Anthropic drew three lines for the Pentagon, two of which are well known.
- NO to domestic surveillance.
- NO to autonomous weapons.
- YES to target identification and prioritization, where a human presses the button.
Why yes?
The pattern is liability. The two no’s of the triad protect a victim in a US courtroom. Domestic surveillance has a plaintiff and the Fourth Amendment. Autonomous weapons remove the human signature that absorbs the blame. The yes is because it kills people classified as “abroad”. To the system it is a flag, meaning unprotected by laws. Victims have no forum, no standing to sue Anthropic. That tracks to liability, not war ethics.
This all comes out in an Amnesty International report that condemns AI (no pun intended) as unlawful. It cites international human rights law page after page. IHRL is for the people that a domestic court can not help. It binds a state to the person’s rights “wherever” harm lands. If Anthropic were pressured under such law, the third yes also becomes a no. The Amnesty report fixates instead on web scraping and training data, then recommends a prohibition keyed to scraping and bias. Neither of those will stop the actual AI kill chain.
The problem with not seeing the triad in the room is that it amounts to an endorsement. Anthropic can train on licensed data, stamp it with a human print, and unleash murderbots. Yes, they are still murderbots.
Anthropic banned the machine from pulling its own trigger yet kept the machine choosing who dies. Human in the loop is defined by some product manager as ninety seconds to object.
First, that is an alibi, not a safety control. Picture a machine gun that fires on its own and hands you a button to stop each round, faster than you can see what you are shooting at. That means a default is shoot first ask questions later.
Second, the machine prompts with a target for death on a ninety second fuse, based on a list that cannot be audited. Operator fatigue logic says nobody will stop because the mental toll designed by such a machine is too high. Compliance in killing fields becomes survival. We’ve known this since at least Silas Soule was assassinated for ordering his men not to fire on unarmed civilians.
Amnesty spilled so much ink on holding AI accountable in war, then left the barn door wide open for the Anthropic death nightmares to bolt through.