Veteran data journalists urge fellows to use AI to enhance their analysis, not outsource it

The topic of artificial intelligence generates polarizing reactions among journalists. Some love it, others abhor it, a few have already been replaced by it.

The veteran journalists serving as mentors at the USC Center for Health Journalism’s 2026 Data Fellowship have their own misgivings about AI, but agree that no matter how reporters feel about it, every newsroom should be learning how to best harness it.

In a panel discussion on the promises and pitfalls of artificial intelligence, they outlined tips for leveraging tools like ChatGPT and Gemini to deepen and accelerate reporting, while avoiding the dangers of a technology prone to making things up. 

“I do see its power, but I think with any emerging technology that's extremely powerful, we need to really interrogate it, and we need to figure out how it impacts our lives and other people’s lives,” said Yoohyun Jung, co-editor of The Boston Globe’s StoryLab.

Jung was joined on the panel by Washington Post data reporters Andrew Ba Tran and Caitlin Gilbert, and Minnesota Star Tribune data editor MaryJo Webster. 

Collectively they have decades of experience using coding and machine learning for data reporting, and they’re now helming efforts to apply AI at their respective publications. 

“What I've discovered in the last few months is that these tools are now allowing us to do things that we in data journalism have struggled with for decades,” said Webster. “Things I wish I could have done 25 years ago now you can do very quickly with some of these tools.”

But before reporters dive into applying AI to their investigations, the panelists said it is essential they first understand the basic mechanics behind this technology.

Large language models such as ChatGPT, Claude and Gemini are all trained on a massive corpus of text and data documents, explained Webster. They don’t understand the information using logic like a human brain, but rather predict likely responses based on how frequently information appears in the training data. 

Knowing this can help reporters understand what sorts of tasks the technology is best suited for. 

For example, ChatGPT can be a helpful source for suggesting experts to interview for a story. However, reporters should be aware that it will provide the most commonly cited experts and therefore be biased against suggesting more diverse perspectives and underrepresented voices.

I do see its power, but I think with any emerging technology that's extremely powerful, we need to really interrogate it, and we need to figure out how it impacts our lives and other people’s lives.” — Yoohyun Jung, co-editor of The Boston Globe’s StoryLab

On the other hand, when it comes to tasks where you want standardization — such as nailing the precise legal language of a public records request — LLMs are pretty fantastic tools. 

It’s also essential to know that LLMs are designed to always provide an answer, even if the system isn’t confident in the accuracy of the response. Therefore, it is helpful to inform AI tools that you are a journalist and place a premium on vetted, sourced and accurate information.  

For ChatGPT, Webster suggests entering the following instructions in the preferences section:

“I’m a journalist. I want you to give me fact-based information. Prioritize information from reputable sources like news outlets, government records, and academic research over social media. Explain your logic, not just your final answer. Do not guess. If you're unsure about an answer, say so. You're allowed to ask questions when the task is unclear. Include links to sources when obtaining information from web search.”

She also urges reporters to always check the links and citations provided by AI to independently verify information before adding it into a story.

Additionally, it's useful to be aware of what types of tasks different AI tools are best suited for.

ChatGPT has wide versatility for everyday tasks and research, while Claude is great at holding extremely large document and data caches and doing coding and other complex tasks. Gemini Notebook is a great place to store and study your own compiled materials, because it can act as a closed-loop system and will not automatically pull in external information.

One of the most exciting things about AI tools is that they enable reporters, with the right prompting and discernment, to accomplish tasks that previously required extensive training. 

For example, while it will always be helpful to have a background understanding of how coding for data scraping and analysis works, reporters no longer need to dedicate hours to studying how to write languages such as Python, JavaScript and SQL. 

Instead they can use LLMs to “vibe code” by prompting the system with instructions on the kind of code they want it to write and then running it in interfaces like RStudio or Jupyter Notebook. 

Of course, doing so also comes with risks of AI making coding mistakes, which is why it is important to prompt the system precisely and check its methodology manually where possible.

For example, Tran advised reporters to avoid having AI directly perform data analysis tasks and instead ask LLMs for step-by-step instructions on how to perform the data analysis. 

That way reporters can then perform the analysis themselves and, in the event that something goes wrong, they can determine at what step the error was introduced and correct the course.  

“Instead of having ChatGPT go from A to Z, you get the steps for A to B, B to C, C to D, all the way to Z,” he said. “You can then understand the methodology and explain to others everything that happened.”

Panelists shared several other complex tasks where AI can be useful to reporters. 

It can greatly increase efficiency for weekly tasks, such as pulling and searching various meeting agendas for different civic bodies. 

It can also be particularly helpful for searching a large amount of text by theme. For example, in the past a reporter looking through hundreds of police reports may need to command F search several terms such as “shooting,” “lethal incident,” “ammunition”, but now LLMs can perform a single search for all incidents involving gun violence.

Tools like ChatGPT also have a “deep research” mode that can be helpful for digging up background information on people that do not appear in simple social media or news searches.

While it's great to use AI as a partner for brainstorming and for optimizing tasks, it is essential to remember that it cannot use logic in the same way as a human and will always need to be fact-checked.

“You want to use it in a way that is supplementing and enhancing your own thought process and your instincts,” said Tran. “Do not offload those instincts onto AI.”