US politicians are trying to change what chatbots say about them

WASHINGTON – Dustin Lloyd, a US Democratic primary candidate for Missouri’s state legislature, was a well-known member of his community when he started his political campaign.

But there was one important group that did not seem to know him at all: Artificial intelligence chatbots.

If voters quizzed a chatbot like OpenAI’s ChatGPT or Google’s Gemini about Lloyd, they would find only basic information that failed to convey his focus on helping small businesses.

Fortunately for Lloyd, there was a fix. He tweaked his online presence, publishing a Q&A about himself on his campaign website. Later, when the chatbots were quizzed again, they dutifully connected that personal history to his policy goals.

Editing AI’s output proved to be easier than he imagined. Since then, Lloyd said, “I’m constantly working on it every day”.

Candidates have long had to worry about their reputations among voters. Now, they have to worry about what AI thinks about them too. And a new industry has sprung up to help them navigate this world.

Lloyd, 33, relied on a report on his AI presence from CampSight, a tool launched in June by Run for Something Action Fund, a progressive group that recruits young candidates to run for office.

It determined that Lloyd’s website and his presence on Wikipedia and Ballotpedia – sites on which voters regularly find information about political figures – were lacklustre, advising him to make changes.

It even recommended posting threads on Reddit so the chatbots would begin including information from the online forum in their responses.

“Everyone’s just grasping for some sort of tool to at least inspect these results and ideally influence them,” said Jordan Haines, chief technology officer for Run for Something.

The industry, known as answer engine optimisation, or AEO, is a response to changes in how AI chatbots generate their answers.

The earliest chatbots, released several years ago, were trained on huge amounts of information that had already been published online. That made many of their answers about news events almost instantly outdated.

Modern chatbots solved that problem by searching the internet and pulling in fresh content to answer questions about current events, including the midterms and candidates.

Recommendations from AEO tools like CampSight are typically limited to rewriting content on a candidate’s own website or ensuring Wikipedia is up to date. But well-financed candidates may ultimately try to publish a wider array of content online that would be aimed squarely at getting noticed – and parroted – by chatbots, experts said.

“There is a certain degree of panic around it, because we understand it so little and it’s changing so fast,” said Beth Simone Noveck, a professor at Northeastern University who is studying democracy and AI tools, known as large language models or LLMs.

“Even if there are people who claim to understand it, even if they master some of the techniques, the changing and unpredictable nature of LLMs just means it’s really difficult to control this process,” she added.

Politicians are also weighing another risk: Getting smeared by AI search results that are not just unflattering, but also flat-out wrong.

An experiment before Scotland’s parliamentary election in May found that more than a third of AI-generated answers to questions about the upcoming vote were inaccurate.

Researchers for Demos, a British think-tank, found that multiple chatbots were inventing candidates, fabricating nepotism accusations or concocting financial scandals. Some falsely claimed that an incumbent was running when they were not, while others misstated candidates’ positions.

Still, voters are increasingly turning to chatbots for answers to their political questions. Caucus AI, a political AEO research company, estimated that at least 16 million voters were getting election information from AI, either through chatbots or from AI-generated answers on search results pages.

“As people get more used to this, and as the reliability of the information coming from these chatbots increases, we do think that there will be a significant uptick in 2028 and beyond,” said Meg Schwenzfeier, a founder at Caucus AI and former chief analytics officer of Kamala Harris’ presidential campaign.

Caucus AI ran an experiment to see how long it took for content newly published to Wikipedia to end up in a chatbot’s answer. The result: About 12 minutes. That suggests that candidates can already insert their own information into an AI’s responses by tweaking public information “quickly, unvetted and without anyone on the other side watching”, the company wrote in a blog post about its findings.

Caucus AI is tracking what chatbots are saying about all Senate, House and gubernatorial candidates, finding that some candidate messaging has an unusual ability to break through.

For example, when the group asked three different chatbots about Mary Peltola, the Democratic challenger for US Senate in Alaska, they all cited her unique slogan: “Fish. Family. Freedom.”

“We’re exploring what causes that to get picked up,” Schwenzfeier said. “There’s still a lot we don’t understand, and these models themselves are black boxes.”

If candidates and political groups can nudge chatbots to produce more palatable answers, experts in digital strategy worry that foreign influence operatives and others might also try to manipulate AI searches.

“Disinformation actors are, by definition, early adopters – they will use every technique and every technology they can,” said Tim Chambers, who runs the digital and social media arm of Dewey Square Group, a public-affairs firm in Washington whose clients include labour unions and government agencies.

The firm has been helping clients audit their websites to see how accessible they are to chatbot scrapers; in research published in February, Dewey found that far-right sites and sites with “very low” factual integrity had set up their websites to be easily seen by AI tools, compared with outlets that were centre-left or highly factual.

Chambers said that his clients were thinking about “needing to rebuild their websites for machines as much as for humans”, and considering how to show up in chatbot queries posed in multiple languages.

For Lloyd, the Democratic primary candidate in Missouri, the changes he made had an instant impact. When the chatbot was first asked who voters should support in the primary, it recommended his opponent, Tanya Lakins, citing her focus on small businesses. After Lloyd made tweaks to his website, the chatbot changed its tune.

“If you’re voting in the Democratic primary,” it wrote, “Dustin Charles Lloyd appears to have the strongest and most explicit small-business platform.” NYTIMES

• This article originally appeared in The New York Times.