[pde3]

We did think about those issues before writing the piece.

It's safer to figure out how to get AI systems to be robust, reliable and safe in civilian contexts before rushing to weaponize them. We need to understand how to avoid technical problems like adversarial examples, and how to recognize and avoid accidental action-reaction-escalation pathways, before militaries start deploying this stuff.

Objectively speaking, the US is one of the planet's more belligerent nations. But if there was evidence that other belligerent countries were already deploying AI weapons systems, there might be an argument for the US keeping pace. If there isn't such evidence, the US should think more carefully about whether to move first, and how to move first, or whether certain kinds of restraint in this space might be in its long term strategic interest.