On April 13, more than 500 CEOs and other power brokers gathered in Washington, D.C. to join Semafor World Economy. Across the five-day event, hundreds of speakers took the stage, including Goldman Sachs President John Waldron, Singapore’s Deputy Prime Minister Gan Kim Yong, and nine sitting U.S. cabinet officials.
In the weeks since the event wrapped, Semafor has been experimenting with how AI models can help produce new insights — and editorial content — about Semafor World Economy. The result is a new AI-assisted editorial product called Semafor Intelligence. Drafted by journalists, the report is based on an analysis by a custom-built AI tool. The first edition boils down countless hours of transcripts from the flagship event into nine key themes about the global economy and where it’s headed.
Each banner topic — including supply chains, the Iran war, and the AI race — leads to a bottom-line analysis and relevant quotes from onstage speakers.
“This is AI doing something that’s really hard for people to do,” Ben Smith, co-founder and editor-in-chief of Semafor, told me. “Even if we weren’t moderating interviews, editing stories, talking to sources, and doing all the things journalists do [during an event], it would be impossible to really consume all the information — all that speech — and walk away with a really clear sense of everything that was said and what the prevailing opinions were.”
Smith is quick to clarify that Semafor Intelligence is not an AI-written product. While the tool initially output more than nine central themes, journalists reviewed, consolidated, and curated the final list. Humans also wrote and edited the copy. Each quote in the report links to timestamped YouTube videos of the event, ensuring accuracy.
“AI tools are incredibly powerful, but we’re also intensely aware of our responsibility to our audience that we’re giving them really high-quality material that is not hallucinated,” said Smith.
In a thread on X, Reed Albergotti, Semafor’s tech editor, said that he built the first iteration of Semafor Intelligence in less than an hour using OpenAI’s coding agent, Codex. That prototype was later refined and tested with help from Alastair Clements, Semafor’s senior director of data and insights. The current version of the tool leans on several different machine learning models, including an embedding model from Voyage, an AI company owned by MongoDB. Embedding models can convert text datasets into vectors — lists of numbers that capture the meaning of each piece of text. For journalists, these vectors make it possible to map out the ideas in that dataset and see which themes naturally cluster together. This process, called “vectorizing,” has been used by other data journalists to analyze giant text corpora, including Elon Musk’s entire tweet history.
While these past reporting projects might have taken a team of journalists weeks or even months to conduct, Semafor claims it built its custom tool for Semafor Intelligence in a matter of days. Meanwhile, the whole product pipeline only cost a few hundred dollars to run, according to a more detailed blogpost about the methodology published on Semafor.
“Nobody sat through all 250 sessions, and nobody could have read all 250 transcripts and analyzed them in this way,” wrote Albergotti on X. “AI created a lot more work for us but also allowed us to give readers something valuable.”
The new tool will likely help the newsroom push out more editorial content to support its growing events business. (According to The Wall Street Journal, roughly half of Semafor’s $40 million in revenue last year came from events.) Semafor Intelligence is slated to create similar reports about the next editions of Silicon Valley & The World and The Next Three Billion, the publication’s two other signature events.
“Our audience doesn’t want AI slop, and they also don’t want human slop, which there is also an enormous amount of on the internet,” said Smith. “They expect really high-quality analysis and whatever tool we’re using to get there is secondary.”
