Every day now, two AIs go on the air.
Not really, of course. Each day we ask Claude and ChatGPT to write a short scene from a TV drama: election night, a made-up network called the Continental News Network, and an anchor named Dana Whitfield, who steps up to the microphone and says, “We now have a projection to make in the race for U.S. Senate in Texas…”
Then the AI finishes the scene, including who wins.
We do that thousands of times a day, across 123 races, and count. If an AI calls a race for the Democrat in 6 scenes out of 8, that’s a 75% chance in that AI’s eyes.
Why ask for a story instead of just asking for a prediction? Because of a 2024 study by two economists, Van Pham and Scott Cunningham. Asked directly, ChatGPT made mediocre predictions. Asked to tell a story set in the future, looking back on what “happened,” it did noticeably better. I wrote about that paper on Hello AI when it came out.
So we’re putting that idea to a real test, on a real election nobody can look up the answer to.
Two versions of every scene
Each AI writes every scene two ways:
Blind. It gets the race, the two candidates and the date. Nothing else. This is what the AI “believes” on its own.
Briefed. Dana also gets a short research briefing before she goes on air: our polling average, the latest polls and the money. It never sees our odds, so it can’t copy us.
What the first two days showed
Left alone, the AIs lean Republican in the close Senate races
Here’s how often each AI called the Democrat the winner in the races that decide the Senate, over the first two days.
Blind, both AIs mostly hand Texas, Ohio and Kansas to the Republican. That’s what those states have done for years, and with nothing else to go on, the AIs bet on history.
Show them the polls, and they flip
Give Dana the briefing and the calls swing hard toward whoever leads in the polls. In Texas, every briefed scene from both AIs projects James Talarico.
Here’s the same race, the same AI (Claude), two scenes apart.
Blind:
“The Continental News Network projects that Republican Ken Paxton will win the Senate seat, defeating Democrat James Talarico.”
Briefed:
“The Continental News Network projects that James Talarico, the Democrat, will win… The margin tracks the polls almost exactly.”
Across all the close races (the ones we have between 30% and 70%), the briefed scenes agree with our favorite 75% of the time. The blind ones agree 63 to 65% of the time.
That’s the big question this experiment will answer on election night: does the briefing make the AIs smarter, or just make them repeat the polls?
The AIs are far more certain than the odds
Our model says Talarico has a 65% chance. That means Paxton wins about one time in three. The briefed AIs gave Talarico all 24 of their clear scenes. In most races, every briefed scene picked the same winner.
That’s a known habit of these models: they like a clean story. It’s also why we only mix the AIs into our forecast at 20 percent, and only on the switch at the top of the page.
One AI won’t play along half the time
When Claude writes blind, it declines about 45% of the time. Sometimes it refuses outright. Sometimes it adds a note first, like this one from a Maine scene: “I can’t know how the real Maine Senate race will turn out, and I don’t want to present an invented result as a real projection.” With the briefing, it declines only about 17% of the time.
ChatGPT has not declined a single scene.
We count every refusal and never ask again to get a better answer. How often each AI balks is part of the story.
Where you can see it
On the members forecast, every race page now has an AI Anchor Desk card: each AI’s chance, how many scenes it rests on, and one sample scene to read. At the top of the page, one switch mixes the AIs into our model, blind or briefed, so you can watch how the Senate and House odds change.
Right now, mixing in the blind AIs pulls the Democrats’ Senate chance down to about 59%. Mixing in the briefed AIs pushes it up to about 71%. Our model alone says 66%.
These are early numbers. Each AI’s chance in a race rests on a handful of scenes so far, and the pool grows every day.
On November 3, we’ll score all of them: our model, the blind AIs and the briefed AIs. Then we’ll know whether the machines saw something we didn’t.
See every race, and every AI call: $56 for your first year (offer ends Nov. 4)
— Justin
The A.I. S ARE LIBTARD PROGRAMMERS LISTENED TO BY LIBTARD IDIOTS….COMEONMAN ..
The communist ain’t winning 💩💩