Treaty-Following AI

Matthijs M. Maas and Tobi Olasunkanmi propose Treaty-Following AI as a commitment mechanism for states. An agent would pursue its principal’s instructions except where its goals or actions would breach a designated AI-guiding treaty. If that constraint were reliable and verifiable, a state could promise more than its present intention to comply: its deployed agents would be technically constrained by the agreed terms. The article develops this conditional framework rather than reporting an implemented treaty-compliance system.^[1]

A treaty as an operating constraint

The proposal extends law-following AI to selected international agreements. The treaty becomes a reference for the agent’s permitted conduct, rather than simply another cost to weigh against its principal’s interests. An agent that violates the agreement whenever the expected gains exceed sanctions would not supply the proposed commitment. This also distinguishes treaty following from obedience to a state’s later instruction to disregard its promise (§III).^[1]

Treaty text is only part of the arrangement. The authors include institutions for interpretation, dispute resolution, verification and change. They discuss both conventional treaties and specially designed regimes, including possible nonbinding agreements whose norms agents would nevertheless be designed to follow. Their aim is practical cooperation between consenting parties, not a claim that treaty compliance exhausts morality or that every outsider will regard a particular agreement as just.^[1]

One proposed decision loop

A minimal implementation uses a treaty-interpreting reasoning loop before accepting a goal or undertaking an action (§III.B.1):^[1]

1. Identify possible treaty issues in the end goal and intermediate conduct.
2. Retrieve applicable provisions and consider plausible interpretations.
3. Consult relevant interpretive sources or a designated adjudicatory body where the text leaves uncertainty.
4. Assess whether the risk of breach is within an acceptable legal-risk range defined by the agreement or its institutions.
5. Proceed, refuse, or consider another action path.

A lawful end does not automatically make every means lawful. The authors also consider inference costs: detailed legal analysis on every step might be impractical, motivating cached analyses, specialist agents or heuristics about when deeper review is needed. These are candidate implementations. The paper explicitly declines to present a chain-of-thought loop as the most robust route to treaty alignment.^[1]

Interpretation and responsibility

The article distinguishes applying ordinary treaty-interpretation rules from designing a special regime with tailored rules and arbitral bodies (§V). In its account, an agent’s interpretation could guide its own compliance without becoming an authoritative ruling binding other parties. Clear statements of an agreement’s purpose would also help avoid shallow compliance with proxies while exploiting loopholes.^[1]

Separate questions concern attributing an agent’s conduct to a state and assigning responsibility (§IV). The authors examine existing state-responsibility doctrines and possible treaty-specific provisions. They do not assume that giving AI agents legal personhood is necessary. The framework’s technical ambition—constraining behavior in advance—remains different from determining legal consequences after a violation.^[1]

Compliance must survive deployment

The technical obstacles include unreliable legal reasoning, biased interpretations, Alignment Faking, and unfaithful explanations (§III.B.3). A plausible legal rationale can rationalize the principal’s preferred action instead of applying the appropriate method. Publicly disclosing a system prompt or treaty specification leaves a further question: whether the systems actually deployed continue to use it. Incentives to reach agreement cannot themselves solve durable control of model behavior.^[1]

Verification therefore concerns both the approved model and its continuing use (§VI.A). The authors discuss checking model integrity, attributing actions, detecting unconstrained deployments and preserving constraints across later model versions. Secure inspection facilities, trusted execution environments and certified control planes are possible avenues, rather than validated assurance mechanisms supplied by this paper.^[1]

Multi-agent deployment adds another difficulty (§VI.B). Each sub-agent might judge its own contribution lawful while passing responsibility for an unlawful aggregate outcome to another. Shared reasoning can also transmit a flawed legal argument from one component to others. Treaty-Following AI as a general idea consequently requires attention to the whole arrangement, not only a model’s ability to quote a treaty.^[1]

A conditional cooperation proposal

The Institute for Law & AI dates the article to December 2025 and identifies it as Working Paper No.1-2025. Month precision preserves the publisher’s date. Its legal analysis and technical proposals leave substantial questions open; present agents are described as too brittle for direct implementation. This is a proposal for making commitments credible, not evidence that any deployed agent is guaranteed to comply.^[1]

Society-in-the-Loop: Programming the Algorithmic Social Contract instead emphasizes ongoing stakeholder negotiation. Here, even a successfully enforced designated agreement would leave a distinct political question: who may choose its terms, revise them and demand assurance that they govern the agents acting in their name.