The Wall Street Journal’s new AI tool, “Doppler,” forecasts the issues that will shape the midterm elections

More than 100 million voters are expected to cast ballots in the 2026 U.S. midterm elections.

In recent election cycles, political reporting based on voter polling has drawn criticism, prompted scrutiny, and found defenders. As the debate over the value of polling continues, The Wall Street Journal is exploring new ways to cover public opinion and break out of traditional horse race political coverage.

In late September, the Journal published its first story using Doppler News AI, a custom tool built to help expand the publication’s midterms coverage. So far, Doppler has ingested more than 25 million TikTok videos, roughly 500,000 podcast episodes, and thousands of Instagram videos and cable news segments. The Journal uses the tool to monitor trends across the internet and to identify voter priorities, concerns, and sentiments that traditional polling might miss.

“We’re not trying to predict the election, but people encounter politics through a lot of different lanes: culture, work, money, family, community. We wanted to tap into that and listen in on those conversations,” said Aja Whitaker-Moore, the deputy editor-in-chief of the Journal. “We’re taking a massive set of information…and telling you, from a statistical standpoint: ‘Here’s what people are talking about. Here are the questions they’re asking. Here are the places where those conversations are happening.'”

Other newsrooms have leaned on AI “visibility tools” in the past to support their reporting. That includes using large language models (LLMs) to transcribe and summarize “manosphere” podcasts, school board meetings, and local talk radio shows. But Doppler goes further.

Rather than using AI-generated summaries to generate leads or help beat reporters monitor online content more efficiently, the Journal uses Doppler to conduct large-scale AI analyses quickly — and publishes those insights for readers.

So far, analyses produced by Doppler have led to four major midterms stories in the Journal, with more expected to run before Election Day. For now, each story requires a custom treatment, with engineers and data journalists working alongside the D.C. politics team to refine the methodology, toggle on and off different features, and audit the results.

“The tool is only as good as the questions that you ask it,” said Whitaker-Moore. The Journal holds a weekly editorial meeting leading up to Election Day to workshop story angles and find new threads to explore.

@wallstreetjournal

A new in-house AI tool lets WSJ reporters analyze trends in online conversation to help understand the political mood of the country ahead of the midterm elections. A previous post paired a visual of TikToks from across the internet with a headline that identified them as conservative podcasters.

♬ original sound – The Wall Street Journal

A story published on October 8 analyzed over 30,000 hours of podcasts and documented how, in recent weeks, right-leaning podcasters had started discussing “affordability” as often as left-leaning podcasters, despite President Trump calling the affordability crisis a “hoax.”

Another story published on October 1 traced anti-AI sentiment, including opposition to Flock surveillance cameras, by analyzing more than four million TikTok videos and thousands of hours of podcasts. The Journal found that posters who shared anti-Flock sentiments disagreed on other political positions about 12% of the time, three times the average rate across other political topics tracked — putting the issue “at the top of the list of positions that unite creators who are otherwise at odds.”

For a story published on October 15, the Journal pulled content about 67 Democratic politicians from its Doppler database. The analysis showed that the Democratic politician currently receiving the most favorable commentary is Angie Nixon, the progressive Senate nominee in Florida. About 70% of TikTok videos posted about her in recent weeks were favorable. She was followed by Georgia Senator Jon Ossoff at 54%. Meanwhile, Senate Minority Leader Chuck Schumer ranked the lowest in the analysis; just 3% of videos posted about him were favorable.

[img wide]Courtesy of The Wall Street Journal[/img]

Whitaker-Moore said the idea for Doppler was inspired by conversations with Washington bureau chief Damian Paletta and politics editor Christopher Matthews. They thought the coverage of the 2024 U.S. election missed important undercurrents in public opinion that never surfaced in polling. And though reporters acknowledged the power of podcasts in that election cycle, there weren’t many standardized ways to capture what was being said across those channels.

Coming out of those conversations, Whitaker-Moore posed a question to Rob Barry, who leads the Journal’s data journalism team: “Can you build something that can eat the internet and then feed it back to us?”

The Journal had already built its own podcast transcription and summarization tool, Orca. (The name refers to the tight-knit pods in which orcas travel.) Orca had collected countless hours of audio and YouTube videos, not just of political podcasts but from genres like health and wellness.

The Journal’s experience with TikTok-based investigations made the social media platform’s videos another natural source of information for Doppler. It’s nearly impossible to create a comprehensive database of TikTok videos, or even a truly random sample. Instead, Barry’s team used a monitoring tool that relies on TikTok’s own recommendation algorithms to identify the accounts it should track.

“Every day we’re adding tons of data across all the different platforms that we’re looking at,” Barry told me. “The database is growing in the background constantly.”

To make sense of that data, Doppler has a tech stack that includes different types of AI models, including LLMs. The model used depends on the specific reporting question.

Embedding models have been central to some of stories published so far. These AI models convert text or images into vectors — lists of numbers that capture the meaning of each piece of content. By vectorizing the content, Doppler can map out the ideas in a dataset and see where they naturally cluster together.

The Journal’s data journalism team has used vectorization in the past, including in its Pulitzer Prize–winning story that mapped out Elon Musk’s entire history of tweets. In Doppler, vectorization facilitates a kind of internet radar, helping reporters identify where patterns are forming across its social media and podcasting database.

For now, the Journal’s political reporters can’t just start using Doppler on their own. The exact methodology has been tailored to each story published, with the help of engineers.

“There’s a lot of human judgment in this. My team would not have aimed for a tool that seeks to just produce stuff that we then put directly on the internet,” Barry said. “You still need that human intermediary, that human to apply the judgment and to ask the question.”

The Journal is attaching a short methodology to every story assisted by Doppler. It makes clear that AI was used to produce the piece, briefly explains how the analysis came together, and notes some of the tool’s limitations.

Despite manual spot checks of data classifications and other safeguards to make sure the tool’s outputs are accurate, Barry said, “there’s almost always going to be some level of error and uncertainty and blur around the edges.” The disclosures are one way to “give readers a sense for what that level of issue is.”

Looking ahead, the Journal is considering bringing Doppler to other desks in the newsroom. The data being collected is not inherently political, and it could be used to help answer a host of other reporting questions, drawing on insights scattered across the internet.

“We definitely think that [Doppler News AI] can live beyond the midterms and live beyond elections,” Whitaker-Moore said.