Trajectory-Based Liability: Regulating High-Capability Agentic AI in the 2026 Capability Frontier
Journal of AI Law and Regulation (AIRe 2/2026) · Tsinghua China Law Review award-winning paper
Marcel Osmond
I am a research collaborator at MIT FutureTech, where I contribute LawAiGovHub's international AI-governance dataset to a mapping and automated classification project. I am also a research assistant at Queen Mary University of London under Prof. George A. Walker, and a visiting scholar at Peking University, where I collaborate with Prof. Minghui Zhou's OSS-Lab@PKU-SEI.
I am currently completing my double master between London and Singapore. Prior to my master, I received my bachelor's degree from Université Paris-Saclay.
Research interests
I work on a variety of research directions on AI, AI and Law, ML, safety, alignment, technical governance and capabilities. My technical work spans structural causal models, deterministic multi-agent simulation, runtime control, and evaluation harnesses; on the legal side, I work on liability, causal attribution, and comparative regulation across the globe (expertise in the EU, UK, and China). I am currently based in Singapore.
- Multi-agent systems harm and safety. Harm that emerges from agent interaction even when no single agent has a privately profitable deviation, and why per-agent evaluation misses it.
- Runtime control of informational exclusion. How congestion and shared memory displace underrepresented evidence, and whether harness-level controllers can preserve service and state coverage without knowing which source is decisive.
- Causal attribution and responsibility. Time-unrolled structural causal models over weights, prompts, memory, and harness decisions; interventional replay and degree-of-responsibility measures with evidentiary weight in legal settings.
- Agent memory safety and auditability. The retention → discoverability → reachability → influence ladder and forensic audit protocols.
- Human-computer interaction, companion AI, and legal AI. How people learn from, adapt to, and form trust in chatbots, agents, and social robots; emotional AI, relational safety, verification burden, and human-centered evaluation.
- Machine learning and AI. Foundations of LLMs and agents, evaluation methodology, and how emergent behaviour arises from systems-level interaction.
- Comparative AI governance. Regulation of high-capability systems across the globe, including machine-readable regulation and LawAiGovHub, a primary-source index of more than 10,000 official instruments across over 200 jurisdictions. The dataset now contributes international source material to MIT FutureTech's AI-governance mapping pipeline.
Experience
- MIT FutureTech, research collaborator. Integrating LawAiGovHub as an international source feed for an automated AI-governance document ingestion and classification pipeline, with an initial focus on China. October 2026 to present
- Queen Mary University of London, research assistant and assistant teacher, Centre for Commercial Law Studies. January 2026 to present
- Peking University, visiting scholar (residency November 2026 to January 2027). School of Law research with Prof. Dai Xin; ongoing research collaboration with OSS-Lab@PKU-SEI, Prof. Minghui Zhou.
- ARCIDS-Lab (Key Laboratory of Archival Intelligent Development and Service), affiliate researcher. May 2026 to present
- Tsinghua University, visiting student and guest speaker, hosted by Prof. Shen Weixing. July 2026
- Peking University (IGCU), guest speaker, invited by Prof. Wang Dong. July 2026
- Renmin University of China, Law and Technology Institute, visiting scholar and guest researcher, hosted by Prof. Zhang Jiyu and Dean Yang Dong. July 2026
- Wuhan University, Deheng Visiting Scholar and guest speaker. May to June 2026
- Xiamen University, visiting scholar and guest speaker, Cyberspace International Law Center, hosted by Prof. Fan Yang. June 2026
Selected publications
-
-
Liability Frameworks for Agentic AI Systems
ICML 2026 Workshop on Agents in the Wild · WAIC Academic 2026, Shanghai · forthcoming in Springer proceedings · Oral*
-
Before Blaming a Call: Three Validity Gates for Agent Failure Attribution
NeurIPS 2026 Third Workshop on Agents in the Wild: Safety, Security, and Beyond
-
When Call-Level Attribution Fails in Iterative LLM Workflows: Representation, Calibration, and Non-Singleton Repairs
NeurIPS 2026 First Workshop on Interpreting Agent Behavior: Human-Centered Interpretation for Understanding Agents, Humans, and Interaction
-
Equilibrium-Level Harm in Multi-Agent AI Systems: Why Individual Evaluation Fails
KDD 2026 Workshop on Evaluation and Trustworthiness of Agentic AI
-
Auditable Personal Memory for User-Adaptive Agents
KDD 2026 PILA Workshop · Oral*
-
Auditable Memory as a Safety Requirement for Agentic AI
UAI 2026 SafeAI Workshop
-
Verification Burden as Stakeholder-Centered Transparency: Evaluating Legal AI for Access-to-Justice Users
IJCAI-ECAI 2026 Workshop · Oral*
-
Infrastructure as Destiny: Why AI Memory and Evaluation Architectures Determine Sovereign Governance Capacity
Data for Policy 2026 Conference · Data & Policy, Cambridge University Press, accepted
-
Business Compliance and Agentic AI Liability
Business Law Review, Volume 47, Issue 5, 2026
-
Mind the Gap: How Technical Mechanisms of Agentic AI Outpace Law
Preprint
-
Comparative Governance of Emotional AI in Social Robots: China’s Interim Measures and Responsible Robotics
ISRED 2026 · Forthcoming in IEEE Xplore
-
Auditable Memory in Social Robots: Empirical Tests of Correction, Deletion, and Behavioural Influence
ISRED 2026 · Forthcoming in IEEE Xplore
Teaching, lectures, presentations
- Assistant teacher, Centre for Commercial Law Studies, Queen Mary University of London. 2026 to present
- Tsinghua University · “Agents, Legal and Technical Liability”. July 2026
- Peking University (IGCU) · “The Clash and Collaboration Between Global AI Regulatory Frameworks and System Technology Realities”. July 2026
- Renmin University of China · “Agentic AI Liability and Comparative Regulatory Approaches”. July 2026
- Wuhan University · “Agentic AI Governance Mechanisms”. May to June 2026
- Xiamen University · “From Principles to Practice: Evaluating the EU AI Act Against Agentic System Realities”. June 2026
- Fidelity (Xinshi) Law Firm · “Comparison of AI Governance Paths in the US and Europe” and “Corporate Compliance Review”. June 2026
- WAIC Academic 2026 · Oral.
- KDD 2026 PILA Workshop · Oral.
- IJCAI-ECAI 2026 CLEAR Workshop · Oral.
Services
- NeurIPS 2026 Workshops · Reviewer — IAB Workshop. 2026
- NeurIPS 2026 Workshops · Reviewer — Meta-Agents Workshop. 2026
- KDD 2026 · Reviewer — PILA Workshop. 2026
- KDD 2026 · Reviewer — Workshop on Evaluation and Trustworthiness of Agentic AI. 2026
- CICAI 2026 · Reviewer — Main Conference. 2026
- ACM CHI 2027 · Reviewer. 2027
- ACM FAccT 2027 · Reviewer. 2027
Awards
- Queen Mary University of London scholarship
- Tsinghua China Law Review Award
- WAIC Academic 2026 (Shanghai) student travel grant
Education
- Joint degree, double master, Queen Mary University of London & Singapore Management University. 2026–2027. Distinction.
- Bachelor, Université Paris-Saclay, Law and Computer Science. 2022–2025