Since 2023, the Pulitzer Prize - the most prestigious award in the US press - has required candidates to publicize the use of artificial intelligence (AI) in their entries. This season, one winning work and three finalists have revealed how they use AI in their investigation.
Elon Musk and the political language map
This year's Pulitzer Prize-winning work for the National Report category of The Wall Street Journal attracted attention with a unique analysis: Monitoring and visualizing the changes in Elon Musk's statements on the old Facebook platform X. After buying the social media platform, Elon Musk has used his personal account - with more than 200 million followers - as a tool to demonstrate his political leanings.
To conduct the analysis, the team of reporters began with a dataset from the Clemson University Digital Communications Center, which included more than 41,000 Musk posts and interactions since 2019. From there, they apply the technique of converting text and images into vectors - a machine learning method that helps encrypt content meaning into digital values.
The results show a clear shift: From Tesla-related content, technology and Internet culture to dividing topics such as immigration and politics. This shift really shows that Musk has gone from businessman to political figure, data journalist John West said.
West also stressed that his team does not use generative AI but uses embedding models (bottlenecks) of images and text developed by Google. We want to understand the memes Musk posted a field where computer vision is much better than it was a few years ago.
Forgotten historical records: 40 land models and lies
The work 40 Acres and a Lie (40 soil samples and lies) - which reached the final round of the Explanatory Reporting category - is the result of cooperation between investigative press organizations such as Reveal and Mother Jones. Based on nearly 2 million digital documents from the post-Civil War reconstruction period, the reporter team traced more than 1,000 people who were arrested as slaves and were granted land and then deprived of it after a short time.
The special thing is that many important documents handwritten in Spencerian - a popular form of calligraphy in the 19th century - are difficult to read and have not been indexed. Data journalist Pratheek Rebala has built an image identification algorithm, trained from verified documents, to Read this huge amount of document.
As a result, the team discovered about 500 more individuals who had been granted land - double the number previously found. These documents became a platform for the group to trace the lineages of those who once owned land and interview them about their stolen heritage. AI helps us expand the scope of the investigation, touching untold stories, said reporter Alexia Fern Fernards.
When war is inaccessible: AI and remote investigation
Unable to enter Gaza, The Washington Post reporters used satellite images, drones and AI to verify the Israeli military's explanation after the attack that killed two Al Jazeera journalists in early 2024.
Based on the object detection model provided by the military AI company Preligens, the investigation team analyzed satellite images and did not find any military vehicles within a radius of 16km around the scene at the time of the attack - denying the claim that the Israeli army was targeting a threat from flying equipment.
In the context of Israel's ban on most foreign reporters entering Gaza, news agencies' application of computer vision - such as identifying tanks, craters to bombs or locations of artillery - has helped strengthen the role of remote information verification.
AI does not replace on-site investigation, but it is an extremely useful tool when the battlefield is closed to the press, commented a member of the investigation team.
Building a national database from handwritten documents
After George Floyd's death in 2020, the AP began investigating deaths from lethal subvention such as strangling and hand tying which are rarely recorded in the official data system.
The Lethal Restraint project - which reached the final round of the Investigative Press category - lasted 3 years, collecting more than 200,000 pages of digital documents, including court records, forensic reports, death certificates, most of which were poor quality handwritten or scan. AI plays a key role: The group uses optical symbol recognition (OCR) tools such as Amazon Textract to digitize and extract information.
Sean Mussenden - data editor at the Center for Journalism Investigation at the University of Maryland - said that AI results are never published directly but have gone through many rounds of verification, correction and professional assessment. Data is just the foundation. Finding meaning is in the hands of journalists, he affirmed.
AI does not replace journalism but expands its borders
Although the AI tools used in the above series of works are not generative models like ChatGPT, their impact on the press making process is undeniable. AI, when used responsibly, is adding flexibility, depth and accuracy to journalism, which was not possible 10 years ago, says Marjorie Miller, President of the Pulitzer Prize Council.
In the context of AI raising many questions about ethics, accuracy and copyright, this year's winning works show another direction: Using AI is not to "replace journalists", but to help journalists access truths that cannot be seen by human eyes, cannot be reached by human hands, whether in the depths of historical times, or in the war zones covered in news.
Nowadays, an investigative journalist not only needs a pen and a recorder, but may also need a powerful set of AI tools. But like all other technologies, what determines the value of a press work is still the heart, mind and honesty of the writer.
