Congress’ favorite AI tool? ChatGPT

The House of Representatives has voted with its wallet, and the winner is OpenAI. In the twelve months ending March 31, congressional offices, committees, and institutional accounts funneled $100,580 into ChatGPT across 798 separate transactions. That represents roughly 90 percent of all identifiable AI spending — $113,740 total — according to House disbursement records analyzed by CNBC. Anthropic's Claude, the nearest competitor, collected a mere $13,160 over 37 purchases. The gap isn't competitive; it's structural.

This isn't a market test. It's a lock-in. ChatGPT arrived first, integrated fast, and became the path of least resistance for overworked staffers who need a memo summarized, a bill compared, or a constituent letter polished before lunch. The tool works well enough that nobody bothers to evaluate alternatives. When the House spends 798 times on one vendor and 37 on the next, the decision has already been made — not by policy, but by habit.

The partisan split sharpens the picture. Democratic offices accounted for $54,165 in AI purchases, triple the $15,782 spent by Republican offices. That disparity hints at deeper currents: different trust levels in generative models, different staffing cultures, different tolerance for opaque systems. Republicans may be relying more on free tiers, or on AI embedded in broader software suites that don't show up as line items. But the spending gap is real, and it means the two parties are effectively building different digital infrastructures for legislative work.

What the disbursement records hide is arguably more important than what they reveal. The CNBC analysis explicitly excludes free accounts and AI bundled into larger contracts. Every staffer with a personal ChatGPT Plus subscription, every office using Microsoft Copilot or Google Gemini through an enterprise license — none of that appears. The $113,740 floor is almost certainly a fraction of actual usage. Congress may already run on AI to a degree no ledger captures.

Staffers now use these models to write the first draft of legislation, summarize hundred-page bills, generate talking points for hearings, and compose social media posts that reach millions. The prompts are private. The outputs are unverified. The training data, the model weights, the moderation filters — all belong to a private company with no obligation to the legislative branch. When a House office asks ChatGPT to "draft a response to a veteran denied benefits," the answer shapes a citizen's experience of government. That answer bears no official seal, no audit trail, no accountability mechanism.

OpenAI has not lobbied for this monopoly. It didn't need to. The product arrived ready-made, priced low, and integrated into workflows before any procurement policy could catch up. The House's decentralized purchasing authority — each office holds its own budget — guarantees fragmentation and defaults to the easiest option. No chief information officer mandated ChatGPT. No ethics committee reviewed its use for drafting law. The adoption happened at the margins, one expense report at a time.

That's how institutional capture works. Not through conspiracy, but through convenience. The same dynamic played out with BlackBerrys, then iPhones, then Slack. Each became the default because switching costs exceeded evaluation effort. But AI isn't a phone. It produces text that becomes policy, correspondence, and public record. The stakes are different. The opacity is dangerous.

Congress could fix this tomorrow. The Committee on House Administration could require disclosure of every AI tool used in official work — free or paid, standalone or bundled. It could mandate that any generative output entering the legislative record carry a watermark or metadata tag. It could fund an open-source alternative trained on public domain legal text, giving staffers a sovereign option. It has done none of these things.

The silence is telling. Lawmakers who grill tech CEOs about algorithmic bias and data privacy have quietly outsourced their own cognitive labor to the very models they summon to hearings. They trust the black box to write their words, then pretend the authorship remains theirs. It doesn't. Every memo drafted by ChatGPT carries OpenAI's architectural choices — its safety filters, its tone calibrations, its hidden instructions. Those choices now echo through congressional offices daily.

The $100,580 is a rounding error in the federal budget. The precedent it sets is not. A legislature that leans on a proprietary model to formulate its output surrenders a slice of its sovereignty to a company answerable only to shareholders. The House doesn't need to ban ChatGPT. It needs to treat it like what it is: a contractor with no security clearance, drafting laws in the dark.