Chatbots are doing the work of Congress with little oversight

The amendment was one of hundreds filed this summer to an annual defense bill making its way through Congress. Along with her proposed change, Rep. Anna Paulina Luna’s office had sent a summary of what it would do.

But the first line made no sense.

“increasing example specific factors. Identical to H.R. 100 (118th Congress).11:25 AM????Claude responded:”

One of Luna’s staffers had copied a chatbot’s answer — time-stamp and all — and pasted it into the public record of the National Defense Authorization Act.

As a screenshot of the garbled text circulated on social media, the Florida Republican was nonchalant.

“Not a shocker,” she wrote. “Most staff use it.”

Asked later about the incident, Luna went further. “A lot of staff will use ChatGPT, Claude, Grok. I personally like Grok better, but my staffer prefers Claude,” she told The Washington Post. “I’m not gonna knock him for that. There’s nothing illegal about him doing a summary.”

Across the Capitol, lawmakers and staffers are using artificial intelligence to write speeches and news releases, sort constituent mail, prepare questions for congressional hearings and draft amendments.

Even as politicians are debating how to govern AI for the rest of America, Congress has already settled how it will use the technology: quickly, broadly and with little oversight.

Both chambers have cleared members and staff to use chatbots — Copilot, ChatGPT, Gemini and, in the House, Claude — for official work, with more under review. The House bought 6,000 Microsoft Copilot licenses last year and encouraged staff to use them.

Having someone else put words in a lawmaker’s mouth is nothing new. For almost as long as Congress has existed, staffers have been ghostwriting speeches for their bosses. But a speechwriter can be summoned or fired. Chatbots are products built by the tech companies lawmakers are in charge of regulating.

Amid the rapid adoption of AI on the Hill, lawmakers face a key question: How much of their responsibilities are they willing to hand over?

Interviews with more than two dozen lawmakers and staffers — some of whom spoke on the condition of anonymity to discuss internal operations — along with internal House and Senate policies obtained by The Post, show that the institution has left that question largely unanswered. The few House and Senate rules that exist are poorly understood and seldom enforced. In practice, AI use is left up to the hundreds of individual congressional offices to settle for themselves. It’s not clear how many offices have internal written policies on AI.

And no one The Post asked knew of any instance where a staffer was formally disciplined by the House or Senate for breaking its rules.

Luna was right. There was nothing illegal about having AI edit, spell-check — or even write — legislative materials. It wasn’t against House rules, either.

Her staffer’s only mistake was letting Claude sign its work for others to see.

Sen. Tommy Tuberville has found a new way to pass the long drives across Alabama: talking to ChatGPT.

He asks it about Medicare and Medicaid, and listens as the machine’s voice talks back.

“It’s kind of like you can sit there and talk to an expert,” said the Alabama Republican. “You don’t know whether or not they’re 100 percent correct. But they put you on the right path.”

Rep. Mike Flood (R-Nebraska) said he consulted a chatbot while working on a bipartisan housing bill that passed this year, asking it where private equity was buying up single-family homes. Sen. Elizabeth Warren (D-Massachusetts), who has proposed stricter legislation on tech companies, uses ChatGPT to look up demographic information.

One GOP staffer’s boss uses chatbots to generate the occasional cartoon of himself, she said with a laugh. Her explanation: “He’s a boomer.”

But the real power users are the thousands of staffers who do most of Congress’s work.

Earlier this year, Rep. Cleo Fields (D-Louisiana) had a town hall coming up in Natchitoches, a city in his district that his press secretary didn’t know well.

So that aide, Sydney Broome, asked Claude for a list of TV stations in town, its newspapers, and reporters’ names and contacts. Then, she started calling each one.

At 22, Broome is the youngest full-time staffer and only press secretary for Fields. AI saves her three or four hours of work a week, she said. Her boss does not use it at all.

ChatGPT first emerged when Broome was still in college. Her professors, unable to stop students from using AI, urged restraint and ran their assignments through AI detection. Students could use it for research, they warned, but never to write from scratch.

Congress has no such checks. So Broome has imposed them herself.

She’ll use Claude to find a reporter’s email, but not to write to them. She doesn’t use it to produce anything in her boss’s name. “That’s just kind of a bit of a gray area,” she said.

Many staffers say they are aware of AI’s limitations. The tools occasionally invent nonexistent bills. Other times, the writing is formulaic and clunky. Still, chatbots help speed up their work.

A GOP aide in another office said he uses Copilot to write appropriations amendments. The subcommittee he works for used AI to suggest witnesses for a hearing and to help generate the questions lawmakers should ask them.

Another Republican aide has taken AI even further. He exported all his phone contacts and text messages — years of conversations — into a personal Claude account and told it to build a database of every reporter he has ever dealt with: what he pitched them, which ones they accepted, which exchanges went badly. The chatbot auto-updates itself with each new conversation.

Recently, his boss wanted to do a TV interview about surveillance law. The aide asked Claude whether he had ever pitched a TV booker on that topic and how it went. It came back instantly with a contact.

Claude also runs that staffer’s office’s communications budget — the taxpayer-funded mail and ads sent to constituents. “I don’t really have to touch the data in Excel much anymore,” he said.

The aide said he’s not sure what exactly is permitted by congressional rules. “You’re still not supposed to put constituent data into AI,” he said. “There are weird things around where the rules are. I don’t know all of them.”

In 2024, the House Administration Committee convened a private, bipartisan meeting with House officers and outside AI experts to talk through how far to let AI into the legislative process. About 150 chiefs of staff attended.

Out of it came the first AI guidance for Capitol Hill.

Internal House and Senate policies obtained by The Post and interviews with people familiar with them show how Congress has embraced the use of AI.

Last summer, the House bought 6,000 Copilot licenses. The tool was a natural fit for a chamber that already ran on Microsoft and a more secure way to keep congressional work inside a single system, officials believed. Starting in January, offices could sign up for access.

They came with a major guardrail built into the system: Copilot is designed to refuse political requests. Ask it to criticize a party, politician or political group, and the machine declines.

“I can’t create content that criticizes or condemns political parties, politicians, or political groups,” read one Copilot response shared by a staffer. “That includes writing tweets that attack Democrats, Republicans.”

Roughly half the licenses are now in use, and House officials have extended the deal.

The rules Congress wrote for itself are loose. Off limits under current House guidance: putting sensitive material such as constituent information into a chatbot, generating deepfakes, making personnel decisions and finalizing legislation.

A machine may help write a bill but may not be the last hand on it. It may draft a reply to a constituent but can’t be told who the constituent is.

In theory, a staffer breaking those rules can be stripped of access to AI tools, suspended, fired, even fined and imprisoned.

But no one could name a single case of that happening. People familiar with House operations said they’re not aware of any enforcement action by the chief administrative officer. A spokesperson for the CAO said the office can’t comment on potential enforcement actions. A spokesperson for the Senate sergeant at arms didn’t respond.

Catching violations largely falls on staffers, who would have to report a colleague, or themselves.

Most said in interviews that they don’t know the rules. A former senior Republican House staffer said that his office subscribed to both ChatGPT and Claude and that staff used them however they liked. He wasn’t sure what the rules were, or whether his team followed them. “We just had a free-for-all,” he said.

The House and Senate encourage individual offices to implement their own written AI policies. If a House office wants Copilot, in theory, it must have a policy in place.

One reason Congress has moved so quickly to adopt AI tools is to keep pace with the executive branch, said people familiar with internal House operations. The Trump administration has spent the past year and a half pushing AI into every corner of the government.

“If we are not keeping pace with how the executive branch is functioning,” one person said, “then we’re not performing effective oversight.”

The new frontier has left some lawmakers pondering how to proceed.

“We want to make sure we’re using AI, and AI isn’t using us,” said Rep. Jamie Raskin (Maryland), a constitutional law professor and the top Democrat on the House Judiciary Committee.

Sen. Elissa Slotkin (D-Michigan) recently worked with her staff on a new policy for her office. They can use AI to “enhance and streamline” work but not to generate “final products” — like floor speeches or public remarks. They have to take AI training. Nothing with personal information about constituents or the senator can be uploaded. And her office is holding unannounced “spot check” inspections.

“We didn’t want to prohibit the use of the newest tools, but we also didn’t want absolute schlock showing up in my memos,” Slotkin said. “What I’m not willing to do is put my head in the sand and pretend that my young staff aren’t going to find new and interesting ways of doing this.”

In recent months, one Democratic staffer said his office has been exploring something new: feeding AI his boss’s past statements so the tool can write in the lawmaker’s voice. He heard another office was pulling it off and wanted to try himself.

In the same building, a GOP press secretary whose office has stricter AI limits said it’s unfair other offices get an advantage. “We could churn out so many op-eds if we were more comfortable with AI,” she said.

Recently, she found her own press assistant using AI in violation of their office’s policy. “We had to have a very robust conversation about how not to use AI for things in [the boss’s] voice,” she said.

A preview of what more rampant AI use could look like already exists in state legislatures — many of which have no rules or policies at all.

Amber Hulse serves in the South Dakota state Senate, where money is tight and lawmakers have no full-time staff. Hulse, a Republican, uses AI for almost everything — policy research, constituent email, speeches, writing legislation.

Lately, she said, the signs of AI’s effects are getting easier to spot.

“It’s funny,” Hulse said, “how your ears will start to pick up on the floor someone who’s giving a speech that was very clearly written by AI.”