By Aldana Vales, AI Journalism Labs Program Lead
Offered in partnership with the Nordic AI Journalism (NAIJ) network and supported by Microsoft, the AI Journalism Lab: Builders brought together 23 journalists, technologists and data practitioners from around the world. Over four months, they went from pitched ideas to tested prototypes. All with the goal of answering one core question: Can we build better products in news with AI?
This, however, won’t be a recap of what they built. Instead, this is a window into how they did it. These are six lessons on what it takes to build AI-powered products for journalism.
(The Builders met in person at a kick-off in New York City. Photo credits: Sari Goodfriend)
- Understand your user before you build
This is a standard product process, and yet it still needs to be highlighted. Discovery work can take many shapes: listing what technology your users have access to, what devices they use, when they reach for this kind of tool, what they are actually trying to achieve. It all comes down to one simple, empathetic step: put yourself in their shoes.
Suzana Souza, head of Product & Engagement at Reset, did this while building a source-recommendation tool for reporters. By talking to them, she was able to surface the real pain point: “Most of them wanted to find diverse sources, but time constraints pushed them back to familiar, easily accessible experts. This completely reframed my solution from a ‘diversity audit tool’ into a real-time recommendation system embedded in their workflow, rather than another standalone platform,” she explains.
- You need really good data
This was a common realization. Good data is not a technical detail you handle later, but the precondition for everything that follows.
“An AI tool is only as good as the data you feed it. Before any LLM ran, we classified a sample of episodes by hand to refine our editorial taxonomy, cleaned 321 records, and replaced a fragile private data source with a public, sustainable one. With clean inputs and a vocabulary the team had road-tested itself, we built an automated search engine that returns what we mean, not just plausible-sounding guesses,” says Natalia Ramírez Jaramillo, Director of Communities & Product at Radio Ambulante.
- Get feedback from the people who will use it
If speaking with people at discovery tells you where to start, this will let you know if you’re still on track.
Melle Denthre, Digital Innovation Lead at Mediahuis, experienced this as he was building a no-code app to help journalists enrich their articles. “Involving test users was probably the most important, and most humbling, step in the entire process. I was under the impression I had built a launch-ready product, but it turned out I had overlooked some essential things, which became clear the moment people actually started using it,”
- Trust is a design mandate
A tool that the newsroom does not trust will not be adopted, no matter how well it works. Especially when it’s powered by technology that people fear or your tool asks them to give something up.
That’s exactly what Chen Wang, senior data editor at The Globe and Mail, had to keep in mind: “Building a newsroom knowledge-sharing tool taught me that any product asking people to contribute knowledge needs to earn that contribution by making the value of sharing feel larger than the cost of it.”
(After four months, they presented their final projects to the rest of the cohort and reflected on the lessons learned. Photo credits: Aldana Vales)
- Adoption can’t be an afterthought
The difference between a demo and a tool is adoption, and this has to be planned from the start.
That’s how Natalie Holly Purviance, freelance journalist, approached this work: “If a tool adds even a little cognitive load, reporters quietly stop using it, no matter how good it is. So the design problem was making it feel additive: it meets reporters where their work already ends, asks them to confirm what the AI inferred rather than do new work, and keeps editorial judgment with the journalist.”
- Plan for what comes after
What happens to this after the program ends? Who pays for it, who maintains it, how does it survive contact with a real budget and a real organization?
As Bryan Davis, Director of Product at the Associated Press, says, the real test starts once the prototype works, especially when you’re building your own product: “That’s when you push the boat off the bank and find out if it floats with you sitting in it. The challenge then isn’t money or a team yet, it’s whether you keep finding the stolen hours, the late nights and odd free moments, to build the idea into something solid enough to meet this moment. I’m sticking with https://plumbtrace.com/ because I haven’t found a better answer to the question I keep asking every time I sit down to build.”
—
Are these six lessons only applicable to AI? Definitely not. This might just be good product work in the end: start with good inputs, understand the people you are building for, plan for use and, equally important, plan for what comes after. And yet no AI-powered product can be sustained without it. This matters beyond anything that one cohort can produce. It becomes a method that these participants can now apply to anything they build, regardless of the technology powering it.
Watch what five Builders created and keep up with the AI Journalism Lab by signing up for our mailing list.
The AI Journalism Lab: Builders was offered in partnership with Nordic AI Journalism, with the support of Microsoft.