The Principle of Proportionality is a cornerstone of Just War Theory, yet the advent of Lethal Autonomous Weapon Systems has intensified debates regarding its application. Conventional scholarship suggests proportionality requires qualitative, subjective judgment, which algorithmic systems inherently lack. Our talk reframes the inquiry, proposing that the difficulty of encoding proportionality into machines is not a technical failure but an epistemological "mirror" reflecting the principle's fundamental conceptual instability. We proceed in two stages. First, we provide theoretical and empirical reasons to suspect the comprehensibility of the proportionality principle. Second, we analyze various machine learning paradigms, including supervised, unsupervised, and reinforcement learning, to show that the inability to formalize a stable objective function prevents machines from learning the principle (a.k.a the "alignment problem").
We argue that if a principle cannot be taught or formalized, its comprehensibility to humans must be rigorously reconsidered. Rather than revealing machine limitations, the problem of teaching Lethal Autonomous Weapon Systems proportionality highlights the inherent indeterminacy of proportionality itself.