This series constructs a systematic doctrinal and criminological framework for understanding and addressing adversarial AI misuse through concealed or weakly attributable accounts. Drawing exclusively from existing primary law, binding case law, peer-reviewed science, and established regulatory instruments (EU AI Act, GDPR Art. 17, NIS2, US UPL doctrine, ABA Formal Opinion 512, and related jurisprudence), the series identifies the structural conditions under which human intent is amplified, concealed, and made difficult to attribute when mediated by large language models. The analysis is organised in classical Roman-style form: axioms, definitions, and propositions. It does not propose new legal categories, does not constitute legal advice, and does not advocate for any specific regulatory outcome. Instead, it maps the existing legal architecture — its strengths, its internal tensions, and its current blind spots — with particular attention to the growing gap between technical capability and legal attribution. Key contributions include: A precise triadic model of adversarial misuse (concealment + context manipulation + harm) Behavioural and forensic signatures that enable better detection and intent reconstruction Comparative analysis of US and EU regulatory approaches to AI-mediated harm A protective architecture grounded in resilience rather than prohibition The series is directly relevant to judges, regulators, platform governance teams, corporate counsel, AI safety researchers, and legal scholars confronting the practical challenges of attribution failure in AI-assisted conduct. All documents are fully open access (CC BY 4.0) and designed to serve as a practical reference framework for those responsible for developing, enforcing, or interpreting rules governing AI systems in legally consequential domains
Oleg Zmiievskyi (Wed,) studied this question.