It is only a little over half a year since the Society for Advanced Business Editors and Writers (SABEW) last convened a panel on AI tools for journalism. But the pace of change in what generative AI can do, and how business journalists are using it, is rapid.
Even a year ago, reporting applications for generative AI beyond the automation of simple tasks like transcription were limited in scope or relatively inaccessible. And concerns over accuracy, hallucinations and ethics further hindered the widespread use of AI.
While those concerns remain, powerful new features on platforms such as Claude, Gemini and ChatGPT have given journalists tools to dig deeper, work faster, and acquire new specialisms.
With that in mind, the Society of Business Editors and Writers (SABEW) recently hosted a panel on the topic in collaboration with the National Association of Hispanic Journalists. The February 17 panel came just six months after SABEW’s last AI-focused workshop, illustrating the rapid development in AI and the need for business journalists to keep up-to-date on developments.
Moderated by Forbes Senior Editor Jeff Kauflin, the panel included three business journalists who shared how they’re using these tools today: Isabella Cota, Mexico City journalist at the International Consortium of Investigative Journalists (ICIJ); Mago Torres, training director at the Google News Initiative and former data editor at The Examination; and Sebastian Herrera, a technology reporter for The Wall Street Journal.
NotebookLM
One key tool being rapidly picked up in newsrooms is Google’s NotebookLM, a specialized research platform used to probe huge volumes of information from various sources and in various file types. Unlike more general-purpose large-language models like Google’s Gemini or ChatGPT, NotebookLM is designed to pull answers only from the resources the user adds directly for each project.
Researchers at IBM have described the difference as equivalent to an open-book versus a closed-book exam. Tools like NotebookLM “respond to a question by browsing through the content in a book, as opposed to trying to remember facts from memory.” As a result, they throw up significantly fewer hallucinations, and provide citations for which line of which document each answer is drawn from. (For this article, The Reynolds Center used NotebookLM to probe the transcript of the SABEW panel, asking for each mention of different AI tools. NotebookLM pointed to the timestamps of each quote, so its accuracy could be checked.)
As a tech reporter in the Bay Area, Sebastian Herrera covers AI companies extensively. But it took his first use of NotebookLM last Fall for AI to become useful “in a big way” to his reporting. After database computing company Oracle’s surprisingly buoyant earnings call in November, Herrera’s Wall Street Journal editors wanted him to write a profile on the company’s founder, the increasingly important Trump ally Larry Ellison. And they wanted it in a couple of days.
“I had never written about Larry Ellison,” Herrara told the SABEW panel. In effect, he was starting from scratch on understanding the 81-year-old’s extensive business career and background.
At the suggestion of editors, Herrera created a NotebookLM “notebook” — the web app’s name for a project — and fed into it a series of Ellison’s prior media appearances, including magazine articles, news clippings, and YouTube videos.
Then he started asking questions: “What has Larry Ellison said about Bill Gates, because they were big rivals back in the day” and “What has Larry Ellison said about running a company into his 80s?”
“I was able to put it together pretty efficiently using my own research, my own quotes, with a lot of other background that I had gathered,” Herrera said. “It really sped up the reporting process for me.”
Mago Torres has also used NotebookLM for help crafting story ideas around breaking news. When reports came out of a drunken raccoon raiding a liquor store in Virginia, Torres uploaded some quick-turn news stories into a notebook and asked for different angles for deeper reporting around it.
NotebookLM gave some suggestions, including a question based on the best ways to humanely treat an animal in this situation. “And that was super interesting!”
Deep research
To get the most from NotebookLM, it helps to have already collated a wide range of sources. And journalists are beginning to make use of generative AI to do this, too.
The three leading LLM platforms, ChatGPT, Claude, and Gemini, all now have “deep research” modes. Deep research splits up the model’s thinking pattern into two stages and gives it far more time to work.
In the first stage, the model employs a psychological tool described in Daniel Kahneman’s 2011 book “Thinking Fast and Slow.” Humans have two thinking systems, Kahneman writes: “System 1” is used for simple tasks like brushing your teeth, while “System 2” takes over for more challenging tasks, like solving a complex logic problem or making a difficult decision. For a simple query like “who was president in 1837?” AI models stay in their “System 1,” and spit out “Martin Van Buren.” But for a deep research query, the model will first enter its version of System 2 to break down the steps it needs to take to produce the right answer.
Once its plan is defined, the model will hunker down for as long as 30 minutes or more, digging up information and outputting it in a useful report. (For this article, The Reynolds Center used the deep research mode in Anthropic’s Claude to find academic papers, blog posts and press articles covering the structure of different AI systems.)
Herrera has used deep research queries to identify examples of people building their own software tools with AI for a recent article on selloffs of software company stocks, “to get a lay of the land of what’s out there.”
ICIJ’s Cota, meanwhile, has used deep research to look for potential human sources for articles, surfacing people that “I didn’t find on a normal Google search.” But she reports more useful results out of “more sophisticated questions,” like quizzing a model on what companies are the primary buyers in a given industry.
Active prompting
For all these tools, understanding and using the full spread of available prompting techniques can result in more targeted answers.
Google News Initiative Trainer Torres points to five elements available for use in a good prompt: tasks, personas, audiences, formats, and restrictions.
Every prompt has a task. The task in “who was president in 1837” is simple trivia, but the task could be more complex, like “make me a research report,” or “spot trends in the data file I provided,” or “using this company’s quarterly earnings report, pull out all potential risk factors in the coming months.”
Personas and audiences help inform how the model thinks and how it communicates its findings. Telling the model “I’m a CPA accountant” will lead it towards giving different outputs than saying “I’m an investigative journalist.” And asking it “describe this to me like I’m six” will give different outputs than “describe this to me like a PhD-level expert.”
Telling the models which formats you want it to output in can also be helpful. This could be as simple as asking only for bullet-point answers or responses formatted in a table. But leading models have become more and more flexible in their output formats — including the ability to visualize data in graphs.
Finally, restrictions can be used to tell the model not to gather information from certain types of sources, or to limit its search only to government press releases.
You can even use the models themselves to improve your prompts, Torres said. English is her second language, so she finds Gemini can help sharpen the prompt she’s asking in its “System 1” thinking, before she sends it on to a deep research query.
A note of caution
ICIJ’s Cota struck a more cautious note in the panel. “I feel very strongly about being cautious and even a little bit skeptical of generative AI,” she said.
One major concern is about protecting sources and confidential information. Tech teams at the ICIJ have for years had security-focused non-AI tools for analyzing confidential data sets, and Cota warns, “If this is the document that has been leaked, I will never expose it to one of the open-source tools” or online LLMs.
Accuracy concerns remain top-of-mind, and the need to check and “triple-check” AI outputs was raised several times in the panel. For Cota, that means avoiding overreliance on AI as a journalist, since working on complex investigations requires becoming a true expert in the field you are investigating. “I think AI can give us the illusion that we’ve become experts so quickly because it can just give us the answers,” she said.
“To me, that’s a risk.”
