Nieman Lab: how newsrooms are using AI agents and personas in investigations
In late September 2026, Hacks/Hackers and the New York Times hosted roughly 150 journalists, researchers, and technologists at an AI x Investigative Journalism Forum at the Times building. The Times’ AI Initiatives team, a group of nine people, co-organized the event. Andrew Deck reported for Nieman Lab on what was discussed.
What the forum covered
The event surfaced several practices already in use at investigative newsrooms. One is what participants called “Uber agents” — large composite AI setups that chain multiple LLMs and specialized tools together to handle different stages of an investigation, from document discovery to pattern recognition to draft synthesis. Another approach involves AI personas: agents configured with a specific research identity, context, and source base that can be pointed at a particular beat or document corpus over time.
The Times’ own AI work came up in the discussion. A Pulitzer-winning investigation into the SEC’s crypto enforcement actions used LLMs to audit case documents at a scale that would not have been feasible for a small team working manually. This year set a record: eight Pulitzer awardees disclosed AI use to the judges, a number that suggests AI assistance in investigations is becoming common enough to normalize disclosure.
Challenges raised
The event was not a showcase. Speakers brought up recurring problems: hallucinations that reach draft form and require close editing to catch, errors introduced by shortcuts taken during development, and the rising cost of running many LLM calls at scale for an organization that is not a technology company.
One researcher framed a point that ran through much of the discussion: AI performs well at surfacing leads — finding patterns, flagging documents worth reviewing, identifying connections across large corpora — but remains weak at producing evidence that could survive fact-checking and publication on its own. It supports the early and middle stages of investigations, not the final output.
Who it is useful for
Investigative journalists and newsroom editors who want a grounded account of where AI stands in practice, not a vendor pitch. The piece covers real outcomes, real failures, and the current state of AI disclosure in journalism awards.