Skip to content
Article Poynter Mar 2026

Poynter: What journalists who built AI tools learned by actually doing it

Alex Mahadevan’s article for Poynter reports on the AI x Journalism Day session held at SXSW, where the organizing theme was the gap between discussing AI and actually building with it. The speakers had already moved past the debate and were sharing outcomes.

The examples from the session

At Pew Research Center, an audience editor noticed her team was spending most of its time writing formulaic social media posts. The content followed a clear structure — same format, different data — so they built a WordPress plugin that generates drafts automatically. The editor reviews and publishes; the drafting work is no longer a bottleneck.

KQED used AI to identify notable moments in radio broadcasts. This addresses a problem of volume: radio archives are large, and interesting clips that warrant repurposing go unnoticed when staff capacity is the limiting factor. The system flags moments for human review rather than making editorial decisions itself.

The Texas Tribune built a chatbot trained on its school voucher coverage to answer reader questions directly. This is audience-facing AI rather than newsroom workflow AI, but the session treated it as related — both involve the same question of where AI output requires human oversight before it reaches anyone outside the organization.

At SWR, a German public broadcaster, community managers now review AI-flagged comments rather than scanning every comment submitted. The system handles the sorting; humans make the decisions on content that needs a response or action.

What the speakers agreed on

Several practitioners distinguished between using AI for research, brainstorming, and the first pass at structured content versus using it for original editorial writing. The consensus was not that AI should be kept out of editorial work but that the point of human oversight needs to be named explicitly in any workflow — who reviews what, and when, before something publishes.

A recurring practical observation: individual journalists can now prototype functional AI tools without a dedicated engineering team. The barrier to building a simple automated workflow has dropped enough that someone with basic coding curiosity and access to an LLM API can test an idea. Testing something small before proposing it organization-wide was presented as a more effective path than waiting for institution-wide AI strategy to materialize.

Useful for journalists and editors who want to start implementing AI in specific parts of their workflow rather than adopting it wholesale, and for newsroom managers thinking about how to structure oversight when AI enters the production process.