Asia-Pacific Journal Series – Can We Design Military AI to Follow IHL by Default?

Editors’ note: This post appears as part of a series highlighting articles in a recent volume of the Asia-Pacific Journal of International Law, on which presentations were made at a workshop hosted by National University of Singapore’s Centre for International Law. Introductory post by Samuel White, Danielle Yeow, and Jonathan Kwik can be found here.

Can we design artificial intelligence (AI) systems to automatically follow core principles of international humanitarian law (IHL) in similar ways to Asimov’s Three Laws of Robotics? A recent proposal from the Institute for Law & AI suggests that future agentic military AI systems should be designed to consult “AI-Guiding Treaties” and to refuse human instructions that would violate the designated treaties.

This idea of Treaty-Following AI (TFAI) is appealing for military agentic AI systems. It would effectively sidestep the long overdue responsibility gap debate regarding ex post attribution of responsibility and turn the focus toward an ex ante design duty that could prevent an AI henchman from loyally carrying out human orders without questioning their legality.

This post highlights my argument from a previously published article contending that some areas of IHL are particularly difficult to codify as legal constraints because States deliberately left their meanings ambiguous.

Why is IHL a Particularly “Hard Case” for Legal Automation?

In quantum mechanics, light can exist as both a wave and a particle, a phenomenon known as superposition. Light collapses into one state only when it is observed (the observer effect). I argue that IHL exists under similar conditions because it holds two logically irreconcilable visions at once: military necessity; and humanity. The meaning of these rules is suspended until we decide which rule wins out.

The difficulty in applying the TFAI proposal to IHL lies precisely in the fact that IHL’s fundamental provisions, such as the principles of precaution, distinction, and proportionality, can be legitimately interpreted differently according to two opposing moral reference points. The danger of TFAI is that it rests on the need to collapse the superposition ex ante to provide objective guidance to the AI, which would inevitably privilege one vision over the other as an infrastructural reality while obscuring the political nature of that choice.

Strategic Ambiguity as “Meta-Indeterminacy”

There is nothing novel in the claim that some rules of IHL are indeterminate. My article distinguishes four established forms of indeterminacy in existing accounts: (1) linguistic indeterminacy, arising from open-textured languages; (2) legal indeterminacy, arising from questions of sources, interpretation, and application (such as whether military manuals count as State practice? What counts as “direct participation in hostilities”? How many civilian casualties are considered “excessive” to the military advantage anticipated?); (3) structural indeterminacy, arising from IHL’s dual role in both legitimising and restraining violence; and (4) practice-based indeterminacy, arising from a divided IHL interpretive community between military and human rights lawyers. Each account describes a form of legal indeterminacy, but none fully explains why indeterminacies in IHL have proved so resilient.

My answer to this question is that the indeterminacies are not gaps awaiting to be filled. States actively preserve these indeterminacies as a matter of strategic choice. I used the term strategic ambiguity as the meta-indeterminacy that explains and sustains the superposition of IHL and hence the four identified indeterminacies.

This idea was built upon the more familiar concept of “constructive ambiguity,” which describes the creative use of ambiguous wording that allows States to reach an agreement they could not otherwise reach. Strategic ambiguity goes beyond that to cover tacit agreement between States not to reach any agreement at all to maintain interpretive flexibility, plausible deniability, and operational latitude.

Notable examples of strategic ambiguities include the absence of more rules governing espionage, the deliberate exclusion of the discussion on the use of nuclear weapons from the two Additional Protocols, and the definitions of engagement and deployment in the “carry arms openly” requirement for Article 44(3) of Additional Protocol I.

Apart from States, international courts also play a role in creating and/or maintaining strategic ambiguity. For example, the International Court of Justice’s 1996 advisory opinion on Nuclear Weapons declined to establish or elaborate upon an objective standard (paras. 84-85). Also demonstrative is the International Criminal Tribunal for the former Yugoslavia’s Appeal Chamber decision in Gotovina et al. to overturn the 200-metre standard that distinguished specifically-directed attack from indiscriminate attack (paras. 58-59).

Here, strategic ambiguity runs contrary to the idea of objective constraint promised by the TFAI. Further, when ambiguity is precisely the agreement itself, it calls into question the TFAI assumption that State consent to treaties could provide a more objective and legitimate anchor than aligning AI with human ethics or ideologies.

Why This Matters for TFAI?

This does not make the TFAI proposal worthless. Many of the gravest IHL violations stem not from legal ambiguity but rather from how humans interact with complex environments. While TFIA might prove difficult in assessing the legality of an airstrike against combatants embedded among civilians, it could prevent it from carrying out the massacre of civilians or prisoners of war due to human hatred, for instance. Against violations of that kind, a system that cannot be influenced by human factors offers a significant improvement.

Conclusion

Ultimately, I argue that the appropriate response is not to abandon TFAI but instead to introduce a contestability-by-design feature, in which AI would flag indeterminacy rather than resolve it, preserving the interpretative space for human military lawyers. In other words, indeterminacy in IHL is not a defect to be engineered away but a constitutive feature of a legal regime that must reconcile violence with humanity.

Thus, while some IHL constraints can be coded, others must remain contested. The task ahead for IHL scholars and lawyers is precisely to map out the exact scope and extent of IHL’s contested space.

***

Dr Sze Hong Lam (Ocean) is a postdoctoral research fellow at the Centre of International Law (CIL) at the National University of Singapore (NUS).

The views expressed are those of the author, and do not necessarily reflect the official position of the United States Military Academy, Department of the Army, or Department of Defense. 

Articles of War is a forum for professionals to share opinions and cultivate ideas. Articles of War does not screen articles to fit a particular editorial agenda, nor endorse or advocate material that is published. Authorship does not indicate affiliation with Articles of War, the Lieber Institute, or the United States Military Academy West Point.

Photo credit: U.S. Air Force Research Laboratory, Matthew C. Clouse