Newsrooms Deploy 'Uber Agents' for AI Investigations

Overview

Newsrooms across the United States are increasingly integrating sophisticated AI tools into their investigative workflows, moving beyond simple search functions to deploy complex autonomous agents. This shift was the central theme of the AI x Investigative Journalism Forum, hosted recently by Hacks/Hackers and The New York Times at the paper’s headquarters in New York City.

The event drew approximately 150 journalists, researchers, and technologists interested in how machine intelligence is reshaping the landscape of hard-hitting reporting. With a record eight Pulitzer Prize-winning teams this year disclosing AI assistance in their work, the industry is witnessing a pivotal moment in how large-scale document analysis and lead generation are handled.

The Rise of 'Uber Agents' and Deep Integration

A significant portion of the forum focused on the evolution of AI from a passive search tool to an active participant in investigation, often described as "Uber agents" or undercover AI personas. These systems are designed to not only retrieve information but also to synthesize, cross-reference, and interpret massive datasets without constant human intervention at every step.

“At the Times we’re not doing ‘robots tell the news,’ but a number of different [kinds of AI adoption] that can make the work we do more efficient, more effective, and able to do things that otherwise would be very, very hard to do.” — Juliana Castro Varón, Senior Design Editor of AI Initiatives, The New York Times

Castro Varón emphasized that the most valuable aspect of these tools lies in their multimodal capabilities. While generative AI is often praised for creating text or images, its true investigative power lies in its ability to understand and connect disparate forms of media, including complex legal filings, financial records, and archival photographs.

The discussion highlighted that these "Uber agents" allow reporters to pull threads from enormous corpora of data that would be physically impossible for a human to review manually. By automating the initial organization and filtering of evidence, journalists can focus their energy on verification, source development, and editorial judgment—tasks that remain fundamentally human responsibilities.

Lessons from the Epstein Files Release

The January 2026 release of the Epstein Files served as a catalytic case study for many attendees. The Department of Justice dropped over 3.5 million pages of documents, along with tens of thousands of images and videos, creating one of the largest public data dumps in recent history.

Duy Nguyen, who leads AI science research at The New York Times, illustrated the scale of the challenge by noting that the PDFs alone would stack as high as the Empire State Building if printed. For a single reporter, reading the text would take eleven years, while reviewing all audio and video clips would require another month of continuous work.

In response, newsrooms deployed specialized engineering teams to build custom tools:

1. NPR’s Internal Search Tool: Designed by the NPR AI Labs, this system allowed reporters to find relevant documents across various beats. It classified files by type, grouped them by location, and extracted entities to ensure searches for one name, such as “Prince Andrew,” also surfaced references to “the Duke of York.”
2. The Epstein Files Engine: Built by The New York Times, this internal chatbot enabled reporters to dive deep into the corpus using natural language queries. The interface was constructed using LibreChat, an open-source platform, allowing for rapid deployment and customization.

Reporters used the Times’ engine to execute complex prompts, such as cross-referencing existing coverage with newly released names to identify gaps, or constructing detailed timelines of events with cited witnesses and quotes. More than 100 Times journalists posed over 5,000 questions to the engine, demonstrating the scalability of AI-assisted investigation.

Balancing Promise with Persistent Limitations

Despite the enthusiasm surrounding these technological advancements, the forum did not shy away from the significant limitations and risks inherent in current AI systems. Hallucination remains a critical concern, as large language models can generate plausible-sounding but factually incorrect information.

Afrooz Mosallaei, a research associate at the Center for News, Technology & Innovation, cautioned that while AI excels at generating leads, it is far less capable of establishing publishable evidence. She noted that tasks such as textual understanding and verification require rigorous human oversight to avoid errors that could compromise journalistic integrity.

Other challenges discussed included:

• Vibe-coding errors: The risk of relying on code generated by AI without thorough understanding or testing, which can lead to subtle but damaging bugs in investigative tools.
• Budget constraints: The rising costs of AI tokens are placing pressure on newsroom budgets, forcing editors to weigh the expense of advanced model access against other resource needs.
• Ethical and transparency concerns: With eight Pulitzer awardees disclosing AI use this year, the industry is establishing new norms around transparency, yet questions remain about how much AI involvement should be disclosed to the public.

The consensus among speakers was that AI is a powerful assistant rather than a replacement for the investigative journalist. The tools are best utilized when they amplify human capabilities, handling the heavy lifting of data processing while leaving critical analysis and ethical decision-making in the hands of trained reporters.