Treaty-Following AI

Matthijs M. Maas and Tobi Olasunkanmi’s December 2025 article proposes agents that follow their principals’ instructions except where those would breach a designated AI-guiding treaty.^[1]

What it adds

Treaty-Following AI makes an international obligation part of an agent’s operating constraints. If feasible, that could help states make credible commitments about deployed agents.^[1]

The proposed implementation checks planned goals and actions against treaty provisions before proceeding. The authors identify fragile legal reasoning, unfaithful reasoning traces, Alignment Faking and verification across deployments as obstacles.^[1]

“the TFAI framework remains dependent on further technical research into embedding more durable controls on model behaviour”

— Section III.B.3(a). This qualification limits the proposal’s promise of reliable compliance.^[1]

What remains open

Legal interpretation, technical verification, political acceptance, and enforcement remain unresolved. The article proposes a framework, not an implemented system guaranteeing compliance. Compare Society-in-the-Loop: Programming the Algorithmic Social Contract for a different approach to collective governance.

Evidence inspected

Review covered the original web article’s abstract, definitions, caveats, implementation and technical challenges, deployment-verification discussion, and conclusion. Its legal authorities and every supporting study were not independently audited.

Historical context

The Institute for Law & AI dates Maas and Olasunkanmi’s research article to December 2025. Month precision preserves the date provided by the original record.^[1]

See Treaty-Following AI.