How can artificial intelligence (AI) be used in investigative journalism? At a session at this year’s Nordic Media Days in Bergen, journalist Martin Folkvord gave the audience a glimpse into how the journalists were able to save lots of time using AI in a huge and complicated investigative project.
But the process was not without stumble stones, as Folkvord explained.
“There is potential for saving enormous amounts of time using AI in investigative journalism. But a human still needs to control everything manually in the end,” he said.
VG is in the lead compared to most newsrooms in using AI. Readers meet the technology in many ways, including AI-generated summaries of most news stories. VG has chosen to be transparent about how it uses AI and has published its editorial AI guidelines.
Portfolio of editorial AI tools
But in parallel its journalists now have a long list of internally developed AI tools designed to help them be more effective and better journalists. One example is JoJo, a tool to transcribe audio and video files which has also been made available to the public. Other examples are “Melderen,” which writes a news story based on the content behind a link, or “Innsynsboten,” which assists the journalist in writing an appeal after being denied access to a government document.
Martin Folkvord is part of VG’s investigative reporting team. Since January 2023, the team has published a series of articles documenting illegal adoptions to Norway. In one of the articles, VG documented that Norwegian authorities had received at least 88 warnings about possible illegal adoptions.
6,000 pages of documents
At Nordic Media Days, Folkvord explained how the journalists used AI to find these warnings.
“We had 6,000 pages of documents. We knew there were many warnings in between the pages, but to go through them all with that specific purpose in mind would take months,” he explained.
Could ChatGPT do the work more efficiently?
Martin started by bringing all the 6,000 pages into one huge document which he uploaded to ChatGPT. Through different prompts, he tried to ask the chatbot to identify warnings of illegal adoptions. However, the exercise did not produce any results. ChatGPT did not find indications of illegal adoptions.
Built his own GPT
Martin decided to go one step further – to build his own GPT. This was introduced as a new functionality in ChatGPT last November.
Again, he had to experiment with different instructions for his GPT. But this time he did not ask it to find indications of illegal adoptions. Instead, it was given seven indications of illegal adoptions and asked to review each document for each of the indications.
“We realised that ChatGPT is not good at doing evaluations, for instance determining if an adoption is illegal. But it is good at summarising documents,” Folkvord says.
The new approach finally gave meaningful results. The journalists still struggled with the large number of documents, however.
Developed script to automate
Fortunately, VG has a team of very competent software engineers who are at the forefront of using AI and other new technologies to enhance the power of journalism. One of them, Jari Bakken, developed a script that automated the process using ChatGPT’s API and presented the results in a structured format.
The script ran the whole night, processing the 6,000 pages of documents, and the journalists were presented with 900 findings laid out in a Google sheet.
After spending four days systematically reviewing the findings, the journalists concluded that the documents included 88 warnings of illegal adoptions.
Saves time
So what were the learnings from this project? How will AI influence investigative journalism?
Martin Folkvord underlines that both the start and the end of this particular project involved massive manual work. The journalists had to ask for each of the documents individually using the Norwegian Freedom of Information Act, and they had to spend days carefully reviewing the results from ChatGPT.
“The biggest potential we see so far is that AI can be helpful in taking massive amounts of unstructured data and giving us an output that is structured. This can save a lot of time. But it still requires us as journalists to know the material and the documents well enough to ensure that the final result is up to the highest standards of VG’s journalism,” Martin says.
Five takeaways
Some of Martin’s takeaways for using ChatGPT as a journalist:
- ChatGPT is NOT good at evaluating
- ChatGPT, however, IS good at summarising
- Give ChatGPT all premises if you are looking for something specific
- Don’t make prompts too long – and be specific
- You must still review everything manually
