Of the 440 applications Texas Tribune chief product officer Darla Cameron received for an AI engineering role, 90% were junk — many apparently written using the very tools candidates were supposed to understand. One application literally read: “Here’s a short response that’ll work for this.”
Key Takeaways
- Poynter’s SXSW takeaways: start with a real pain point, draw the line, name the human.
- Texas Tribune saw 90% of 440 AI engineer applicants submit junk written by AI.
- Practical newsroom AI is winning over cargo-cult strategy and hype.
That anecdote, from Hacks/Hackers and Poynter’s AIxJournalism Day at SXSW, captures the split-screen reality of AI in newsrooms right now: genuinely useful when deployed against real problems, already producing absurd side effects when cargo-culted. Poynter’s Alex Mahadevan spent the day listening to newsroom builders — not theorists — and distilled five lessons worth reading carefully.
Start with a specific pain point. Every AI tool actually being used in newsrooms started with a complaint, not a mandate. Pew Research Center’s audience team was spending 95% of its time writing formulaic social posts and 5% engaging with audiences. A WordPress plugin now drafts the formulaic posts automatically; the team reviews options and moves on. That freed time goes to reading comments and responding to what readers actually ask. “Where is the mundane bullshit work that you’re sick of doing?” said Upasna Gautam of the News Product Alliance. “That is a great pathway.”
Draw the line between AI for thinking and AI for writing. Nebraska Public Media’s chief innovation officer Chad Davis stopped using AI for writing entirely — not good enough. But he uses it constantly as a “curiosity partner” for research and brainstorming. His labs team uses AI for concept art and music prototyping: instead of pitching an idea in a meeting, they show a working prototype. Most newsrooms haven’t drawn that line clearly. If you don’t decide, your staff will make their own inconsistent rules.
Be specific about where the human sits. KQED tested AI to identify notable clips from their hourlong radio show Forum. “I’m not ready to say the AI can choose the four most notable moments,” said editor-in-chief Ethan Toven-Lindsey. “But if you put a producer in the loop to make sure those are the right moments, that felt doable.” SWR, a German public broadcaster, has community managers review AI-flagged comments before anything gets acted on. “Human in the loop” is too vague — the newsrooms getting results name exactly where.
Use AI to get closer to audiences, not further away. The Texas Tribune trained a chatbot on its school voucher coverage. When readers asked questions the Tribune’s reporting hadn’t addressed, reporters got story ideas. Pew studied what creators do differently from traditional newsrooms and found a feedback loop: they read comments, answer questions, and build new content from what audiences ask. Any AI strategy that doesn’t free up staff time for that kind of direct audience engagement is probably solving the wrong problem.
Learn to build things yourself. This is the one that cuts across every other lesson. Mahadevan — not a software engineer — built a fact-checking research tool prototype for PolitiFact on vacation using AI coding assistants. Agentic coding tools like Claude Code have lowered the floor dramatically. Trei Brundrett of New_Public said it reminds him of the early web, when gatekeeping vanished and anyone could just build. “The newsroom people who pick up that skill are going to have an outsized advantage,” Mahadevan writes. The ones waiting for someone else to build it for them will be waiting a long time.
