering hate speech, the HATEDEMICS project is entirely committed to the concept of preserving freedom of expression.
When highlighting the importance of counter speech, Guerini says the idea is the premise that if we want to fight hate speech, what is needed is not less speech, but more speech: “We try to de-escalate the conversation and try to put forward the reasons for which that particular content is not acceptable.
“We try to have a rational approach to discussion, a dialogical approach to solve the conflict of different points of view.”
Marco explained how one of the other advantages of counter speech is that it can also apply to bystanders online.
“If they are also able to see that there can be an alternate point of view, they might understand that the content was hateful,” he said.
The next steps
The project is currently being trialled in the four member states to design and deploy interactive training. There are plans it will expand beyond Italy, Malta, Poland and Spain to other EU countries.
The project will run until March 2026.
