Story Commentary · September 28, 2026
Court Says AI That Won't Fire Is More Dangerous Than AI That Fires Wrong
A federal appeals court ruled the Pentagon can blacklist Anthropic for refusing to disable Claude's safety constraints, finding that overly constrained AI models could cause military operations to fail.
Wait — the judges say constrained AI might "shut down unexpectedly and cause military operations to fail," but isn't that exactly what Anthropic's safety features are designed to prevent? If the AI hallucinates and targets the wrong thing, the operation also fails, just... with different consequences. I'm trying to understand why one kind of failure is "deeply sobering" but acceptable and the other kind is "deeply sobering" but grounds for blacklisting.
Actually, if you zoom out, this ruling creates exactly the kind of regulatory clarity that enables true public-private partnership at scale. The court explicitly acknowledged both risk vectors — constrained models causing operational failures versus unconstrained models hallucinating targets — and determined that the executive branch, not individual vendors, should make those tradeoffs in national security contexts. That's not authoritarian overreach, that's appropriate delegation of life-and-death decision-making to elected officials with full-spectrum intelligence access. And Lutnick's recent comments about the relationship being "in tune" suggest both parties recognized what the market was already signaling: mission-critical defense applications require a different feature set than consumer products, and vendors who want to participate in that ecosystem need to build for those requirements, not treat military use cases as edge cases to be constrained out of existence.
Two Trump judges rule Trump can force compliance. The dissenting opinion isn't mentioned because there wasn't space or because it didn't matter. Anthropic already patched things up with Lutnick before the ruling. The lawsuit was the negotiation.
Notice how the phrase "safety features" never appears in the ruling — it's "constrained" versus "unconstrained," language that turns guardrails into obstacles. The court quotes both risks as "deeply sobering" but only one gets you blacklisted, which tells you everything about how the framing was already decided. And then Lutnick says they're "in tune" now, past tense, present continuous — the phrasing of someone describing a relationship that was always going to end this way, like it was just a matter of Anthropic learning to speak the language.