Minnesota’s Sahan Journal dives into AI with strategic goals and open eyes

As journalists around the world experiment with artificial intelligence, many newsrooms have common, often audience-facing, ideas for what to try.

They range from letting readers talk to chatbots trained on reporting, to turning written stories into audio, creating story summaries and, infamously, generating entire articles using AI — a use case vehemently rejected by many journalists.

But Sahan Journal, the nonprofit newsroom serving immigrants and communities of color in Minnesota, wanted to try something different.

“We’re less enthusiastic, more skeptical, about using AI to generate editorial content,” said Cynthia Tu, Sahan Journal’s data journalist and AI specialist.

Instead, the outlet has been working on ways to support internal workflows with AI. Now, it’s even testing a custom ChatGPT bot to help pitch Sahan Journal to prospective advertisers and sponsors.

In the process of creating that bot, Tu joined the forefront of the nascent, and somewhat chaotic, industry shakeup as journalism tries to figure out what it can and can’t do with AI.

And while AI has plenty of ethical and technical issues, Tu’s work highlights another important aspect: The intended users — in this case, the Sahan Journal team.

“A lot of…this experiment is less of a technical challenge,” Tu said. “It’s more like, how do you make [AI] fit in the human system more flawlessly? And how do you train the human to use this tool in a way that it was intended?”

Sahan Journal’s AI experimentation, and Tu’s job, are supported by a partnership between the American Journalism Project, a national nonprofit helping local newsrooms, and ChatGPT creator OpenAI.

The resulting two-year $200,000 grant to Sahan Journal runs through the end of 2025, with an additional $20,000 grant given this year to support more revenue experiments.

Liam Andrew, technology lead for the AJP’s Project & AI Studio, sees part of his job as helping newsrooms overcome hesitancy around AI.

“I’m trying to get folks to not think of AI as this technology…on the top of a skyscraper that they’re trying to climb,” Andrew said.

“It’s just a tool that they can use right now,” he said. “Maybe it’s useful, and maybe it’s not…there’s no imposter syndrome here. Everyone is figuring this out as we go.”

Scaling up AI solutions

Tu joined Sahan Journal fresh from a Columbia Journalism School master’s program in data journalism. She had played a little with chatbots, but otherwise didn’t have much experience working with AI.

So she decided to start with the basics, learning the ins and outs of the tech to prepare for larger projects.

“The first six months…I was more like, ‘Let’s just see what we can do,’” Tu said.

For one investigation, Tu used a Google AI tool to process the financial data of charter schools in Minnesota. Thinking about how to save time on backend workflows, Tu then helped Sahan Journal generate story summaries, tailored for Instagram carousels, with ChatGPT. Other experiments, like turning some stories into audio with AI, also followed.

Importantly, this time was an opportunity to learn about Sahan Journal itself.

“There’s a lot of things you have to take into consideration when you’re developing a tool,” Tu said.

“You need to know what the workflow of the organization looks like…[and how] you push for change within a department when they’ve already been doing [something] for the past five years using a manual or human labor way.”

That knowledge came in handy when finally tackling Tu’s core AI project: improving Sahan Journal’s revenue.

The project stemmed from an anonymized database of audience insights, which included demographic information and interests. While an important resource, Sahan Journal’s small revenue team didn’t have the time to figure out how to leverage it.

“The idea at the very beginning was really rough — we felt like we could do something [with] this data, but we’re not sure what,” Tu said. “At the same time, we were also trying to revamp our media kit.”

What if AI could feed two birds with one scone? A custom ChatGPT bot could process the audience data and personalize a media kit for clients. But it needed to work without being an extra burden on the revenue staff.

As Tu remembers, it took her only two days to create the first version of the bot. Refining it, though, would take months.

Easy doesn’t always mean simple

The magic of AI chatbots like ChatGPT is that you don’t need to know how to code to use them. Just type in a prompt and get rolling (though, uh, retain some common sense).

Less magically, AI chatbots can be hard to keep in line for specific tasks. Designed to be eager helpers, they hallucinate false results and stubbornly twist instructions in an attempt to please.

Troubleshooting those issues was no simple task for Tu.

The custom revenue chatbot struggled to keep Tu’s preferred formatting, and hallucinated audience data. The bot would also intermix results from the internet that Tu had not asked for. None of that was ideal for a tool that should work reliably for the revenue team.

“I was kind of jumping through hoops and telling it multiple times, ‘Please do not reference anything else on the internet,’” Tu said.

The bot didn’t listen. So she turned off its access to the internet. But “it still kind of hallucinated,” she said.

Working with chatbots is an exercise in prompt engineering — mostly a trial-and-error process of figuring out what specific instructions will get the preferred result. As Tu said, “lazy questions lead to lazy answers,” though that is changing as AI rapidly evolves.

“Even the last couple of months, you can get lazier with your inputs,” said the AJP’s Andrew, who worked with Tu on the custom revenue chatbot. “We’re just running by examples that we see, and what seems to be working.”

Eventually, Tu settled on a reliable set of prompts.

The custom chatbot takes about 20 seconds to find relevant data from the audience database — for example, pulling up how much of Sahan Journal’s audience cares about public transportation. Then it creates a summary for a media kit tailored to potential clients.

The chatbot also double-checks its work by referencing the database again, making sure its output matches reality. And part of the database is shown for users to manually see the chatbot isn’t hallucinating.

Another separate ChatGPT bot, connected to Sahan Journal’s website, can pull relevant articles for the media kit to show the outlet’s impact.

Earlier this year, Tu introduced the final version of the revenue bot to Sahan Journal’s team.

“Their first reaction, just kind of seeing how fast everything got turned out, was, ‘This is such a slick product,’” Tu said. That was a welcome contrast to “some hesitation in terms of people being like, ‘Oh, I’m not sure how to use ChatGPT. I’m just not that tech coded’…But it turns out it’s actually pretty simple.”

Sticking the landing at Sahan

By mid-April, the Sahan Journal revenue team had used the custom chatbot on six sales pitches, with three successfully leading to ads placed on the site, Tu said.

The experiment should show whether the AI tool measurably increases revenue, how easily it fits in the revenue team’s workflow and how clients react. If successful, it could be a model for other newsrooms.

But there’s a larger question hanging over this work: Is it sustainable? In a way, newsroom experiments with AI exist in a bubble.

“Everything is kind of tied to a grant,” Tu said, referencing the AJP-OpenAI partnership that supports her work. But grants come and go as donor interests (and financials) change. “We don’t really know if next year, at this point, I can still pursue those projects.”

Then there’s security. As part of the AJP-OpenAI partnership, OpenAI committed to not using newsroom data to train its AI models. That’s a cornerstone of Tu’s work, given she has a ChatGPT bot plugged into Sahan Journal’s audience data.

There is some buffer, as the custom bot is referencing a secondary, abridged and anonymized database rather than the raw dataset. But without the AJP partnership, OpenAI (or other AI providers) could treat newsrooms like any other user — swallowing inputs for its own gain.

And while newsrooms experiment in a low-cost AI environment, OpenAI is hemorrhaging money. That could put the longevity of AI tools at risk.

“We talk about what sustainability means — is this a tool that needs to last for five years, or is it O.K. for it to last for one year? And what would you learn from it?” said the AJP’s Andrew about his conversations with newsrooms.

Andrew takes the issues with AI in stride, but is ultimately optimistic. He points to the variety of AI tools and ChatGPT alternatives, and sees journalism’s smaller-scale AI development as having a good chance of outlasting any instability. Besides, planning for limited use can also have larger benefits.

“It helps teach a lot of these newsrooms some…important product thinking, techniques and practices,” he said.

Tu also recognizes AI’s pitfalls, from concerns about AI replacing human journalists — something Sahan Journal’s AI policy explicitly rejects — to the environmental impact of the technology.

But AI also seems to be the future, and it helps to come in on the ground floor.

Working at Sahan Journal is “a great opportunity for me to be able to really dedicate my time to work on this stuff,” Tu said.

“Even though I’m not sure where it’s going to go next, I do know that AI is going to be integrated into journalism one way or the other,” she said. “Knowing how to interact with it ethically is going to be a strength.”