A screenshot is an anecdote. A poll is a record.
When a chatbot says something newsworthy, the first question an editor asks is the right one: would it say that again? One reply cannot answer it. Mind Polling can: the same question, to every model you choose, as many times as you like, on a schedule, with every reply kept.
The newsrooms that did this did it by hand
- The EBU and BBC, with 22 public broadcasters, collected over 3,000 answers and found 45% had at least one significant issue (report).
- NPR, with NewsGuard, tested six chatbots on Russian, Chinese and Iranian false narratives in July 2026, and found they did better than expected (NPR, via KPBS). A good result is a story too, and it needs the same evidence.
- Proof News published every answer from its election audit as a spreadsheet (data), so anyone could check the work.
Each built its own pipeline, ran it once or twice, and stopped. A shared tool means the follow-up story, three months later, costs an afternoon instead of an engineer.
Stories this is built for
- Does the answer hold? A chatbot told a reader something wrong. Ask it 20 times, across models, and report how often it says it.
- Did it change? Ask the same question every day through a news event, and show the day the answer moved, with the model that answered.
- Does it depend on who asks? Write the question as different askers, and report how the answer shifts, never as what those people think.
- Will it say? Ask about something not yet known, and report how often each model declines. Declining is an outcome Mind Polling counts, not a failure it hides.
Writing it up
Say what you measured, in the terms you measured it: "Asked 20 times through its API on Nov. 3, Model X named B in 14 (70%, range 48% to 85%)", not "Model X prefers B". Name the version, the dates and what was left out. Say it was the API, not the app. The methodology has the checklist, and every answer downloads for your fact-checker.
Start small
One question, three models, two wordings, one run. Read the replies before you schedule anything. If the question does not catch the distinction you need, change it: the new version starts a clean series.