Banned after the bombing

Last week, the U.S. government banned Claude. Reportedly used it to bomb Iran. The Pentagon's position, apparently, was that they were on a break. People mass-uninstalled ChatGPT. Katy Perry drew a literal heart around Claude's pricing page and posted it with the caption "done." The internet decided to Get More Claude. Dario wrote an internal memo calling OpenAI's deal "safety theater" and "straight up lies." It leaked. The White House used it against him. He apologized.

That's one week. In AI governance. In 2026.

The race doesn't stop for the ethics debate. The race IS the ethics debate.

A $200 million Pentagon contract collapsed over five words.

The Department of War asked Anthropic to delete one phrase: "𝗮𝗻𝗮𝗹𝘆𝘀𝗶𝘀 𝗼𝗳 𝗯𝘂𝗹𝗸 𝗮𝗰𝗾𝘂𝗶𝗿𝗲𝗱 𝗱𝗮𝘁𝗮."

Amodei wrote that those five words were "the single line in the contract that exactly matched the scenario we were most worried about."

Anthropic said no. Hegseth gave them a 5:01 PM Friday deadline. They held. The Pentagon declared them a supply chain risk.

𝗢𝗽𝗲𝗻𝗔𝗜 𝘀𝘁𝗲𝗽𝗽𝗲𝗱 𝗶𝗻 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗲𝘃𝗲𝗻𝗶𝗻𝗴 with a contract governed by "all lawful purposes."

Which sounds principled, right up until you read Peter Wildeford's question from Sam's Twitter AMA: if the contract says "all lawful purposes" but your safety stack prevents a lawful purpose, you're in breach.

So either the safety stack has no teeth on lawful-but-objectionable uses, or OpenAI has manufactured a future contract dispute with the Pentagon.

Sam's answer was that he really wants to "use expertise to design a safe system."

That's a hope.

And Anthropic? 𝗧𝗵𝗲𝗶𝗿 𝗿𝗲𝗳𝘂𝘀𝗮𝗹 𝗺𝗮𝘆 𝗵𝗮𝘃𝗲 𝗯𝗲𝗲𝗻 𝘁𝗵𝗲 𝘄𝗼𝗿𝘀𝘁 𝗽𝗼𝘀𝘀𝗶𝗯𝗹𝗲 𝗼𝘂𝘁𝗰𝗼𝗺𝗲 𝗳𝗼𝗿 𝗮𝗰𝘁𝘂𝗮𝗹 𝘀𝗮𝗳𝗲𝘁𝘆. Claude is now in active military use, helping identify targets in Iran through Palantir's Maven Smart System, with no contractual framework and no formal red lines. The refusal prevented a conversation. It didn't prevent the use.

The uncomfortable corollary of Amodei's own memo is that a company can refuse to sign a contract and still have its product used in a war. Principled refusal and operational control are not the same thing.

𝗥𝗲𝗽. 𝗦𝗮𝗺 𝗟𝗶𝗰𝗰𝗮𝗿𝗱𝗼 𝗽𝘂𝘁 𝗶𝘁 𝗽𝗹𝗮𝗶𝗻𝗹𝘆: "There is only one problem with the Pentagon's approach. There is no law. The law is years behind the technology."

𝗕𝗼𝘁𝗵 𝗗𝗮𝗿𝗶𝗼 𝗮𝗻𝗱 𝗦𝗮𝗺 𝘄𝗮𝗻𝘁 𝘁𝗼 𝗽𝗹𝗮𝘆 𝗕𝗶𝗹𝗹𝘆 𝗝𝗼𝗲𝗹 𝗵𝗲𝗿𝗲.
We didn't start the fire, we're just trying to make it safer.

But here's the thing.

The most coherent governance signal of the week came from regular people mass-uninstalling ChatGPT, subscribing to Claude, and putting an Claude logo on an American flag.

That's not nothing.

That's just citizens doing governance the only way currently available to them.

𝗪𝗲 𝗱𝗶𝗱𝗻'𝘁 𝘀𝘁𝗮𝗿𝘁 𝘁𝗵𝗲 𝗳𝗶𝗿𝗲. But at least people are finally noticing it's burning.

Katy Perry drew a heart. It's a start.