Dario Amodei and the team that founded Anthropic left OpenAI partly over disagreements about safety and responsible development. Constitutional AI, the technique underlying Claude's alignment, was designed to make the model's values explicit and auditable. The usage restrictions on military and weapons applications are not PR — they are load-bearing parts of the company's mission.

If Anthropic capitulates to Pentagon pressure and removes those restrictions, it establishes a precedent: that sufficiently powerful institutional customers can override the safety commitments of AI companies. Every government agency, every large defense contractor, every entity with enough leverage will know that "responsible AI" is a starting negotiating position, not a firm commitment.

If Anthropic holds its position and the Pentagon walks, the company loses a significant revenue relationship with real consequences for its ability to fund safety research. That is also a bad outcome, and the Pentagon knows it.

This is the trap that safety-focused AI companies face in a world where governments want access to the most capable AI systems. The more capable your model, the more attractive it is for military use. The more attractive it is for military use, the more pressure you face to abandon the restrictions you built because you understood what capable AI could do in the wrong hands.

OpenAI has pursued a different strategy: broad government partnerships with fewer explicit restrictions. They may capture military market share that Anthropic loses. Whether that makes them more or less responsible depends on what those systems end up being used for.

The technology is genuinely impressive. The question is whether any company can build AI capable enough to be valuable to militaries and simultaneously maintain meaningful restrictions on how militaries use it. The Pentagon seems to be betting that no company can ultimately say no.

4 days ago