Poynter: what a Duke journalism class built when failure was the assignment

The AI Journalism Lab at Duke University takes a practical-first approach to AI in newsrooms: identify a specific editorial problem, then match it with the appropriate technology. The course resists the reverse — picking a technology and building toward it — and that choice shaped the outcomes in ways worth studying.

Fourteen students partnered directly with local media organizations to identify real bottlenecks in their editorial workflows. They built five working tools: a grocery price tracker, a city council meeting analyzer, a public records request monitor, a newsletter summarizer, and a cultural events calendar builder. None required deep technical background. Students learned what was necessary to make specific things work, not a broad engineering curriculum.

The most instructive finding was about generative AI’s limits in this context. Many projects converged on traditional automation — rule-based sorting, categorization, and structured summarization — rather than large language models, because hallucination risk made generative approaches unsuitable for tasks where factual accuracy was non-negotiable. Students who tried generative solutions on data-intensive problems discovered this under real conditions, then pivoted. Instructors considered the pivot a success and awarded a “bomb trophy” to the team whose original concept failed and whose revised tool turned out more durable than anything they had initially planned.

The course also makes a quiet argument about pedagogy. When students have an actual newsroom as a client and a specific problem to solve, the gap between AI’s marketing claims and its operational reality becomes apparent quickly. That gap is where the real learning happens, and it produces more experienced practitioners than a classroom built around demos of what AI can do under ideal conditions.

For journalism educators, the course model is transferable: partner with local media, constrain scope tightly, accept failure as part of the process, and evaluate tools by whether they solve the specific problem — not by whether they use the most advanced available technology.

Useful for journalism educators designing AI curricula, and for newsrooms structuring internal AI experiments with low tolerance for wasted effort.