As part of its ongoing Webinar series with OpenAI, the INMA Generative AI Initiative recently offered an in-depth look at how news media companies are using generative AI to improve their research.   

From The Philadelphia Inquirer’s new archive search tool to the American Journalism Project’s field guide for local reporting, newsrooms are experimenting with GenAI to reimagine how journalists conduct research. With help from OpenAI’s Deep Research feature in ChatGPT, newsrooms are finding new ways to improve research, efficiency, and accuracy. 

The Inquirer’s digital librarian

Like many newsrooms, The Philadelphia Inquirer began discussing last year how to make the most of GenAI.  

Matt Boggie, chief product and technology officer, explained that as the newsroom started exploring its potential, internal conversations focused on “what GenAI was useful for, what concerns we had about the technology, and how we might best use it to support and expand the work of our newsroom.”

Looking at how the technology could support research, the team’s flagship project, Dewey — a nod to the Dewey Decimal System — was designed to search the Inquirer’s archives dating back to 1978. Built as a retrieval‑augmented generation (RAG) application, Dewey allows reporters to ask natural language questions and receive conversational answers complete with inline citations.

Kevin Hoffman, the Inquirer’s AI engineer, noted that Dewey rewrites ambiguous queries into precise research questions. For example, “Who was the mayor last year?” becomes “Who was the mayor of Philadelphia in 2024?” Dewey then restricts its search to relevant date ranges and surfaces articles that justify its answers. Reporters can follow up with conversational queries, such as asking about campaign policies, with Dewey resolving references and expanding searches accordingly.

Evaluation of Dewey showed an 82% accuracy rate across dozens of newsroom queries, with success tied directly to citation relevance.

“What we found was that Dewey was really good at being able to answer questions over large periods of time [because] there's so much data to be able to enrich its answer,” Hoffman said.

But the system had a surprise added benefit, they discovered: “Despite having designed this specifically for the news reviews, [its] usefulness extended beyond it. Some of our marketing team had been able to use it for pulling quotes, for instance.”

Although reporters praised its ability to handle large spans of time and conversational context, they voiced concerns about incomplete coverage, as it was limited to retrieving 20 articles, and its subjective understanding limitations.

Adoption peaked during pilot launches and company‑wide rollout, but usage plateaued, and Hoffman said there are plans to explore knowledge graphs to improve comprehensiveness.

That, he said, will allow reporters to “get all the facts so they are able to go through and verify all of the different articles pertaining to a particular topic.”

The next project the Inquirer focused on is Scribe, a tool that aggregates and transcribes public meeting recordings from 34 municipalities and school districts.

Steven Stirling, data editor, introduced the tool, which uses AssemblyAI for transcription and OpenAI for summarisation. It then produces bullet‑point reports highlighting the most newsworthy items, scored by a bespoke index developed with reporters and editors. Scribe has enhanced coverage, allowing it to track more stories, he said.

“The Philadelphia Inquirer covers hundreds of school districts and municipalities,” Stirling said. “Each of those entities hosts public meetings on a monthly basis, several of them a month. We cannot possibly be at all of those meetings.”

However, Scribe can transcribe the video feeds of those meetings and provide the newsroom with summaries that inform coverage. Reporters now receive tailored e-mail reports, searchable databases of meeting summaries, and even Slack alerts for high‑value events.

AJP’s field guide for reporting

At The American Journalism Project (AJP), GenAI serves a dual purpose: strengthening editorial coverage and enhancing the organisation’s fundraising capabilities.

Founded in 2019, AJP provides grants and coaching to small non-profit newsrooms, many of which operate with limited staff and little to no engineering support. Two years ago, with funding from OpenAI, AJP launched its Product and AI Studio to help these organisations experiment with AI tools and adopt them responsibly.

Liam Andrew, technology lead, said the studio combines grantmaking with hands‑on coaching, building a community of product staff who share solutions and best practices.

“A lot of what we’re doing is helping folks adopt AI in smart, responsible ways,” he said. AJP also streamlines the adoption of AI by sharing best practices: “We’re doing it full-time. A lot of these folks have a lot of other jobs they’re doing and AI is just something that they want to wrap their heads around while they’re doing it.”

On the editorial side, AJP has focused on civic information, using AI to expand coverage of local government meetings. With many municipalities posting lengthy or inconsistently formatted records, small newsrooms struggle to keep up.

GenAI tools are being tested to transcribe, summarise, and highlight the most newsworthy moments, and AJP recently published a field guide to AI for local reporting. These cataloguing tools can help newsrooms scale meeting coverage, identify patterns, and streamline research.

That’s critical in a time when many local news outlets are shrinking and resources are disappearing: “These newsrooms … are finding that they need to do more to cover their community, things that other newsrooms might’ve been doing in their city that they aren’t able to cover. So [with AI] there’s more breadth of coverage,” Andrew said.

AI is also proving transformative in fundraising and revenue growth. Newsrooms are using Deep Research tools to identify prospective donors, build standardised profiles, and uncover connections to boards or community organisations.

“This used to take weeks and now it takes days, even hours,” he said. “It’s a game changer for fundraisers.”

Outlets like The Salt Lake Tribune are applying these methods to expand into new markets, evaluating audience demographics and sponsorship opportunities with AI‑driven insights.

Across 39 projects to date, AJP has found GenAI most effective at the early stages of workflows: “I really recommend it very early in the pipeline, especially because later in the pipeline is where mistakes are more costly. Hallucinations can be a major challenge,” Andrew said.

By embedding AI into research processes, whether for fundraising or news coverage, small non-profit newsrooms can cover more ground, grow sustainably, and strengthen their ties to the communities they serve.

“AI can really help the research process across the whole organisation,” Andrew said.

Deep Research in action

To conclude the Webinar, members heard from Yiren Lu, solutions architect at OpenAI, who shared how to best use the Deep Research feature in ChatGPT. Introduced about two years ago, it has emerged as a powerful tool for newsrooms seeking to streamline complex reporting tasks.

“The goal of Deep Research is really not just to inform you, but to equip you with a well-organised narrative that’s ready for presentation and decision-making,” Lu said.

Deep Research acts like a thorough research assistant, scanning and synthesising information from across the Internet — including public filings, press releases, and profiles — and then presenting findings in a structured, citation‑rich format.

Unlike traditional search engines, it doesn’t just return links; it reads sources, identifies key themes, contrasts viewpoints, and produces outputs ready for editorial review, she said.

Lu walked attendees through the workflow, which begins with a clear prompt. She emphasised that the success of the results depends on how precise the prompt is: “Similar to talking to a human research assistant, your clarity will determine their success,” she said, adding there are three key principles to follow:

1. Set a role and a goal. “Tell the model who it should act as and what it should deliver.”
2. Add structure with specific questions. “You want to avoid vague queries.”
3. Be explicit about preferred sources, such as peer-reviewed articles, specific newspapers, government white papers, etc. “The model will respect these constraints and you'll get more credible citations as a result.”

She offered examples of how Deep Research can save newsrooms hours of work, but cautioned that, while powerful, it is not infallible. Hallucinations can occur, such as outdated job titles or duplicate LinkedIn profiles, so outputs should be treated as drafts and verified.

“You can mitigate the risk by being very precise in your prompt, by being very precise in the types of sources you want it to look at, and just being very prescriptive in what you tell it to do,” Lu said.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .