Perplexity, an AI search engine that has courted controversy by lifting liberally from news articles and skirting web-scraping rules, this week promised to serve as a reliable source for live information on the tightly contested US presidential election.

Perplexity promised that its Election Information Hub would serve as “an entry point for understanding key issues, voting intelligently, and tracking election results.”

“There is only one AI that can do this,” Perplexity’s CEO, Aravind Srinivas posted on X. Srinivas appeared to troll the publisher of The New York Times by posting a message on X offering to help while Times Tech Guild workers strike during contract negotiations; he later posted that the offer was for infrastructure rather than AI-generated content.

Perplexity’s tool did not end up making any gaffes last night, providing mostly accurate voting information and also accurately tracking the results as they came in—but largely because it dialed down the use of AI. Perplexity is currently finalizing a funding round worth $500 million that would give the company a valuation of $9 billion, a source familiar with the situation confirmed to WIRED yesterday.

Using a large language model to help summarize information from the web is problematic because when models are unsure they tend to fabricate facts. Perplexity placed tighter guardrails around this by ensuring that information on candidates, ballot measures, and polling sites were summarized using information sourced from Democracy Works, a nonprofit founded to provide voter facts. Live results were provided through a partnership with the Associated Press, which offers such data through an API. Additional information came from a carefully curated list of trusted sources.

Alon Yamin, cofounder and CEO of Copyleaks, a plagiarism detection company, says it is good to see Perplexity providing more guardrails around the election information it shares, as well as clear citations. Last week Copyleaks published an analysis showing that Perplexity appears to summarize information from behind some publishers’ paywalls.

Yamin adds, however, that errors are possible, and users need to verify the source of information. “Nothing that is created by AI is fully authentic,” he says. “The main issues are the same—hallucinations are still an issue, information could be partially correct, and so on.”

Perplexity’s Election Information Hub might also blur the line between verified and free-wheeling AI-generated information. While some results come directly from trusted sources, searching for more information triggered open-ended AI-generated results from the wider web.

Other AI companies appear to be taking a more cautious approach to the election. In WIRED’s testing, ChatGPT Search, a newly launched service from OpenAI, often declined to provide information about voting. “We’ve instructed ChatGPT to not express preferences, offer opinions, or make specific recommendations about political candidates or issues even when explicitly asked,” Mattie Zazueta, an OpenAI spokesperson, told WIRED.

The results were often inconsistent, however. For instance, the tool sometimes refused to provide talking points to help persuade someone to vote for one candidate or the other, and sometimes willingly offered some.

Google’s search engine also avoided providing AI-generated results in relation to the election. The company said in August it would limit use of AI in relation to the election in search and other apps. “This new technology can make mistakes as it learns or as news breaks,” the company said in a blog post.

Even regular search results sometimes prove problematic though. During voting on Tuesday, some Google users noticed that a search for “Where do I vote for Harris” provided the location of voting information while a search for “Where do I vote for Trump” did not. Google explained this was because the search interpreted the query as one related to Harris County in Texas.

Some other AI search upstarts are, like Perplexity, taking a bolder approach. You.com, another startup that blends language models with conventional web search, on Tuesday announced its own election tool, built in collaboration with TollBit, a company that provides AI firms with managed access to content, as well as Decision Desk HQ, a company that provides access to poll results.

Perplexity appears to have been particularly bold in its approach to upending web search. In June, a WIRED investigation found evidence that a bot associated with Perplexity was ignoring instructions not to scrape WIRED.com and other sites belonging to WIRED’s parent company, Condé Nast. The analysis confirmed an earlier report by developer Robb Knight concerning the behavior of bots operated by Perplexity.

The AI search engine is also accused of cribbing liberally from news sites. For instance, also in June, a Forbes editor noted that Perplexity had summarized extensive details of an investigation published by the outlet with footnote citations. Forbes reportedly sent a letter threatening legal action against Perplexity for the practice.

In October, News Corp sued Perplexity for ripping off content from The Wall Street Journal and the New York Post. The suit argues that Perplexity is breaching copyright law because it sometimes fabricated sections of news stories and falsely attributed words to its publications.