Webinar presenting veraAI outcomes, focussing on research matters: a summary
More than 70 participants came together online on 24 June 2025 to join for the second “end of series” vera.ai webinar. In this 3-hour-event, public to anybody who had registered previously, project participants presented key research outcomes and technological advancements developed over the past three years. The individual sessions primarily targeted researchers and professionals, focussing on the cutting-edge AI models and methodologies that power the verification tools offered by vera.ai.
Welcome and Introduction
Olga Papadopoulou (project manager based at veraAI coordinator CERTH-ITI) opened the webinar at exactly 14:00 CEST, welcoming participants and outlining the agenda. She then gave the floor to Symeon (Akis) Papadopoulos (also of CERTH-ITI, the acting project coordinator), who offered a high-level overview of vera.ai. This included project goals, components, and the underlying research challenges tackled throughout the project.
Next: presentations and talks by veraAI partners
Text-based Disinformation Detection
Olesya Razuvayevskaya (USFD) demonstrated how multilingual credibility signals can support the verification of suspicious articles and how the use of large language models can enhance the detection and explainability of such signals. Ivan Srba (KInIT), in turn, shared results from a comparative analysis on whether large language models (LLMs) generate or resist disinformation. Findings showed that open-source models are more resistant, while commercial models are more prone to generating high-quality disinformation.
Image-based Disinformation Detection
Luisa Verdoliva (UNINA) opened this section by contextualizing the challenges of synthetic image detection, especially considering the ease of generation through text prompts and diffusion models. She highlighted the evolving threat landscape and limitations in current detection approaches.
Then, Christos Koutlis (CERTH-ITI) presented a series of AI models developed since 2023 under vera.ai, showcasing how the team addressed technical constraints and advanced detection capabilities across datasets and architectures.
Synthetic Speech Detection
Milica Gerhardt and Luca Cuccovillo (both from Fraunhofer IDMT) introduced their innovative approach in audio-based disinformation detection. Moving beyond emotion and prosody, their method focused on formant analysis, leading to improved performance in detecting AI-generated speech — a fast-emerging area of concern.
Video-based Detection: Synthetic and Out-of-Context Content
Luisa Verdoliva (UNINA) returned to the screen a second time, this time to present on the detection of fully synthetic videos, emphasizing how generator-specific traces can serve as signatures for synthetic video detection.
Key Frame Selection and Enhancement
Kostas Apostolidis (CERTH) explained how the identification of key frames within a video, combined with the detection and enhancement of key visual elements, can significantly aid in debunking older videos that are misused in the context of breaking news events.
Spatio-temporal Narrative Detection
In the final session of the day, Xingyi Song (USFD) presented a spatio-temporal narrative detection method, tracking how topics and documents evolve over time.
Martin Hyben (KInIT) showed how to extract check-worthy claims from user content and visualize the information for better insight.
Andrey Tagarev (ONTO) closed off this part by demonstrating the advanced search functionalities of the Database of Known Fakes (DBKF), including cluster-based search and a newly added chatbot interface for querying debunked content.
Recap of the afternoon
As the session moderator, I (Olga, also the author of this article) wrapped up the event by directing the audience to resources for deeper exploration, pointing to:
- the vera.ai Zenodo Community, where all our research publications and datasets can be found;
- the Project Website, with dedicated pages for:
Thank you so much to all attendees and contributors for an engaging and impactful session!
Author: Olga Papadopoulou (CERTH-ITI)
Editor: Jochen Spangenberg (DW)
Addition: a recording of the entire 3-hour event is also available on / via the project website and the project's YouTube channel.
