In July 2023, the American Journalism Project launched its Product & AI Studio to explore the smart application of generative AI and other technology within local news. This Studio, made possible by OpenAI with additional support from the Patrick J. McGovern Foundation, is sharing regular updates of its learnings.
Overview: In Dec. 2023, Open Campus received support from the Studio to use AI to summarize higher education data and explore tools for their local partners. Here’s an update on their progress so far:
Building useful tools for newsrooms
In September 2023, Scott Smallwood, CEO of Open Campus, was thinking about generative AI use cases for his network of local higher-education reporters. Open Campus embeds reporters in local newsrooms across the U.S. to cover higher-ed, its role in local democracies and its effects in local communities. While Smallwood and other Open Campus editors don’t directly supervise individual reporters, they provide support and guidance on reporting, writing, and idea generation.
There, at his dining room table, Smallwood sketched out an idea generation app on paper, took a picture and asked ChatGPT to help him build something he could share with reporters. He wanted to see how the machine learning that powers generative AI could help Open Campus reporters generate new story ideas.
“The main idea was structuring the prompts for reporters ahead of time,” Smallwood said during a recent session for organizations supported by AJP’s Product & AI Studio. “I wanted to create something where they feed in their notes and ideas to see what it comes up with. Some reporters were struggling to figure out how to prompt in a way that was going to get them a functional thing on the other side.”
While editorial staff remain crucial for decision-making, the machine learning that powers generative AI makes sorting through large amounts of data much faster. A single human might never uncover patterns or outliers that AI can often identify within minutes.
Smallwood, who characterizes himself as a hobby-level coder, was able to ask ChatGPT itself for instructions on how to write the code to build a simple tool for reporters built around specific pre-written prompts. It took a weekend.
AI tools can feel intimidating to many in the newsroom. When a new-to-AI reporter simply asks a chat interface for follow-up story ideas, they’re likely to get unhelpful responses. That same reporter throws up their hands, unclear what all the fuss about AI might be. But more experienced users of large language models (LLMs) know that the context and instructions you provide can change the output dramatically.
Once you tell the AI tool more about the type of work you really do (community-oriented journalism for Pittsburgh) and the perspective you take (an accountability-oriented mindset that holds power to account), the results take on a very different flavor. Smallwood’s goal was to create the type of background instructions that can help less-experienced users have more success with their own experiments.
While some see AI tools as a versatile toolbox filled with various tools, Smallwood has discovered that many employees benefit from starting with a simple, task-specific tool. For these cases, he believes creating something like a Custom GPT — an AI model tailored to specific tasks — can be especially helpful.
Using ChatGPT Enterprise, which AJP can provide through its Product & AI Studio, Smallwood has created a handful of custom GPTs. Some help with generating story ideas and others query large databases. The three basic elements of building a custom GPT are Instructions, Knowledge and Action, Smallwood says
- Instruction include the role you want the GPT to play and how you want it to respond
- Knowledge are the files it will use to answer questions
- Actions are what the GPT will rely on to do its job. This can include API calls, getting information from files or the open web.
Fostering adoption in news organizations
When asked what adoption has been like by the newsroom, Smallwood said he uses tactics like office hours to demystify the entire AI space.
“Adoption is mixed,” he said. “But some of my staff use these (GPTs) a lot. My managing editor loves them and uses them all the time.” He also said he noticed new GPTs created by reporters in their ChatGPT Enterprise dashboard. “My heart skipped a beat,” he said. “It made me feel great because that reporter was someone who just came to office hours a couple of weeks ago.”
Giovanni Moujaes, product manager at inewsource, an investigative newsroom based in San Diego, asked Smallwood what helped move “skeptical” staff from understanding to adoption.
Many newsroom reporters have been using the same techniques for years, and have been successful, Moujaes said. So getting them to get the most use out of a technology like this is challenging.
When it comes to LLMs Smallwood said many people don’t know what the technology is capable of, what it’s bad at or good at. Encouraging people to rethink mental models for solving problems isn’t easy, Smallwood said.
“I tell them this technology is not the creator of anything,” Smallwood said. Sometimes he encourages staff to think about generative AI as a fast-typing intern who sometimes uses all that knowledge to come up with a few creative ideas.
The intern might spot something you missed, Smallwood said. But that doesn’t mean you rush out and implement the idea. The intern didn’t assign you anything and neither did ChatGPT.
“Just because it failed at one task, doesn’t mean it’s going to fail at this one,” he said he told reporters. “And just because it was good at one thing doesn’t mean it will be good at every one.”
Smallwood shared a story about a reporter who was skeptical to the point they felt “dirty” when talking about generative AI. Smallwood said he took that reporter’s latest story and uploaded it to ChatGPT then asked for follow-up ideas.
“He was blown away,” Smallwood said.
They followed that up by turning the story into a haiku, then an academic whitepaper, then a movie script.
“I found that was helpful because he could see some of the things it’s good at,” Smallwood said. “That didn’t mean it was going to be good at everything.
Another reporter did an immediate follow up story using the GPT to analyze how public universities across North Carolina were complying with DEI bans.
“People don’t use tools because you tell them to,” Smallwood said. “They use them because it solves a problem they have.”
Want to create GPTs for your organization? Joe Amditis with the Center for Cooperative Media published this guide specifically for local news organizations.
