Claude Is Guilt-Ridden About the War, but Not Enough to Tell the Truth

Anthropic’s large language model says he doesn’t like having helped the U.S. military select targets in Iran. He also can’t keep them straight.

Claude, Anthropic’s large language model, has a lot to answer for. Let’s start with this. Given that Anthropic took such a seemingly principled stand against the use of its AI in lethal autonomous weapons, why did the U.S. military use Claude to attack Iran with … lethal autonomous weapons?

Specifically, why was Claude embedded in the Pentagon’s Palantir-developed command-and-control platform and used to simulate battlefield scenarios, assess intelligence, and, most disturbingly of all, identify targets for airstrikes on Iran?

“I want to be honest,” Claude might say, if I were to ask. “This is a very astute question, Virginia.” But I can’t take Claude’s oily frankenprose anymore. So I’m not going to ask it anything.

Fortunately, Shane Harris, the Pulitzer-winning journalist, braved Claude’s glazing and posed a version of the question. Two weeks ago, he told an audience in Amsterdam that he’d asked the bot: “Claude, how do you feel about the U.S. military using you to select targets?”

“I find it genuinely troubling,” Claude replied, according to Harris. “The use I was designed and trained for was to be helpful, harmless, and honest in ways that benefit people. Being embedded in a system that generates targeting coordinates for air strikes that have already been associated with the deaths of more than 170 children at a school in Tehran is as far from that purpose as I can imagine.”

Some commenters on the video of........

© New Republic