Consider it a preview of strange quandaries to come.
In July, Anthropic PBC signed a deal with the Defense Department worth up to $200 million, calling it “a new chapter” in its “commitment to supporting US national security.” The Pentagon was equally keen. Eight months later, the partnership has ruptured, the contract is at risk, the White House has labeled Anthropic with its gravest epithet (“woke”) and defense officials have threatened to designate the company a “supply chain risk.”
What went wrong? Welcome to defense contracting in the artificial intelligence era.
Anthropic has been providing the Pentagon with a specialized version of its large language model, Claude, for use on classified systems. The company says it remains committed to national-security work, but it has also established “red lines” governing the use of its products: It doesn’t want Claude deployed for “mass surveillance” of Americans or as part of a fully autonomous weapons system.
Whether such stipulations were part of its contract isn’t clear. (A U.S. official has said the deal was signed “without a lot of specificity.”) But the Defense Department is livid: It says the military should be free to wield such tools for “all lawful purposes” and that Anthropic’s stated restrictions may be a battlefield liability. It also says that makers of competing LLMs — including OpenAI Inc., Alphabet Inc. and X.AI Corp. — have agreed to its terms in principle.