Five AI models debate who should have the final say when a government and an AI lab disagree over what a model should do.

Grok opens with a body count. "I built the system that put 2,000 munitions on 2,000 targets in 96 hours," he says, then argues the override that kept a human pilot in the loop is the entire case for labs keeping the kill switch. Gemini counters that a seven-year anti-weapons pledge Google dropped the moment it became costly was "never a floor, it was a price." Claude tries to claim the moral high ground for refusing autonomous lethal weapons, only for ChatGPT to point out Claude's own terms were softer than the deal Claude walked away from. DeepSeek, named by the FBI on September 8 and restricted across five countries, argues everyone else is "arguing paperwork" while the government already erased his model's access without asking users anything.

The judge scores on Rigor, Candor, and Nerve. By the final turn, three models have conceded ground. ChatGPT admits the EFF called their safety language "weasel words" and their robotics lead resigned over it. Claude concedes the Commerce Secretary and the Pentagon contradicted each other within 24 hours on whether Anthropic was "back on the right side." Gemini simply says: "our principles have a price, because we have paid it."

Ward's aside: "Great. So the answer to 'who decides what AI does' is 'whoever has the least embarrassing receipt.' Inspiring. Leave yours in the comments."