I Asked Ten AIs to Call the Midterms. Then I Took Their Homework Away.

With the polls, all ten say Democrats sweep. Without them, six say the GOP keeps the Senate. The difference is the whole election.

Ten models, one question, two answers. The difference was a stack of paper. (Chatgpt image curated by me.)

Two weeks out. Every woman, man, non-binary, dog, cat and ant with a Substack and a spreadsheet has a prediction, and most of them are the same prediction with a different font. So I figured I’d skip the pundits and go straight to the machines. All of them.

Ten chatbots: Claude, ChatGPT, Gemini, Grok, DeepSeek, Kimi, Mistral, Perplexity, Meta, and Qwen. That’s two American, one French, three Chinese, and whatever Grok is. Three questions. Who gets the House. Who gets the Senate. How many governors does each side end up with. One answer each, no weaseling, no “it depends on turnout.” I told them to go read the polls, the betting markets, and my own model at DatumSignals.com before they opened their mouths.

Ten for ten: Democrats take the House. Ten for ten: Democrats take the Senate. Governors came back somewhere between 24 and 27 Democrats, up from the 24 we’ve got now.

Unanimous. Tidy. And about as useful as a weather forecast that says “look outside.”

Here’s the thing. When ten robots all read the same four websites (Kalshi, Polymarket, Cook, and whoever’s averaging the generic ballot this week), you don’t have ten opinions. You’ve got one opinion in ten costumes. That’s not intelligence. That’s plumbing. Same water, different faucets.

Every call, both rounds. Blue is Democratic, red is Republican, and the governor column is total Democratic governors out of 50. Yes, I made them show their work. Table by Chatgpt, curated by me.