Inside The New York Times’ AI newsroom strategy

Onstage at the Digiday Publishing Summit in Miami, Fla., last week, Seward outlined how his team is working with reporters, what tools they’ve built and what the key use cases for AI are so far. Seward has a team of eight, including four engineers, a product designer and two editors. 

Using AI for research and investigations is “by far the biggest use of our resources and I think the biggest opportunity right now when it comes to AI in media,” Seward said. His team mostly works by helping a reporter use AI technology for one project, and then creating a repeatable process from that experience for others in the newsroom to use.

Building the ‘Cheat Sheet’ AI tool 

A Times reporter came to Seward’s team with an impossible task – 500 hours of leaked Zoom recordings of an election interference group to go through before Election Day. AI tools were used to transcribe 5 million words and identify parts of those transcripts that were of interest to the reporter.

“[The election interference group] wasn’t so dumb as to say, ‘We’re going to spread misinformation on the internet’… then you could Control F” to find that information in the transcripts, Seward said. “Where AI becomes useful – typically it’s referred to as semantic search, or sometimes vibes-based searching – where you’re looking for topics, concepts, things that are similar. And that’s hugely useful when looking through enormous corpuses of text,” Seward said. That led to a big story before the presidential election last year, he added.

Those efforts were then developed into a spreadsheet-based AI tool built internally called Cheat Sheet. Reporters can pick (with the guidance of Seward’s team) which LLM model to use with Cheat Sheet, Seward said. It’s now being used by several dozen reporters.

Seward declined to share other specific AI tools The New York Times newsroom was using, though he said it was “pretty much all the commercial AI providers as well as open source models.”

Cheat Sheet also helped a reporter who had an unorganized list of 10,000 names of people who had registered for a tax cut in Puerto Rico.

“You can’t Google 10,000 names… but a computer can Google 10,000 names. And then using AI, we could analyze those search results for certain markers that [the reporter] was interested in,” Seward said.

Even though the results weren’t entirely accurate, it helped sort the names into more promising leads. The reporter could then call them up and continue reporting out the story, Seward said.

Seward’s team’s approach is to took on an individual reporting challenge – “knotty, huge, messy data sets” with an “immediate deadline” – “but always with an eye toward building up tooling that will make that repeatable in the future,” he said.

How Seward’s team works with the newsroom

How does Seward’s team decide which AI tools to build and where they can help the newsroom?

Seward said it comes down to constant communication with the newsroom. His team hosts training sessions on how to use AI tools for research investigations. Seward’s team has spoken to 1,700 of the 2,000 people in the newsroom so far, he said.

The New York Times also has an open Slack channel that anyone from the newsroom can join to ask questions and share use cases – ranging from “how can I get Gemini?” to one bureau chief inspiring another across the world with an idea for how they’re using AI technology.

“AI… is such a personal technology,” Seward said. “The way people would describe what they want out of AI can be different to the person.” Many have experienced “writer’s block in the chatbot… With a tool that can do anything for me, sometimes the challenge is… what can it do? And so we’re just trying to help answer that question,” he said.

Dealing with skepticism from the newsroom

AI isn’t being used to write articles at The New York Times, Seward said. Reporters are allowed to use AI to draft copy around published articles, such as SEO and headlines, he said. 

Seward said he reminds editorial staff to “never trust output from an LLM. Treat it… with the same suspicion you would a source you just met and you don’t know if you could trust.” 

Some newsrooms have challenged upper management about using AI vendors to generate articles, or fought back when policies are rolled out that push for more use of the technology from editorial.

Seward team deals with any skepticism from reporters about using AI by acknowledging their reservations, and showing them how the technology can be useful, Seward said.

“We’re not trying to be AI boosters. In fact, quite the opposite. I think there’s a lot of caution. A lot of time we spend cautioning people about uses of AI, both [in the] legal and editorial senses,” he said. “But if we can have you leave a session saying, ‘I’m still pretty concerned about this whole environmental issue and maybe like destroying humanity thing – but in the meantime, it’s going to let me transcribe handwritten notes in Arabic that I took messy iPhone photos of while I was on a reporting trip, and that’s pretty cool.’ And no reporter is going to say no to a competitive advantage, which I think is the theme of what we’re trying to build for them.”

What’s the biggest challenge Seward faces in his newly-created role at The New York Times?

“I definitely live in fear of an error that is in some way attributable to AI. To be clear, we also say in sessions with our newsroom we would never attribute an error to AI, meaning it’s always on us,” Seward said. 

“I would 100% feel responsible” if something like that happened, he added.