The Privacy-preserving Hate Speech Detection Challenge (PrivHSD)

Mentored by a researcher from the Technical University of Munich

• The Problem: The challenge operates at the intersection of Hate Speech Detection (HSD) and privacy protection, particularly with the ultimate goal of maximizing the “privacy-HSD” trade-off, where the optimal scenario maximizes HSD performance while minimizing loss to personal privacy protections. Using state-of-the-art techniques and metrics from both research fields, we challenge participants to innovate ways in which HSD can continue to function as urgently needed, but not at the cost of creating harmful re-identification tools. The team that can achieve the best trade-off wins!
• The Objective: Innovate a new HSD method that operates agnostically to the author’s identity. The goal is to maximise the “Privacy-HSD trade-off”— achieving high detection performance while minimising the loss of personal privacy protections.
• Potential Output: A Demonstrable Concept or working prototype accompanied by a research note articulating its rights-based architecture and privacy-preserving features. The winning team should expect to share their full results and take part in publishing their results in a scientific publication.

Coordinated Inauthentic Behaviour (CIB) Identification

Mentored by a senior fellow from the Institute for Strategic Dialogue

• The Problem: Hate-based narratives are increasingly driven by “smarter” coordinated networks rather than organic user behaviour.
• The Objective: Build a tool that detects when hate-based narratives emerge across platforms and accounts in a short window of time. LLM enabled semantic clustering can be used instead of being limited to keyword-based queries to group mentions by intention and meaning rather than simply by identical language. Participants should include at least one gendered campaign (e.g. coordinated harassment of women politicians or activists) among their test cases, as these are among the most documented forms of CIB and provide a concrete benchmark for the tool’s real-world relevance. An important threshold is determining the line between coordination and organic behavior from many authentic users.
• Potential Output: Visualized summary of narrative clusters in a dashboard designed for use by analysts or journalists. The dashboard should allow filtering by targeted identity group, including gender.

From reporting to response: Addressing online hate and gender-based violence

Mentored by a researcher of Democracy Reporting International

• The Problem: The challenge responds to a practical implementation gap during electoral periods, when harmful content is often encountered online but remains difficult for institutional actors to document, assess, and escalate. While legal frameworks and multi-stakeholder mechanisms already exist, reporting pathways are often fragmented, unevenly accessible, or not designed around gender-sensitive and electoral-risk criteria. The goal is to support timely, proportionate, and rights-respecting responses without undermining legitimate political debate or freedom of expression.
• The Objective: DRI invites teams to design a tech-enabled reporting and referral solution aimed primarily at institutional actors, including but not limited to electoral commissions, trusted flaggers, and platform integrity teams. Rather than focusing only on the reporting interface, the solution should help turn user-submitted reports of suspected online hate speech and gender-based violence (OGBV) against women candidates, as well as trans and non-binary candidates, into structured evidence that can support classification, assessment, prioritisation, and clear escalation or referral logic.
• Potential Output: Teams are expected to present a prototype/MVP, wireframe, or proof of concept that shows how reports of suspected online hate speech and OGBV would be collected, structured, classified, prioritised, and referred to relevant institutional actors. The output should include a proposed user flow, key data fields, classification criteria, and referral/escalation logic.

 Think out of the box

Mentored by an academic researcher and lecturer from the University of Strasbourg

• The Problem: What if tackling hate speech was not only about removing harmful content—but about transforming the social dynamics that produce it? Hate speech online is a growing and complex societal issue, affecting social cohesion, mental health, and democratic discourse. While current solutions focus largely on detection and content removal, they often fail to capture context, add ress bias, or empower users. This challenge invites participants to think out of the box and go beyond traditional moderation approaches. Rather than focusing solely on detection, teams are encouraged to explore new ways to prevent, contextualize, counter, and transform hate speech in digital environments.
• The Objective: Design a solution that prevents, contextualises, counters, or transforms hate speech rather than simply moderating it. Teams are encouraged to combine technology, design, policy, and creative approaches, targeting any relevant group from social media users and platform moderators to educators, NGOs, or policymakers. An intersectional approach to the work, including gender, age, origin and/or other grounds of discrimination is encouraged.
• Potential Output: A functional prototype or MVP in any form (web or mobile application, moderation support interface, educational experience, policy framework, or creative concept with social impact), accompanied by a 5-minute pitch and a short concept paper outlining design choices, methodology, and ethical considerations.