A translation of a mass shooter’s journal written in Faux Cyrillic. A scraper that pulled every public meeting minute from a Texas county website. Tens of thousands of leaked emails from a Chinese surveillance firm, made searchable with a large language model.
Each of those reporting efforts won or was a finalist for a Pulitzer Prize this year. And each relied, in some way, on AI.
Nieman Lab’s Andrew Deck reports that eight Pulitzer honorees disclosed AI use to the judging committee in 2026: five winners and three finalists. That’s the most since the Pulitzer Prizes began requiring AI disclosures in 2024—and it points to a shift in how journalists are using the technology.
In the two years before that, reporters mostly used older forms of machine learning — embedding models for data visualization, for example, or pattern-recognition tools to analyze satellite imagery. This year, commercial large language models did more of the heavy lifting, especially when reporters faced mountains of documents.
The approach was similar across this year’s Pulitzer winners. At The Wall Street Journal, computational journalist John West and his colleagues built a custom scraper to collect records from Kerr County, Texas, after deadly summer floods. They then used an internal tool called WSJPT to summarize every page and flag references to previous flooding events. Reporters still read each flagged section themselves.
“We aren’t obviating the need for human investigation of a pile of documents,” West said. “Instead, we’re trying to sort the pile so the most relevant stuff is right at the top.”
At The Minnesota Star Tribune, engineer Dana Chiueh used an enterprise ChatGPT account to analyze screenshots of a Minneapolis church shooter’s journal. The model identified the writing as Faux Cyrillic and produced a first-pass translation of more than 600,000 words.
The team then used Google’s NotebookLM to identify recurring themes and had two Russian-language academics at St. Olaf College verify key passages. Their review caught several errors, including one that mischaracterized the shooter’s motivation.
The reporting won the Pulitzer Prize for Breaking News.
The Associated Press used large language models to make tens of thousands of leaked documents searchable for its Pulitzer-winning investigation into American technology companies’ role in China’s surveillance state.
But reporters did not simply trust what the models surfaced. They manually reviewed documents flagged by AI, independently checked the accuracy of AI-generated summaries and did not quote from those summaries, said AP investigative journalist Garance Burke.
The New York Times flipped the usual workflow. For its investigation into the Securities and Exchange Commission’s retreat from crypto enforcement under the second Trump administration, reporters manually read and classified more than 10,000 documents. They then used OpenAI’s GPT-5 to conduct a second pass, flagging discrepancies for reporters to review.
In this case, AI was used to check the humans — not the other way around.
For newsrooms, perhaps the more revealing detail is what readers never saw.
The Wall Street Journal did not disclose its use of AI in the flood investigation at publication. West said the tools functioned essentially as a more sophisticated search system.
That gap between disclosing AI use to a prize committee and disclosing it to the public is at the center of an increasingly important debate over newsroom standards. It’s a tension that has surfaced repeatedly as news organizations figure out when AI use is significant enough to tell readers about.
Pulitzer administrator Marjorie Miller said the industry now has a clearer understanding of where AI can be used appropriately, such as data collection and analysis, and where its use raises more questions, including writing and editing.
Next year, after controversies over AI-generated text in prize-winning literary works, the Pulitzers will add an AI disclosure question to their book entry forms.
“AI is here to stay,” Miller said. The committee, she added, will continue asking entrants to demonstrate that a human produced the work.
