NLP/LLMs:

Automated Fact Verification (Google Search & Wikipedia): Multi-stage pipelines integrating cross-encoders and LLMs for evidence retrieval, question generation, and veracity prediction. These models significantly enhance accuracy in processing complex claims, extracting relevant evidence, and generating questions.

Toxicity Detection Bias Analysis: Examines biases in NLP models (e.g., representation, selection, and over-amplification biases) and their impact on fairness in tasks such as toxicity detection. It also applies debiasing techniques to enhance fairness by fine-tuning models on balanced datasets.

Poli: Uses NLP to simplify complex financial Terms & Conditions (T&Cs), breaking them down into clear summaries. The tool enhances accessibility for users with low confidence, limited digital literacy, or non-native English speakers.

Graph-Based Techniques:

ARELight: Introduces a graph-based topic modelling framework that links entities and topics through sentiment-rich, directed graphs. This method allows for the identification of abnormal patterns, such as spikes in negative sentiment indicative of disinformation campaigns. It also supports intuitive graph operations for comparing narratives across different timelines or user communities.

Computer Vision and Neural Networks:

AI Detection of Bots: Uses a multi-stage approach to classify bot profile photos from social networks, with neural network pipelines including:

• YOLO for detecting faces.
• Face Recognition for identity verification (including celebrity detection).
• AI-Specific Detectors for identifying synthetic (AI-generated) faces.

Bot Classification: Classifies bots based on photo types (AI-generated, stolen, or anonymous) and analyses their behaviour using metrics such as cost per action, execution speed, quality rating, and ability to deceive moderation. Statistical analyses compare AI-generated bots with other categories.

Qualitative Engagements:

Understanding Young Adults’ Agency in Misinformation-Induced Online Harms: This two-stage, design-led approach engaged 22 stakeholders – including young adults and professionals working closely with them – to explore the contexts in which misinformation is encountered and its real-world impact. Insights from this workshop informed a follow-up co-design session with seven young adults specialising in technology design and social sciences. These participants examined how misinformation spreads in digital spaces and contributed to developing practical, user-centred solutions.

Navigating (Mis)Information in Pregnancy-Related Decisions: Through semi-structured interviews with young women and key stakeholders, including maternal mental health link workers, this study explored how women access and interpret pregnancy-related information. By centring on personal narratives, we identified emotional and decision-making challenges and developed design recommendations to better support women’s self-identified needs during this critical period.

Diverse Stakeholder Perspectives on Children, Young People, and Online Misinformation: In collaboration with Coram Life Education and Roots and Wings CIC, this study gathered insights from diverse, cross-generational groups through individual and group interviews, as well as Vox Pops (Talking Heads) interviews. Participants – including children (ages 8–11), teenagers (ages 12–18), parents, educators, healthcare professionals, advocacy group representatives, and policymakers – shared their experiences with smartphone and social media use, exposure to online misinformation, and attitudes toward the Online Safety Act. The findings contribute to research evidence and media resources designed for PHSE school curricula, supporting efforts to protect young people from online harms.