On March 7, education reporter Hannah Dellinger published a story on the experiences of Michigan LGBTQ+ students since Donald Trump took office. Dellinger spoke to several students who have seen a rise in hate speech at school after the president signed a series of anti-trans executive orders. The lead voice in the story was Sebastian Eaton-Ellison, a gender-fluid senior who has faced relentless bullying, and even physical assaults, at his high school.

As a reporter for Chalkbeat’s Detroit bureau, Dellinger has often found student sources through school board meetings. This story was no exception. Eaton-Ellison had shared his experience with bullying at a recent Traverse City school board meeting — but Dellinger wasn’t in attendance. Instead, she found his testimonial by searching keywords on LocalLens, a database that uses AI to transcribe and summarize local government meetings.

“I’m not able to physically go to every single board meeting. Even watching them online, or looking at all the agendas, can take up an entire day,” Dellinger told me. (Traverse City, for example, is a four-hour drive from Detroit.) In recent months, Dellinger has used LocalLens as an extra pair of ears, listening in on public meetings across the several Michigan school districts on her beat.

“My source, Sebastian Eaton-Ellison, is a kid who in that moment, at that board meeting, for the first time ever, spoke up publicly,” she said. “He’s somebody who really wants to get involved in advocating for his rights…but I don’t know that I would have found him if I wasn’t using this tool.”

Increasingly, AI transcription tools powered by large language models (LLMs), like LocalLens, are assisting journalists who cover local government – whether that be with auto-generated summaries of town halls meetings, school district proceedings, or even hearings in state legislatures.

By most accounts, these tools aren’t able to generate publishable stories, or even publishable transcripts. But they are helping local reporters find sources, stay informed about public meetings they can’t attend in real time, and expand the scope of their coverage.

“We are going to be in the rooms where we need to be, where the big decisions are being made, but we can’t be everywhere all the time,” said Eric Gorski, the managing editor for local news at Chalkbeat, who has been coordinating the newsroom’s use of these tools. “The summaries are springboards for more reporting. It’s not a replacement for coverage, and we’re not trusting AI to get these things right. It’s more like a news tip.”

To build or not to build

Dellinger’s story is only the most recent example of Chalkbeat reporters leaning on AI meeting summaries to find tips. Back in 2023, Chalkbeat’s parent organization, Civic News Company, received a grant to experiment with the technology from the American Journalism Project’s Product and AI Studio. The program is backed by a $5 million donation from OpenAI. At first, Chalkbeat used this grant funding and coaching for a pilot in its New York bureau, which houses four reporters who cover the largest school system in the country.

Originally, Chalkbeat built its tool in-house for this pilot, bringing on a former Meta product VP to consult and eventually hiring a dedicated software engineer, Perrin Myerson. Myerson shouldered the engineering work, product management, and fine-tuning needed to custom-build an AI summary tool. Gorski called their work “indispensable” to the pilot. “We just didn’t have that technical expertise on staff,” he said.

By November of last year, Chalkbeat had a workable prototype. There were early wins in this pilot, including a story published by New York bureau chief Amy Zimmer about parents who were angry that outdoor recess time for their kids was increasingly being replaced with indoor screen time. The story was rounded out by comments from a Brooklyn parent at a local board meeting that Zimmer found through an AI summary.

Chalkbeat made many adjustments to the tool as they went. For one, Gorski said they began prompting it to surface all statistics or hard numbers mentioned during board meetings. After one of Zimmer’s stories was mentioned outright in a meeting, they also prompted the tool to flag any time Chalkbeat was named. Alerts helped document the impact of their stories on local policy, and the reach of stories among citizens and school officials.

The underlying large language models’ shortcomings also quickly became apparent. The summaries and transcripts the tool churned out were far from polished. The tool struggled to accurately identify individual speakers, despite regularly handling meetings in which at least six school board members could be speaking.

And the New York bureau learned any information or quotes coming from the AI summaries needed to be fully fact-checked. “It is a tip, and similar to getting a tip, you need to check it out. You need to confirm the accuracy,” Gorski said.

One of the biggest challenges with the pilot, and building an in-house tool overall, was friction in the editorial workflow. As Gorski puts it, “people are slammed,” and for many this was just one more thing on their plate. Even minor adjustments to the core prompt, for example, required reporters to ask Myerson to make manual tweaks for them.

By the end of last year, Chalkbeat decided to scale up the pilot from New York to all eight of its city bureaus across the country. They also decided to leave their custom built tool behind, for now. So far, third-party vendor LocalLens has offered a more seamless user experience for reporters. They have the option of searching an entire database of public meetings by keywords and setting up custom alerts, alongside access to annotated transcripts. Currently, Chalkbeat is monitoring about 80 school districts in 30 states through its LocalLens enterprise account.

The next hurdle is getting buy-in from the newsroom at large. Some reporters are reasonably skeptical of using automation in their reporting. Others might fear being automated out of their jobs.

Even Dellinger was “apprehensive” when her Detroit bureau chief first mentioned experimenting with LocalLens late last year. In the months since, her perspective has changed.

“I wouldn’t consider myself an advocate for using it, other than being an advocate for anything that helps journalists do their jobs in a time where we’re really short-staffed and there are news deserts,” Dellinger told me. “If it’s being used in a way that is amplifying the voices of students who I cover — which it has in my case — then, yes. I would advocate for that.”

“I would much rather have anything than nothing”

For a small local newsroom like the Midcoast Villager, building an AI transcription tool in-house was never in the cards. Based out of Camden, Maine, the print weekly and digital outlet launched last fall. Its newsroom may lack AI engineering expertise, but it’s full of tenured local reporters. The publication is a roll-up of four legacy newspapers that used to serve the surrounding Knox and Waldo counties.

Even with a merged editorial staff, though, the Villager’s team can be stretched thin. Its market spans 43 towns, many of which are on offshore islands only accessible by ferry. Like Chalkbeat, the team found there are only so many places their reporters can be at once.

A new partnership with Civic Sunlight, a startup providing AI summaries and transcripts of public meetings via email, is an experiment to expand the outlet’s reach.

“Sending a reporter to drive an hour to go sit in a four-hour meeting that will probably not make news 90% of the time isn’t worth the time and effort,” said Alex Seitz-Wald, the deputy editor of Midcoast Villager. Having Civic Sunlight sit in on the meeting is a convenient and affordable alternative.

It also doesn’t hurt that Civic Sunlight’s offices are literally down the street. “I’m looking out the window right now, and I can see the building that their office is in,” said Seitz-Wald, in a recent phone interview.

Civic Sunlight was co-founded by two local residents, including Tom Cochran, the former CTO of The Atlantic and director of new media technologies in the Obama White House. Midcoast Villager is its first newsroom customer. “We linked up because we know everyone in town — it’s a really small town,” said Cochran. “We said this is important to you and you are important to us. You are the media organization that matters here and we have expertise that can support you.”

Civic Sunlight started as a newsletter brief for the general public that uses a multi-model AI system to summarize local government livestreams. “The equivalent of a Zoom transcript is what we call a dirty transcript,” said Cochran. “Then we clean it up to make sure that it is high quality.” That includes automated fact-checks of proper nouns and names. The final email is a concise roundup of recent public meetings, with link outs to specific timestamps of the recording on YouTube, Vimeo, or other platforms.

“We quickly discovered that [there are only two ways] to monetize this: journalism or special interests,” said Cochran. “We’d rather support journalism and local news.”

Since the partnership launched at the beginning of this year, Civic Sunlight’s email summaries have started to feed directly into Midcoast Villager reporters’ inboxes. Seitz-Wald likens the summaries to the emails he used to receive from campaign field reporters at NBC News (where he was a national political reporter until last month), and said the Villager has already published a handful of news stories using leads or quotes found in the transcripts.

“[Civic Sunlight] probably is not as good as a junior reporter, but it’s on a cost and scalability that is accessible,” Seitz-Wald said. “I would much rather have anything than nothing.”

One way the tool falls short of human reporters is understanding regional dialects. “The Maine accent has been a challenge,” Seitz-Wald admitted. Like Chalkbeat, the Villager expects all reporters to independently confirm anything found in a Civic Sunlight summary prior to publication.

Despite errors, the newsroom already has plans to spin up the experiment. The Villager publishes a network of town columnists — vetted residents who contribute regular local dispatches. Those columnists will receive training on how to use Civic Sunlight and verify information from it, starting with a workshop in Lincolnville, Maine in the coming weeks. Eventually, the Villager plans to run summaries of local government news beneath their regular columns.

Despite leaning on Civic Sunlight to expand coverage, Seitz-Wald isn’t worried about these AI vendors becoming direct competitors anytime soon. He doesn’t even see them as a full replacement for the work of reporters covering public meetings in person, with a recorder in hand.

“AI can never go up to the town official at the end of the meeting and get a fresh quote that has never existed anywhere in the world. It can’t read the facial expression of the town members. It can’t notice when somebody leans over and whispers something in the ear of a colleague, and then follow up to ask about that,” Seitz-Wald said.

“I’m not a doomer about AI in the news, because it only knows what already exists on the internet,” he said. “Life [doesn’t just happen] at town meetings.”