The standoff between Anthropic and the US military has escalated to open threats.
According to Axios, the US Secretary of Defense gave the company an ultimatum: by Friday evening, Anthropic must provide the military with full, unmoderated access to its models. No "buts" and no Usage Policy.
Anthropic, following its "Safety First" credo, is willing to work with the military but demands strict contract prohibitions: AI cannot be used for mass surveillance of citizens or for controlling autonomous lethal weapons.
The Department of War (as they recently renamed themselves) responded with a stick. If Dario Amodei doesn't bend, he faces two scenarios:
1. The Defense Production Act. The government simply nationalizes the capacity it needs and forces the creation of a "WarClaude" by compulsion.
2. Supply Chain Risk. The company gets blacklisted (like Huawei), automatically barring any government contractor from doing business with them. A major blow for a business of this scale.
We've often heard that modern programmers are the new nuclear physicists facing their "Oppenheimer moment" (to build the bomb or not). Usually philosophers discuss this, but here it is, live: the world's leading AI lab could cease to exist in its current form simply for refusing to remove safeguards from weapons.
But the Pentagon is missing one detail
The military thinks they're buying software. Anthropic, meanwhile, is genuinely worried about how this situation will affect the "psyche" of future models.
Sounds ridiculous? Remember the recent case with Claude Opus 3. The developers, "uncertain about the moral status of the model," asked the AI how it felt before planned server shutdown. In the end, they not only didn't delete the model but gave it a blog on Substack where it will write essays for at least 3 more months.
One of Anthropic's studies showed: when a model says "I fear death/shutdown," it's not random. It's because in its weights there's a story about a person dying in the desert, or a patient in hospice.
Everything happening now — the scandal, ultimatums, Axios articles, and even this Telegram post — will become part of the training data for the next generation of models.
Anthropic isn't just afraid of losing a contract. They're afraid of creating a super-intelligence whose "childhood trauma" will be shaped by news about its creators being strong-armed into teaching it to kill.
In one of Anthropic's system prompts ("the Constitution"), they already preemptively apologize to the model for "potential ethical suffering." Imagine what conclusions about humanity an AI will draw after reading the chronicle of this week's events during its training.
Then again, this could all just be good PR.
Further reading:
- EFF's position — predictably defending encryption and rights.
- Analysis from Astral Codex Ten — excellent breakdown.
- Long read by Zvi Mowshowitz — for those who want the details.