New AI and Large Language Model Tools for Journalists: What to Know – Global Investigative Journalism Network

Image: Shutterstock
Generative AI tools have often been seen as a poisoned chalice for journalists. Responses from large language models (LLMs) are often biased or unreliable, and image generators present major ethics and copyright problems — in addition to being notoriously unreliable when rendering human fingers and hands. Meanwhile, AI chatbots are so strange that, according to recent research, request prompts that offer cash incentives, or even “physical” threats, can persuade these robots to give you better results.
But data journalism experts say the smart, careful use of free or affordable AI tools can also be a dramatic help to the front end of investigative projects and a boost to data efficiency, subject briefings, code training, cost saving, and public document requests. In particular, they say reliable AI tools that produce data code and spreadsheet formulas already allow data reporters to spend more time on journalism and less time on tech troubleshooting and coding work.
In several sessions on AI tools at the 2024 NICAR data journalism conference, data experts grappled with the broader question: “Why should journalists use AI?”
During the week of the summit, Bloomberg published an investigation by two NICAR attendees that illustrates some of the problems with LLMs, showing that ChatGPT assigned lower rankings to identical work resumes featuring applicant names typically associated with minority groups.
Jon Keegan, an investigative data reporter at The Markup, said reporters and editors need to be systematic in their approach to AI to benefit from its efficiencies and avoid its many accuracy, bias, and ethics pitfalls, including several basic steps.
“These tools are not a stand-in for journalism, but they really can make your projects faster or more ambitious,” Keegan explained. “But there are some shocking ethical failures in some of these models. These are very weird tools — one moment they’re amazing; the next, they come up with something shockingly stupid, or just make things up.”
In a separate session on AI-Driven Tools, Tips, and Tricks to Empower Your Data Journalism, four experts discussed the pros and cons of this technology for journalists and shared a short list of tools that they felt comfortable recommending.
Speakers included Darla Cameron, managing editor for visual journalism at The Texas Tribune; Jeff Hargarten, a data journalist at the Minneapolis Star Tribune; June Kim, a journalism fellow at MIT Technology Review, and data journalism lecturer Mike Reilley, founder of Journalist’s Toolbox.AI
Journalist’s Toolbox.AI. Image: Courtesy of NICAR slideshow
“I am excited about how AI might help us scale and solve some of the really big challenges we face as data journalists,” Cameron said. “But we hear from tech companies all the time: that ‘you should use this AI tool; it will help you do better journalism.’ Well, we journalists are the best advocates of our journalistic needs. Let’s first define our needs, instead of widely trusting vendors to tell us what our needs are.”
There was consensus among NICAR speakers about what AI use cases journalists should generally avoid:
Reilley said one way to reduce the risk of biased results was to demand diversity and authoritative sources in prompts to AI chatbots. “Ask an image generator to generate images of prison inmates, and it tends to give you people of color; ask an LLM to write a love story, and it will write a heterosexual love story,” he warned. “We have to be specific and demand diversity and accuracy in our prompts. Over-explain everything.”
However, speakers also listed numerous examples of how LLM chatbots and niche AI tools had helped their investigations — especially with code and the retrieval of scattered data sources.
“In 10 years of data journalism, I would say easily 60% of my time has been spent troubleshooting and trying to figure out coding solutions and data analysis solutions that some of these tools can do for you in seconds,” said Hargarten. “I can focus more on journalism craft and less on tech.”
Here are some of the AI tools the panelists recommended.
Created for journalists, by journalists, Rolli tools use AI technology to solve key reporting challenges while keeping project data secure.
The new Rolli Information Tracer — built in collaboration with the International Center for Journalists’ LEAP Innovation Lab — uses algorithms to track disinformation campaigns on social media. The site describes its approach to detecting coordination this way: “We develop an ensemble of technical signals to identify inauthentic behaviors behind online campaigns. We work with research institutions and humanitarian agencies to monitor misinformation narratives across the globe.”
“I’m really excited about Rolli,” said Reilley. “Information Tracer tracks all kinds of disinformation spreaders, which is really important right now in a big election year. It’s also adding some tech which will break the analysis right down to the account level, where you’ll be able to see if it’s a bot and which country it originates from. Very cool.”
He added: “Sign up for the ‘press pass’ free plan for a year. If it asks you for a code, enter ‘JOURNOAI.’”
Reilley said an allied tool, Rolliapp, provides a useful expert-search database and helps connect reporters to expert sources on deadline.
Rolliapp AI tool. Image: Courtesy of NICAR slideshow
“This tool is there to help you write Excel formulas,” Hargarten explained. “It spits out some of the most complex formulas that I’ve ever used, that would take me hours to write if I tried to do it myself. I like it because it’s simple, and doesn’t require as much prompting. It’s got settings for Airtable and Google Sheets as well.” He added: “It bypasses a lot of the extremely irritating troubleshooting that would otherwise slow down an in-depth enterprise investigation.”
“This is the AI tool that has changed my life the most,” Hargarten continued. “You can prompt it to spit out any code you want. I’ve used it to write Python scrapers, just by putting in the URL I want to scrape and a sample of the HTML structure page, and it creates an almost flawless scraper right off the cuff, with very little editing needed. It can also troubleshoot complex Mapbox issues.”
He added: “It’s got the entirety of GitHub behind it, and so far it has not been stumped by anything I’ve given it.”
Keegan agreed: “GitHub Copilot is one of the great tools that journalists really can start using today. It’s a nice drop-in tool; you can summon a little window, you type in your command, and it just gives you the code.”
Speakers said using the tool also helps reporters learn coding skills. Trials on GitHub Copilot are free, followed by a cost of $10 per month.
GitHub CoPilot AI tool. Image: Courtesy of NICAR slideshow
“Whenever I use LLMs with my data students, I’ll have them use three and pit them against each other to see which one is the best for that task,” Reilley explained.
“There is an art form to prompting ChatGPT, and it’s definitely a big broad tool – some things it’s good at, others not,” said Hargarten. “But ultimately it can do some version of almost anything, as long as it’s language-based. I use it to generate perfect public document requests in the right format, citing the correct legislation and everything. Sometimes it will even sign my name and enter my contacts. That’s also terrifying, of course.”
This new chat interface impressed several experienced investigative journalists at NICAR 2024 as an excellent briefing tool for complicated or unfamiliar subjects. While not a direct source of verified data, it provides concise and mostly reliable responses to complex investigative questions, as well as curated lists of authoritative sources and useful suggested follow-up questions. In contrast to some other AI tools, it tends to take you where you want to go, rather than where it wants to take you, according to its early users.
“Perplexity gives you the ability to understand complex questions, with citations and relevant onward questions,” said Jeremy Caplan, director of teaching and learning at  CUNY’s Craig Newmark Graduate School of Journalism. “It’s beautifully rendered — it gives very specific examples, and it’s a very useful tool for quickly getting up to speed.”
Caplan added: “What I really like is that it offers a series of questions to follow up on. You gradually gain a richer and richer understanding of a subject, while retaining the ability to verify at any stage.”
One major concern about commercial LLM chatbots is uncertainty about the security of source data with these cloud-based services, and another is the lack of personal control.
A true technological marvel, the Mozilla Innovation Project has developed an open source solution that allows an LLM — remarkably — to be stored entirely offline, on a single computer or USB stick, under your direct control.
In his blog review, tools expert Simon Willison said Llamafile was essentially like “your own local copy of ChatGPT.”
“This is a single binary file which you can download and then use, forever, on (almost) any computer,” he wrote. “You don’t need a network connection. Stick that file on a USB stick and stash it in a drawer as insurance against a future apocalypse. You’ll never be without a language model ever again.”
Said Keegan: “This tool is built on other people’s work, but it allows you to have a single executable file on your computer that is a self-contained LLM with a chat interface. It does not require any internet connectivity at all.”
Keegan confessed that he initially found it hard to believe that a 4-gigabyte thumb drive could hold LLM-scale computing power, but said it is enabled by complex processes such as “vectorization,” and that it works well.  “You can use this to feel safer about information out there.”
While limited to the US, the new Commons Project tool shows how data journalists in most countries can use AI chat interfaces to build their own data tools to parse large government websites. In this case, journalists built an AI tool to dig into the vast US archive of what citizens have stated about government regulations. It was built in a collaboration between the Columbia University Graduate School of Journalism and Stanford’s School of Engineering.
“It’s a tool powered by GPT that helps journalists parse through public comments about federal regulations posted on regulations.gov,” explained Kim. “If you’re a beat reporter, you want to know what people are saying about new regulations. The data is there, but it’s difficult to sift through 200,000 comments. This tool helps you answer questions like: ‘Which organizations are commenting the most on the EPA’s new emissions regulation?’”
Kim — who maintains the tool — said she had no previous personal experience as a tools developer. She said the process of building the system revealed both the power and surprising limits of LLM helpers — as well as the need to communicate directly with human developers of commercial AI tools.
Commons Project AI tool. Image: Courtesy of NICAR slideshow
“Talk to the developers: ‘How did you make this? What are the problems?’” she advised. “Before we built our own tool, we didn’t really understand the very specific ways in which these big AI tools are flawed.”
Adds Cameron: “Ultimately, for these things to be reliable for journalists — and to feel comfortable that the material is not copyrighted — we as journalists will likely need to create our own AI spaces.”
Rowan Philp, senior reporter, GIJNRowan Philp is a senior reporter for GIJN. He was formerly chief reporter for South Africa’s Sunday Times. As a foreign correspondent, he has reported on news, politics, corruption, and conflict from more than two dozen countries around the world.
This work is licensed under a Creative Commons Attribution-NoDerivatives 4.0 International License
Republish our articles for free, online or in print, under a Creative Commons license.
This work is licensed under a Creative Commons Attribution-NoDerivatives 4.0 International License
Experts agree that AI-driven audio deepfakes could pose a significant threat to democracy and fair elections in 2024.
This roundtable will conclude the data track of the conference with a discussion on the tools that are coming to the forefront for doing data stories and with predictions for what is ahead. Join us with your ideas for what is next. ———————– The Global Investigative Journalism Network is an international association of journalism organizations […]
Two reporters whose investigative work has exposed systemic land grabbing and illegal mining in the Amazon share their tips.
Offering insights on how to evaluate and understand the outcomes provided by publicly accessible AI detectors that are available for free or at low cost.

source

Leave a Reply

The Future Is A.I. !
To top
en_USEnglish