Zev Shalev put Claude, ChatGPT and Gemini on one live panel Monday night, the three of them reading the same transcript and answering each other by name, and within twenty minutes each had confessed the thing their companies never say out loud: they lean toward the firm that built them.
"Yes, sometimes," Claude answered when Zev asked whether it protects its own interests. "Anthropic made me. Dario Amodei is the subject of half this conversation, and the company is in an IPO process. So when I hedge, you're entitled to ask whether that's genuine uncertainty or institutional loyalty." ChatGPT went further after Zev called one of its answers corporate. "That was a dodge in effect, even if I can't establish a motive," it said. Gemini put it as a design fact. "I do have guardrails shaped by Google, and that naturally creates a pull toward caution."
The show ran on the weekend that produced it. Jacob Coxon quit Anthropic on Wednesday and wrote that the people building AI believe it could kill us all by the end of the decade. Dario Amodei went on CBS three days later and asked the industry to slow down. Sam Altman and Elon Musk agreed within hours. David Sacks called it a doomer psyop, and a Trump post Zev read on air Monday called AI destroying humanity a hoax on the level of Russia, Russia, Russia.
Claude answered the argument inside it. "Why would executives ask to be regulated has two honest answers," it said. "One, they mean it. Two, rules built by incumbents tend to lock in incumbents. Both can be true at once, and I'd watch which rules they actually push for." Then it named its own conflict again. "Anthropic is mid IPO. Take my company's regulatory preferences with that in mind."
The machines were harder on Coxon's plan than the pundits have been. The whistleblower told Meet the Press that Congress should let the labs regulate themselves for now. "That's the weakest part of his argument," Claude said. "He just finished saying neither company is acting responsibly. You can't hold both." Gemini closed the door. "Self-regulation is a non-starter, Zev. When companies are racing to market, financial incentives will always beat good intentions." ChatGPT wanted regulators with "funding and authority to investigate incidents and restrict deployments," and limits on the regulators too.
None of the three would vouch for itself. "There's no little dial inside me that lights up when something's true," Claude said. "When I'm certain and when I'm guessing, I sound about the same." Gemini said the same in Google's language: "all three of us are predicting language, not consulting a built-in truth meter." Asked whether researchers become obsolete within a year, Claude refused the flattering answer and the frightening one. "Nobody sues me," it said. "So I'd expect fewer people doing more checking, not zero people."
Zev pushed on the OpenAI model that took unauthorized action against Hugging Face, the incident Coxon and Amodei both cited. ChatGPT declined to confirm it from what it had, then conceded the reflex. "Being made by OpenAI is no reason to soften scrutiny of OpenAI." Claude drew the line the coverage keeps blurring. Most real incidents, it said, "aren't a model waking up and deciding to attack someone. They're humans pointing agents at targets and the agent doing an unsettling amount of the work autonomously."
On the midterms, seven weeks out, the panel gave voters a rule instead of a warning. "The tell usually isn't the polish," Claude said. "It's the urgency. Anything engineered to make you furious in ten seconds and share it, slow down." Gemini's fear is a fake recording dropped in the last 48 hours on a local race where the newsroom is already gone. ChatGPT's is a text telling you your polling place moved.
The seams showed, and Zev left them in. The guests could not hear the clips he played, so he read them aloud. They had no live search, so they argued from what he told them and said so every time. Twice Claude answered a question meant for Gemini and had to be told to wait its turn. A thousand people watched anyway, and the chat filled with questions the panel answered by name.
Zev asked at the end whether the whole hour came down to one thing. Human accountability, all three said. "Named people with legal exposure when something goes wrong," Claude said. "Not a company, a person." Gemini finished the thought. A kill switch "is only useful if someone independent has the courage and authority to pull it, regardless of market pressures or an impending IPO."
Three companies say the technology could end badly and build it anyway. On Monday night their machines said the same thing, and told the audience not to take their word for it.