FEC Filings: US Campaigns Paid AI Vendors $17M Since 2020, Mostly for Texting Tools

TL;DR

Nathan Sanders and Bruce Schneier ran a text match over every itemized expenditure in Federal Election Commission bulk data since 2020 and published the results in The Guardian today. They found at least $17M in disclosed spending on AI vendors across 523 federal candidates and campaigns. The chatbot giants are a rounding error in that total: OpenAI has about $50,000 from at least 80 campaigns and committees since 2024, and Anthropic went from essentially zero to at least 65 paying candidates or committees in 2026. The real money goes to political-specific vendors that automate texting and voter outreach. The whole pipeline is open on GitHub, and its data matches the article's numbers.

What the filings show

Federal campaigns must itemize who they pay and roughly why, on Schedule B of their FEC reports. Sanders and Schneier, who co-wrote the book Rewiring Democracy, matched payee names, purpose text and memo fields against a curated list of AI vendors. They then added four states (California, Colorado, Massachusetts and Washington) for a local view going back to 2022.

The headline number is at least $17M across 523 federal candidates and committees since 2020. The dataset tags each vendor as either "generative" (modern LLM, diffusion or voice-clone products) or "legacy" (pre-generative companies, or older products that bolted on AI features). By that tag, only about $756,000, roughly 4% of the total, went to generative-era vendors. The rest is texting platforms, call-time software and voter engagement tools that were selling "AI" before ChatGPT made it a household word.

ChatGPT vs Claude, sorted by party

According to the Guardian piece, at least 80 federal campaigns and committees have paid OpenAI since 2024, for about $50,000 in total, skewing slightly Republican. The Republican National Committee is the biggest single buyer at nearly $10,000. Most line items read like office overhead: ChatGPT seats for staff, subscriptions, research tools.

Anthropic built a similar footprint almost entirely inside this cycle. At least 65 candidates or committees reported paying it in 2026, with a nearly two-to-one Democrat-to-Republican ratio. The largest single Anthropic customer is still a Republican: Senator Tom Cotton's campaign, which reported more than $4,000 in its June filing.

The detail that jumps out is exclusivity. Only about 12% of campaigns or committees using either tool paid both. Campaigns pick a chatbot the way they pick a party, which is to say once, and then they stop comparing.

In the underlying dataset, Anthropic's high-confidence 2026-cycle total ($57,389) already edges past OpenAI's ($48,773). That comparison is conservative for Anthropic in one direction and generous in another: the pipeline deliberately excludes about $758,000 of ambiguous "Claude" matches, because the word has plenty of non-AI meanings, starting with people named Claude. Elon Musk's xAI barely registers: seven federal and two state-level payers, about $5,000 total, most of it from RFK Jr's 2024 presidential campaign.

Where the money actually goes

The big line items belong to vendors built for politics. Amplify.ai (the Guardian calls it AmplifAI), an automated texting platform, took $4.7M in the 2022 cycle alone, largely from Democrats including Mark Kelly, Joe Biden, Bernie Sanders and Adam Schiff. Its spending has tapered off since, and it is now owned by Triller.

Its Democratic successor looks like Daisychain, at about $300,000 in reported candidate spend this cycle, up from roughly $50,000 in 2024. More than half of this year's Daisychain money comes from one campaign, Abdul El-Sayed's Senate run in Michigan, which is a useful reminder of how concentrated these numbers are.

On the Republican side, Campaign Nucleus, tied to former Trump campaign manager Brad Parscale, draws six-figure spending from the RNC and several Trump-aligned PACs. It is displacing the older Republican texting vendor Prompt.io, which has about $375,000 in 2026 federal spend so far, down from more than $500,000 in the 2022 cycle. At the state level, an Uber-sponsored California PAC alone has spent more than $1M on Prompt.io in 2026.

Think of it like a restaurant's books. The flashy new espresso machine gets the Instagram posts, but the money goes to the linen service and the dishwasher lease: boring vendors that handle a volume job the owner never wants to think about. Campaign texting is the dishwasher.

The deepfake budget is tiny, on paper

For all the coverage of synthetic media, the disclosed spend on it is small. Six federal candidates or committees reported paying ElevenLabs, about $1,400 in total. Five reported paying Midjourney, about $1,600. Across the four states, those two firms combined had less than $100. The deepfake apocalypse, as itemized on FEC Schedule B, costs about as much as a used laptop.

The authors are clear that this is the tip of the iceberg. They cite Wesleyan Media Project data showing at least 164 political ads this cycle included AI-generated media, backed by at least $80M in ad spending. That work mostly happens at consultants, media firms and outside groups, which do not file itemized expenditure reports at this level of detail. A campaign pays its ad agency; the agency pays ElevenLabs; the FEC sees only the agency.

States are growing faster

Across California, Colorado, Massachusetts and Washington, the authors found at least $92,000 in spending clearly attributable to generative AI vendors since 2022, spread over at least 108 candidates and committees. The growth is steep: 2026 already shows about 10 times the state-level AI spending reported in all of 2024. Daisychain leads there too, followed by OpenAI and the AI prospect research tool DonorAtlas.

Caveats

• Undercounting is structural. Campaigns can pay for AI through corporate cards, staff reimbursements or consultants, and the vendor name never appears in the filing.
• Google is probably the most undercounted. The methodology notes that Gemini often rides inside an existing Google Workspace bill, with no separate line item.
• The 2026 cycle is not over. A large share of campaign spending lands in the final weeks before November.
• Categories overlap. One disbursement can match more than one vendor or use category, so category totals are not strictly additive.
• Voters are touchy. The authors cite a September 2025 Pew survey in which more than 70% of Americans said they would think less of a candidate who used AI to help write a speech, which is a strong incentive to keep AI spend buried in "office expenses."

Why builders should care

This is one of the cleanest public views of how a real, non-tech industry buys AI. The pattern is familiar from every vertical market: horizontal chatbots show up as cheap staff seats, while the budget flows to specialists who own a painful workflow (mass texting, call time, donor research) and speak the customer's language. If you are building AI tools for a vertical, the lesson is to sell the job, not the model.

The partisan sorting matters too. Daisychain on the left, Campaign Nucleus on the right, and even the chatbot choice splitting along party lines suggest that in politically coded markets, vendor identity is part of the product. That is a distribution channel and a ceiling at the same time.

The method is also reusable. The repo pulls FEC bulk files and four state disclosure systems, matches vendors from a documented pattern list with confidence tiers, and refreshes weekly through a GitHub Actions workflow. Public disclosure data is an underused market-sizing tool, and this is a working template for any regulated industry that has to itemize its spending.

Key Takeaways

• FEC filings show at least $17M paid to AI vendors by 523 federal candidates and committees since 2020.
• Only about 4% of that went to generative-era vendors; political texting and outreach tools like Amplify.ai, Campaign Nucleus and Prompt.io took most of it.
• OpenAI has about $50,000 from at least 80 payers since 2024; Anthropic reached at least 65 payers in 2026 with a nearly 2:1 Democratic lean, and only about 12% of users pay both.
• Disclosed deepfake-tool spending is tiny (ElevenLabs about $1,400, Midjourney about $1,600), while Wesleyan counts at least 164 AI-media ads backed by $80M, so the real usage sits with consultants and outside groups.
• For vertical AI builders, the budget follows the workflow, not the model, and the open pipeline on GitHub is a reusable template for sizing markets from disclosure data.

Sources: The Guardian: How American political campaigns are using AI, by Nathan Sanders and Bruce Schneier, federal-campaign-disclosure-ai (code and dashboard data), FEC bulk data, Wesleyan Media Project.