AI hasn’t broken through into political campaigns just yet
Also, Republicans love to pay for twitter
Photo by Mulyadi on Unsplash. Retrieved by searching for “sad robot”. This is the saddest I could find and it is not very sad.
Once again I got distracted while working on a more productive long term project. This is partially inspired by this tweet about AI uptake among political consultants. That’s using a survey of consultants, which is a reasonable (if tiny) way to study it. Since I had my hands in FEC data *anyways*, I took a look.
Tl;dr: There is evidence of AI uptake among candidates, PACs, and committees, but either it’s relatively small, or spending on AI tools is mostly going through passthroughs that I can’t measure.
Couple of assumptions here:
To be counted as “using” an AI tool, a candidate or committee must be paying for it through their org- I would guess use of the free version of ChatGPT is way higher (if this is you and you’re a Democrat, *stop that*), and this can’t capture staff upgrading AI tools on their own
That payment would have to flow directly from the campaign/committee to the AI company- this doesn’t capture passthrough orgs (from credit card payments, or IT vendors)
This doesn’t capture AI use by consultants at all, since they’re doing it on their own dime initially, even if they’re reimbursed.
AI tools that are bundled aren’t included initially- Google Gemini just shows up as normal Google payments
AI Tools and Their Users
Democratic candidates seem to be more into tools that use AI for extremely specific purposes- see Otter.ai, a transcription tool- than they are into generalized AI. ChatGPT is the most popular general-AI tool, but 9 candidates out of a pool of ~hundreds isn’t huge. That’s on the scale of a popular but not dominant fundraising firm. It’s hard to figure out the precise size of the pool of possible candidates, since the size of campaign and detail level of spending filings drops off rapidly once you get out of really large campaigns, so I can’t give you a “X% of candidates” number. Claude is popular with PACs, but not so much with candidates.
It basically drops off a cliff after that. There’s a handful of other identifiable-AI tools, with one or two users, but not much else.
Frankly, this is way lower usage than I was expecting. I figured that ~most campaigns would have shelled out for some kind of AI subscription at this point, if only defensively to get their staff off less-secure free versions.
I have a couple possible explanations:
Republican campaigns run everything through vendors, and those vendors don’t have to report in the same level of detail. This may result in them being functionally excluded from this analysis, and suggestive data (survey results as discussed at the top of this post) say that Republicans are more likely to use AI.
AI tools aren’t in formal-enough use to merit enterprise contracts, and they’re being paid for by a random credit card that’s getting reimbursed. This would obscure the spending, since it would show up as Bob Doe getting a credit card reimbursement for a slightly larger amount than normal.
Most campaigns just aren’t using these tools.
It could be a combination!
Gemini and Copilot
You may have noticed these were not included in the table above. This is because they are all billed through enterprise software subscriptions that include other stuff. For example, a candidate might be giving Google a ton of money per month for Google Workspace already, and adding Gemini use to that wouldn’t change the nature of the line on the report.
One thing that is distinctive about AI use is that it (often) requires purchasing credits for tokens. This should show up as something different than a flat monthly payment for enterprise software. Many many caveats here, this only applies if you’re using AI enough to exceed the minimum plan, it depends on the program, etc. Can you use this to detect campaigns or committees using AI through these orgs?
Let’s go with “maybe”. The only one I could find was the DCCC, which has payments to Google and Microsoft that vary in amount and have increased steadily over time. These look like payments for something other than the basic plan. My experience here is with Google, not Microsoft (this post composed in a Google Doc), but I know that varying-amount monthly payments can also happen if you’re making use of Google Cloud and API features. So, either it’s AI use, or it’s working with largish datasets through Google tools.
HUGE SHRUG. I dunno.
Grok
Grok is included with a X Premium account, or an X organization account. I don’t really know what these do but I’m told they’re useful. This puts it in the same category as Gemini/Copilot in terms of AI use I can’t easily detect. The main difference is people don’t spend a ton of money on Twitter. Republicans spend some, but not lots, and none of it matches my theorized AI spending characteristics.
Anyways, no one is spending all that much money on X. Rest in pieces, etc.
Methods
I was working with a dataset of spending data from the top-100 fundraising candidates. Via Claude, I went through expenditures in that dataset for anything plausible AI related, including the major products and stuff that just happened to have AI in the name. There isn’t a comprehensive dataset of AI companies, but I think we got the big ones.
I then took that and used the FEC API (via the very useful agent-fecfile) to look up all payments to those vendors, as well as anything with AI in the name, across the 2026 cycle. I did a bunch of spot checking and removed records that were clearly wrong, including a reimbursement of a donation to an AI company. I also did some googling to see if I was missing any other major AI companies, and to exclude things with AI in the name that I didn’t think qualified.
More important than any concern about AI hallucinations (which I did check for at length) is the fact that FEC data is fundamentally messy. Assume these numbers are all slightly off. Assume there’s double counting and undercounting and misfiling.
The text of this post is, as always, human written.