vera.ai: VERification Assisted by Artificial Intelligence

Recent major events and social trends, such as elections across the globe, the COVID pandemic, and a global rise in populism, along with impressive advances in the field of Artificial Intelligence (AI), such as the automated creation of highly convincing imagery and text (also known as deepfakes), have made clear that the challenge of disinformation can have grave impact on our societies. Media and news professionals, who are still seen as the gatekeepers of accurate and high-quality information, are in dire need of sophisticated solutions to make their ever increasing (both in volume and complexity) verification work more efficient and manageable. To this end, vera.ai seeks to build trustworthy AI solutions against advanced disinformation techniques, co-created with and for media professionals and set the foundation for future research in the area of AI against disinformation. Key novel characteristics of the AI models will be fairness, transparency (including explainability), robustness to new data, and continuous adaptation to new disinformation techniques.

Internal in our organization

Deliver multilingual and multimodal trustworthy AI methods for content analysis, enhancement, and evidence retrieval to assist in disinformation detection and content verification. Deliver multimodal trustworthy AI tools for the detection of deepfake, synthetic media and manipulated content, including AI generated and manipulated content, namely images, videos, audio and text. Enable the discovery, tracking, and impact measurement of disinformation narratives and campaigns across social platforms, modalities, and languages, through integrated AI and network science methods. Provide an intelligent verification and debunk authoring assistant, based on chatbot NLP technology, to explain the AI outputs and guide its professional users through the verification process. Pursue a fact-checker-in-the-loop approach to seamlessly gather new actionable feedback as a side effect of verification workflows. Ensure adoption and sustainability of the new AI tools in real-world applications through integration in leading verification tools (Truly Media and the InVID-WeVerify verification plugin) and AI platforms (AI4EU, ELG) with established large stakeholder communities.

European Union - Ethics Guidelines for Trustworthy AI

Verify multimodal content

Bias-free and explainable AI methods for verification of images, videos, and audio content; explainable multilingual credibility assessment and retrieval of relevant evidence from authoritative sources; multimodal deepfake and manipulation analysis; cross-modal detection of decontextualised content. To ensure the AI models continuously adapt to new kinds of disinformation, vera.ai will also implement a factchecker-in-the-loop approach for seamless gathering of new, human-tagged training data.

Analyse disinformation agents

A network-science approach and tools for analysing communities and spreaders using graph neural networks; deep learning methods for extracting source credibility signals following the W3C guidelines; visualisations of spreader networks and accounts through Truly Media and the verification plugin.

Uncover disinformation campaigns

Deep learning methods for near-duplicate detection of images and videos and clustering of topically and semantically similar messages; knowledge graph models and spatio-temporal analysis of disinformation campaigns; network science methods for discovery of coordinated sharing behaviour.

Assess and reduce impact

Development of a first-of-a-kind AI-based debunk writing assistant; research audits of the effectiveness of social platforms and their algorithms on reducing the impact and spread of mis- and disinformation; research into the effectiveness of debunking and fact-checking.

Assist professionals

Development and integration of novel AI-based verification and debunk authoring assistant that incorporates a chatbot advising professionals on how best to use the new AI tools and interpret their outputs and offers assistance with debunk authoring through automated evidence gathering and snippet authoring; also encouraging users to flag mistakes in the AI outputs, which will then be used for retraining.

Developer, External domain expert, Academic researcher, Industry partners

Workshops, Survey, Interviews, Co-creation process

Yes following a co-creation methodology.

No

Yes

By academic researchers

5

Data minimization (incl. not gathering personal data), Authentication and access management, User control (consent, update, retract)

vera.ai acknowledges that the benefits of fighting disinformation using AI should not come at disproportionately large environmental costs as a result of massive AI model training and deployment. To this end, care will be taken to select good trade-offs between accuracy and model size (which translates into higher energy consumption), employing more compact architectures whenever possible.

Project partners may re-use existing open datasets available online in compliance with their license and applying GDPR provisions. vera.ai datasets will be made available as soon as they are collated and cleared through ethics and data management.