The Voter File Doesn't Have a Party

The Voter File Doesn't Have a Party

The industry has sorted its AI models into red and blue jerseys. ChatGPT is the Republican tool: 74 percent of GOP practitioners call it their most-used model. Claude is the Democratic tool: 82 percent of Democrats pick it first. Grok runs 45 percent on the right and 4 percent on the left. The DNC went further and barred its staff from ChatGPT and Claude entirely — Gemini is the approved model, in part because it integrates with the committee's existing stack.

Read that as an operations question and it should bother you. Campaigns are choosing infrastructure the way voters choose cable news. The model your team uses to parse a county reject file is being selected by tribal signal.

The voter file doesn't have a party. Neither does the work.

Where the model actually sits

Strip the punditry away and look at where a language model earns its keep in a turnout program. It is not on a debate stage. It fills seats — back-office seats, the kind nobody writes about.

The extraction seat: county files arrive as PDFs, scans, and forty different spreadsheet dialects. Something has to turn them into structured records before anyone can act on them. The matching seat: those records have to resolve against your file — this returned ballot is that voter, at this address, with this phone. The analysis seat: somebody has to write and maintain the code that ranks the work queue as new data lands. The drafting seat: per-county instructions, per-status variants, call scripts — drafted by machine, closed by a human, never shipped raw. And the QA seat: checking the other four.

Here is the thing about those seats. None of them ever asks the model its opinion. A model can hold whatever worldview its training gave it and still be the best extraction engine on the market, or the worst. The failure modes are boring and specific: hallucinated rows, missed matches, brittle code, a drafted instruction that cites the wrong county's deadline. Different models fail differently at different tasks, and the rankings shift with every release cycle.

Which means choosing one brand for every seat — and choosing it by jersey color — guarantees exactly one outcome: you never benchmark half the market.

The map doesn't even hold

If the tribal sorting reflected measured behavior, you could at least call it a crude heuristic. It doesn't.

On June 24, The Washington Post published a test of the major models against a battery of political questions designed by researchers at Dartmouth and Stanford. ChatGPT — the Republican industry's most-used tool — answered 80 percent of queries with only left-leaning arguments and offered a right-leaning-only answer 3 percent of the time. Grok, marketed to the right, cited left-leaning arguments more often than right-leaning ones. And the most balanced model in the test — offering both sides in over 90 percent of its answers — was Gemini. The one the DNC allows.

So the party that banned two models kept the most even-handed one, and the party that adopted ChatGPT as its workhorse is running the model that tested most one-sided against it. The sorting doesn't track the testing. Which tells you what the sorting is: a signal, not a decision. Nobody ran the test. Everyone read the room.

There's a second layer to the same survey data that gets less attention. Republicans aren't just picking different models — they're using them harder in some lanes and softer in others. GOP practitioners report using AI for image, video, and audio generation at more than double the Democratic rate, 61 percent to 25 percent. Content is where our side has planted the flag. The seats above — extraction, matching, analysis — belong to whoever bothers to sit in them.

What the build logs say

I came at this as a non-engineer. When I built Campaign Compass, the SaaS I shipped without a programming background, the build logs filled up with a pattern I didn't plan: the model roster kept changing by task. The model that wrote the cleanest code was not the model I trusted to read a long, ugly file without dropping records. The model that drafted the best plain-English instruction was not the cheapest one to run a thousand times overnight. I stopped asking "which model do we use" and started asking "which model holds this seat" — and the answer moved every few months as new releases landed.

The method that fell out of that is a bake-off, and it costs an afternoon. Take one real task from your own program — say, last cycle's reject file from your hardest county. Pull the same hundred records. Run three models against the same prompt. Count the errors by hand. Note the cost and the time. Now you have a number instead of a jersey. Repeat per seat, not per cycle — per release.

That's not a research program. It's the same discipline we'd apply to any vendor: diagnostic first, no tactic before the premise is proven. A model is a vendor. Treat it like one.

The margin math

Now the budget question, because this is ultimately a buying decision.

Run the arithmetic on a hypothetical 40,000-voter chase universe — mid-sized, nothing exotic. If two models differ by two points of accuracy in the matching seat, that is 800 voters mis-resolved: ballots credited to the wrong record, phones attached to the wrong household, doors knocked for people who already voted or letters mailed to people who never requested a ballot. Every one of those errors is money spent against a fiction, and most of them are invisible — the program "ran," the reports filled in, the waste never surfaced. Turnout programs live or die on numbers smaller than 800.

Two points is not a claim about any particular model. It's an illustration of what's at stake in a decision most campaigns are currently making on vibes. The spread between the best and worst model on your specific files, in your specific seats, is an empirical question — and it's answerable for the cost of one staffer's afternoon.

So the buying question for any AI line item this cycle is not "whose model is it." It's this: which seat is this model filling, and what did it score in that seat, on our data? If the vendor can't answer, the tool isn't a capability. It's merch.

The side that treats models as vendors — benchmarked, seat by seat, re-tested every release — will quietly out-execute the side that treats them as team flags. That edge won't show up in a headline. It will show up where the margins live: in the reject file that got parsed right, the match that got made, the ballot that got chased on time.