Information manipulation is the deliberate transformation of a message’s content to steer audience perception and weaponise narratives for strategic gain, eroding public trust and destabilising democratic institutions. This thesis studies information manipulation through three core mechanisms: persuasion techniques, credibility signals, and personalisation. These mechanisms exploit cognitive and emotional biases, mimic perceived credibility, and tailor content to the recipient's beliefs, respectively.
With the emergence of generative Large Language Models (LLMs), the weaponisation of such mechanisms can now be automated at unprecedented scale and speed, while LLMs also offer novel means to detect, analyse, and mitigate these threats. However, existing research is typically confined to individual domains, the English language, and a small set of manipulation signals or exploitation settings.
This thesis by publication has four contributions encompassing analysis, detection, generation, and exploitation of information manipulation using LLMs. Firstly, we provide a large-scale characterisation of 16 persuasion techniques across four domains, revealing how manipulative style is strategically adapted to socio-political contexts. Secondly, we introduce PASTEL, a detection framework that prompts LLMs to extract credibility signals and combines them via weak supervision to classify misinformation. Thirdly, we present AI-TRAITS, the largest multilingual dataset of LLM-generated personalised disinformation, demonstrating that LLMs can tailor disinformation narratives to audience profiles at scale while consistently bypassing their in-built safeguards. Finally, we propose Persuasion Injection Attacks, an adversarial framework that uses LLMs to rewrite false claims into persuasive variants, significantly increasing their evasion rate on automated fact-checking systems by disrupting evidence retrieval and verification.
Together, these contributions systematically characterise the mechanisms of information manipulation across domains, languages, signals, and exploitation settings, highlighting the dual nature of LLMs, which amplify manipulative capabilities whilst providing essential means to counter them. The thesis further contributes novel models and resources to support the development of more resilient defences.
