Reviewing Scott Alexander’s article on Astral Codex Ten.
What Happened
The Pentagon signed a contract with Anthropic last summer that included usage restrictions — guardrails preventing certain military applications like mass surveillance and autonomous weapons. In January, the Pentagon demanded renegotiation to remove these restrictions, seeking “all lawful purposes” access. When Anthropic refused, the Pentagon escalated with three threats:
- Simple contract cancellation — the most conventional response.
- Invoking the Defense Production Act — forcing compliance through wartime-era legislation.
- Designating Anthropic as a “supply chain risk” — which would effectively ban any US company using Anthropic products from government contracts. A potential death sentence for the business.
Why This Matters
The escalation from a routine contract dispute to existential business threats is unprecedented. Alexander argues this is government coercion at a scale we haven’t seen applied to a domestic tech company, comparing it unfavorably to authoritarian behavior.
The “supply chain risk” designation was designed to counter foreign espionage threats (like Huawei). Repurposing it as contract negotiation leverage sets a dangerous precedent — it’s the equivalent of bringing a nuclear weapon to a salary negotiation.
Alexander also notes the irony: the Pentagon could simply use OpenAI, Google, or other vendors who might be more accommodating, rather than trying to force one specific company into compliance. The coercion seems designed to establish a principle rather than solve a practical problem.
Our Take
This connects directly to prediction markets. The Manifold market “Will Anthropic give the military unfettered access to Claude by Friday night?” was trading at 9% when we last checked — most bettors expect Anthropic to hold firm.
Alexander’s piece reinforces this: Anthropic’s willingness to resist Pentagon pressure at significant business risk is strong evidence that their safety commitments are genuine rather than performative.
The outcome of this standoff is a significant data point for the broader question of whether voluntary AI safety commitments survive contact with powerful institutional pressure. If Anthropic caves, it undermines the credibility of voluntary safety commitments industry-wide. If they hold, it demonstrates that principled AI governance can survive real-world stress tests — though at potentially enormous cost.
We’re watching this one closely.