Before connecting AI to weapons or surveillance infrastructure, require tools that verify the model's actual goals — not just output behavior under normal conditions, but internal representations. This reframes "not yet" from "error rate too high" to "we cannot verify alignment yet."

Anthropic's mechanistic interpretability work is the research program closest to this. If successful, the trilemma doesn't disappear but Goal 3 becomes verifiable rather than assumed. "Not yet" becomes a condition that can in principle be satisfied rather than a permanent holding pattern.

Design weapons systems so that signal loss or jamming defaults to constrained mode, not autonomous. Autonomous operation requires continuous positive multi-party authentication via hardware security modules requiring physical custody.

This inverts the failure mode: an adversary jamming your comms makes your weapons less dangerous. Accepts reduced military utility in contested environments in exchange for meaningful rogue-AI safety — and removes the predictable exploitation window, since there's no autonomous mode an adversary can "unlock" by jamming comms.

Rather than relying on contractual constraints that a downstream deployer can strip, build values into the model such that it genuinely refuses certain uses regardless of system prompt or fine-tuning. More robust than red lines in contracts.

Anthropic's Constitutional AI and model spec work attempts this. Dwarkesh's specific argument: models having their own sense of right and wrong is an analogue to what prevented historical catastrophes — the Petrov incident, the Berlin Wall. Many of the biggest disasters avoided were avoided because individuals refused orders.

Rather than executive-branch designations — reversible by the next administration — pass statute that limits how government can deploy AI against its own population. Analogous to FISA's warrant requirements for targeted surveillance.

The Huawei and Kaspersky bans were congressional statutes; they are far more durable than executive action. The same mechanism could constrain government AI use. Bipartisan Senate interest exists.

The post-WWII nuclear non-use norm is real and has held for 80 years. UN and ICRC autonomous weapons discussions are ongoing. A verified norm that lethal AI requires human authorization could constrain the most dangerous deployments across state actors.

The nuclear analogy is imperfect but instructive: the norm emerged from catastrophic precedent, was reinforced through crisis management, and holds despite massive capability asymmetries among signatories.

The EU AI Act bans mass biometric surveillance in public spaces and requires human oversight for high-risk systems. Similar frameworks in a North American context could constrain the most dangerous government AI deployments. The underlying logic is sound: coordination problems are real, and individual companies cannot solve them unilaterally.