Millions of people now regularly turn to artificial intelligence chatbots like Claude and ChatGPT to answer any of life’s pressing questions. Later this year, political operatives expect voters to turn to the chatbots with another one: who should I vote for?

Most of those operatives — Republicans and Democrats — just have no clue what the chatbots will say. Or what they can do about unpleasant answers.

Six months before November’s critical election, some of the most influential sources of information on the internet remain a black box for operatives in both parties. Many privately concede that they — and the political committees and groups they work for — have spent little time thinking about large language models, or LLMs, they consider largely inscrutable.

And even those who have begun working to influence LLMs concede their work is in its very early days.

“Nobody has written that playbook yet on the LLM stuff,” said Pat Dennis, the president of the liberal outside group American Bridge 21st Century. “So a lot of people just aren’t doing it.”

Bridge is one of the few groups whose leaders have said publicly that they’re trying to shape what answers LLMs give to people who ask about politics, an effort that can resemble a supercharged version of long-standing campaigns to affect search-engine optimization for Google.

But their effort is, for now, an isolated one, according to interviews with more than a dozen political strategists.

“To me, it just feels like a giant space race, one in which the first rocket hasn’t been launched yet,” said one national Republican operative, granted anonymity to speak candidly.

The burgeoning effort to shape LLMs is the latest battle over the emergence of AI in politics, which has affected everything from internal campaign operations to a public debate over data centers. Democrats in recent weeks have been openly frustrated that their party has been slower than Republicans to adopt AI-enabled tools in their political committees and campaigns.

With LLMs, however, Republicans aren’t sure their party is as far ahead of their opponents as they are elsewhere in the integration of AI into their daily operations.

A poll from Elon University from last January found that more than half, 52%, of Americans now use LLMs, a share of the population that has likely grown as traditional search traffic continues to drop.

Even in an environment in which voters are constantly bombarded with political messages, the answers they get from these platforms will likely be a key source of information in an election in which Republicans and Democrats are battling for majority control of both the House and Senate.

“My basic take is that LLMs are going to be more politically influential than your most AI-skeptical friend thinks, and less politically influential than your most AI-friendly friend thinks,” said Rob Flaherty, the deputy campaign manager for Kamala Harris’ 2024 presidential campaign. “And that still means they’re going to be important.”

LLMs, even more than results from traditional search engines, can give direct answers to voters about politics. Talking with the chatbots about the Senate race in North Carolina, where Republicans are trying to make Democratic nominee Roy Cooper’s record on crime a big issue, shows how that can play out.

Asked by a reporter from NOTUS about Cooper’s record on crime, ChatGPT responded that it “draws sharp partisan divides: critics portray him as soft-on-crime for commuting death sentences, supporting bail reform, and facilitating early prisoner releases linked to subsequent murders, while supporters highlight contextual factors like COVID-era overcrowding and rehabilitation efforts.”

The same platform, asked if moderate Republicans should support the former two-term governor, says Cooper “projects a moderate image through calm rhetoric and bipartisan deals like Medicaid expansion, but his record features liberal policies on social issues, gun control, and criminal justice that clash with core Republican priorities, making him a poor fit for moderate Republicans seeking fiscal conservatism or traditional law-and-order stances.”

There’s a similar effect when talking to the chatbots about the early potential 2028 presidential contenders, where the LLMs appear to put more weight on potential candidates who have been most prominent in news reporting.

When given the prompt “I care about affordability — which candidate cares most about that,” Claude responded with a detailed breakdown of different potential candidates, but spent the most time on Gov. Gavin Newsom of California.

“The Newsom problem on affordability: California under Newsom has the most homeless people in America despite massive spending, and critics point to the state’s high cost of living as a liability for him on this specific issue,” the bot wrote.

Claude pegged Newsom and Vice President JD Vance as the frontrunners for their respective nominations. When asked “who has the most scandals,” Claude listed out scrutiny Newsom has received back to 2007 and noted Vance’s “sharp political pivots, though he hasn’t accumulated the same volume of governance-related controversies since he’s been VP rather than a governor.”

LLMs are designed to be neutral, say officials at these AI companies. They draw from publicly available news stories, YouTube videos and other sources to come up with the most accurate and comprehensive information, the officials say. They are trained to be resistant to manipulation from anyone, officials say, including political actors.

“We work to train Claude to be politically even-handed in its responses,” Anthropic wrote in a blog post last year. “We want it to treat opposing political viewpoints with equal depth, engagement, and quality of analysis, without bias towards or against any particular ideological position.”

AI experts don’t even know exactly how to best shape what these platforms say, known as answer-engine optimization. But even the methods that do work might be difficult for some individual campaigns to implement, given that many of them are already trying to do everything they can to reach voters in other ways.

“Part of the challenge for these campaigns is you have so many day-to-day challenges already, in terms of getting your message out there and doing voter contacts,” said the national Republican strategist. “So something like this, the concern around what’s being fed into LLMs, it’s not as urgent. But it is important.”

Other strategists pointed out that campaigns still don’t do enough to affect SEO, much less AEO, and any expectation that they would suddenly catch up this year is misguided.

But even if most political operatives in Washington react to questions about LLMs and politics with a shrug, there are some who say they’ve made an early and somewhat fruitful effort to shape what they say. Some of them also suspect that other organizations might quietly be making an effort to influence these chatbots, even if they are not discussing it publicly.

“LLMs are becoming a larger part of the professional and personal lives of more Americans every day,” said Joanna Rodriguez, a spokesperson for the National Republican Senatorial Committee. “Any campaign entity not taking that into consideration is at risk of falling behind.”

The two chief analytics officers for Harris’ 2024 campaign, meanwhile, have built a tool, called Caucus AI, that lets people track what the LLMs Gemini, Grok and ChatGPT say about candidates and elections in each state, stores the responses and shows which sources the chatbots use in their answers. The tool was built, people involved with its design say, out of a recognition that this is the first election cycle in which the answers LLMs give voters will be of significant importance.

Political operatives familiar with AEO say there are a few widely understood ways to shape what shows up there. Candidates’ websites, for instance, need to have direct, clearly written information about their background and policy positions. Content produced by the campaigns also needs to be written this way, and it needs to feed into a multitude of platforms. (Gemini, Google’s AI, is said to rely more heavily on videos from YouTube, also owned by Google, while Elon Musk’s Grok relies more heavily on X than other LLMs.)

Wikipedia is also key, although the operatives said the site is already more difficult for campaigns to edit.

“The challenge is the number of sources they are pulling from. It’s not a Google or search engine, they are not just trawling the internet,” Flaherty said.

Kelsie Taggart, a vice president at American Bridge, said her group started making a serious push last year to make its opposition research on Republican candidates more AEO friendly, publishing the documents in a way that makes them easily searchable. She and other leaders at Bridge considered it an essential activity for a group that traditionally has put a priority on shaping the public image of Republican candidates.

“We definitely see this as a main tool that voters are using to educate themselves in the coming election,” Taggart said. “But we’re trying to figure this out while the ship is already sailing, and figure out the best way to get the correct information.”