It seems voyeuristic to read someone’s breakup text messages, but typically the parties involved in the breakup aren’t a billion-dollar AI lab and the largest agency in the federal government, so you could be forgiven for being interested in taking a look. On Tuesday, court documents including emails taken to Anthropic and the Department of Defense earlier this year were released, and they read more or less as you would expect based on the public account of the situation.
In the exchange, Anthropic CEO Dario Amodei made clear that his main concern was that the Department of Defense use your lab’s AI models for what he and the company were not comfortable with, namely integrating the technology into autonomous weapons systems and home surveillance tools. As noted above, the Pentagon’s position is that it could use the technology for “all legal uses,” which creates considerable room for maneuver.
The emails reveal that the seeds of the split were planted in January, when Defense Department undersecretary of defense for research and engineering Emil Michael contacted Amodei after radio silence for several weeks. In it, Michael said he was “hoping we come closer to committing to your revised view”; basically waiting for Anthropic to be ready to accept the Pentagon’s demands. Amodei responded by reiterating its position that there need to be safety barriers to the use of AI, including a ban on its use for fully autonomous weapons and home surveillance.
Michael said that was “simply not viable” and warned that there was “one more opportunity to align on basic principles that would lead to statutory language” before deciding to part ways. He went on to note that “there is no distinction in our world between defensive or offensive weapons” (a detail that seems noteworthy given the ongoing debate over selling “defensive” weapons to Israel), so Amodei’s attempts to make a distinction about how his company’s technology could be used would be futile.
Amodei noted that the “all legal uses” standard would not apply to Anthropic because U.S. law does allow domestic surveillance. In response to an apparent language proposal from the Pentagon, Amodei pointed out to Michael that the department appeared to “completely eliminate our red lines.” The next day, Secretary of Defense Pete Hegseth announced that his agency would be designate Anthropic as a supply chain riskwhich practically put an end to all that negotiation.
If anything, Amodei comes off pretty well in the emails, given that he stuck to his convictions. Less impressive was Michael, who went to great lengths to obtain the concessions his agency sought. But Michael already seemed pretty invested in this whole situation, given that he was sitting on a big stack of shares for anthropic competitor xAI, between other investments in AI companies. Even if getting rich was only in the back of his mind during his exchange with Amodei, at the front of his mind was figuring out how to keep the door open to using AI to kill people and surveil citizens, so he’s not exactly a guy with a strong moral compass either way.
We have risen the complete 346-page court document which you can consult here.
