Elections · Experiment

Think Past the Number

Sick of the election and dying to know how it ends? Both can be true. Here is a prompt that gets you your own read, not one more poll dressed up as news.

A lone analyst seen from behind in a vast dim data room, dwarfed by a wall of screens full of noisy polling tables, with a single clean amber probability curve cutting through the chaos along the analyst's sightline.
One clean signal, cut from a wall of noise. Illustration generated with Gemini.

By now you hold two feelings about this election at once. Completely done with it, and quietly desperate to know how it turns out. The feeds have figured that out, which is why they keep selling you the same poll in forty different outfits. There is a better way to scratch the itch than refreshing an average at midnight. Build your own read, and stop waiting for someone to hand you theirs.

That sounds like more work than a Tuesday deserves. It took an hour with an AI, treated as a thinking partner instead of an answer machine, and that hour is the whole point.

Stop reading what they say

Pointed at the election, the model did not open with polls. It opened with a distinction. There is what people say they will do, and there is what people and institutions actually do when their own money, time, and reputation are on the line. Polls are the first kind, and they are the cheapest signal money can buy, which is exactly why your feed is drowning in them. The second kind is harder to fake, and nobody bothers to aggregate it for you.

That distinction is the whole premise of Everybody Lies, the book I teach in my marketing research class. Seth Stephens-Davidowitz shows how people tidy the truth for a survey, then confess it, anonymously, to a search bar. His rule for reading human behavior is the one the model reached for on its own: don't trust what people tell you; trust what they do.

Stated

What people say
  • Polls
  • The generic ballot
  • Favorability
  • Pundit forecasts

Cheap to make. Endlessly reheated. Already commoditized.

Revealed

What people do
  • Votes already cast
  • Ad money pulled from a race
  • Where the candidates fly
  • Committee cash moved late

Costly to fake. Nobody aggregates it for you. Where the edge hides.

So we went looking for the second kind. Votes already banked in dozens of special elections. The districts where party committees quietly canceled their ad reservations, having read their own private numbers. The states where a president chose to spend his last nights, which tells you where his own pollsters think the fight actually is. None of that is a poll. All of it is behavior.

Then came the anticlimax. Hunting for a juicy fight between the money and the polls, we mostly caught them nodding along, and boring agreement across signals that cannot copy from one another is itself a finding. The pet theory everyone drags out, that the polls secretly lowball one side, turns out to be a presidential-year habit. In the last two midterms the error hugged zero.

-20+2+4+3.02016-0.52018midterm+4.72020+0.82022midterm+2.92024overstated Dem ↑
Final national polling error, by cycle. Positive means polls overstated Democrats. The midterm years hug the line; the big misses are all presidential.

Ten thousand elections

Pressed for a prediction, I put my gut on the table: a split, one party grabbing the House while the other clings to the Senate. The model did not pat me on the head and agree. It took the opposite side cold, spelled out exactly what would have to break for me to be wrong, then offered to settle it the only honest way, by building the thing instead of trading opinions.

It ran ten thousand elections. Not ten thousand coin flips, which would have been a lie, but ten thousand runs sharing one national swing, so the races moved together the way real races do under a common tide.

00k1k4345474951535551 = majorityDemocratic Senate seats
Democratic Senate seats across ten thousand runs. The pile centers on 51, the barest majority there is, and the tails reach into clear control for either side. Fifty seats or fewer hands the chamber to the Republicans.

The center of that pile sat on the thinnest majority possible, and it spread wide enough that the other result never stopped being credible. That alone reframes the question. The Senate was never a safe call for anyone. It was a coin weighted a hair past even.

Then came the part that pays. Give the whole board one shared shove toward either side, and my split did not merely survive. It stood up as the sturdiest outcome across a wide band of plausible misses. Translation: the confident people on every network are guessing too.

0%20%40%60%80%today-4-3-2-10+1+2National miss (pts) − = toward RepublicansDem sweepRep sweepSplit
Ten thousand simulated elections. The split, one chamber each, is the most robust outcome across the plausible range of misses, and the favorite under even a small miss toward the Republicans.

The forecast it produced mattered less than what producing it revealed. My gut was not a hunch to defend. It was a hypothesis I could finally see the shape of.

Why this matters

Most people point AI at a hard question to make it smaller. Summarize this, decide that, tell me who wins. That is extraction, and it is the least interesting thing the tool does. The creative move runs the other way. You use AI to make a question bigger before you make it smaller, to assemble in an evening a way of thinking you could never have built alone in the time you had.

One conversation became a research desk, a sparring partner, and a simulation engine in sequence, and not one of those steps did the thinking for me. Each one gave the thinking somewhere to go. What you keep after Tuesday's results are called is not a prediction. It is a repeatable way to form your own view on anything a confident number pretends to have already settled. Point it at a diagnosis, a market, a career move, a marriage. The election was just the loudest excuse to try it.

The prompt

Here is the prompt that ran the whole thing. Swap the bracket for any contested call you are tired of hearing about and still want your own answer to: an election, a market, a hiring decision, a diagnosis you want a second frame on.

▸ the thinking-partner prompt
Act as a thinking partner for a question I can't settle by looking up an answer.
The question: [something genuinely uncertain and contested].

Don't hand me the consensus. Build me a way to think about it:
1. Separate what people SAY (polls, surveys, stated opinion) from what they DO
   (where money, time, and reputation actually move). List the revealed signals
   I could go check myself.
2. Hunt the divergences. Where does the behavior disagree with the stated
   consensus? Tell me which gap is real signal and which is noise.
3. Take the other side of my gut. Ask what I believe will happen, then build
   the strongest case that I'm wrong.
4. Turn it into a model. Give the key pieces a baseline and a shared swing, run
   it a few thousand times, and show me the distribution and the two or three
   outcomes that actually matter, not a single point estimate.

Ask me three questions before you start. Push back throughout.
The goal isn't your answer. It's a sharper version of my own thinking.