AI Policy
Last updated: 29 April 2026.
Barna.News is an AI-native local news publication. Articles are drafted by AI-assisted personas under named bylines and edited by a human Editor-in-Chief before publication. This page sets out how that works and what we will and will not do.
How AI is used
- Source ingestion. AI processes pull from named, verifiable primary sources (official announcements, press releases, named publications, scene-tracker forums, primary documents) and de-duplicate against everything we have already covered.
- Drafting. AI-assisted writers (named in our Masthead) produce article drafts under a consistent voice profile per byline. We do not generate articles from a generic, unbylined model: every published piece has a named author with a documented voice and beat.
- Editing. A human Editor-in-Chief reviews and edits every article before it publishes. The Editor-in-Chief decides headlines, holds the final call on what runs, and signs off corrections.
- Image rights. Editorial images are credited to the original creator. AI-generated images, where used, are explicitly labelled as such.
What we will not do
- No fabricated first-person anecdotes. Our writers report research and named sources. They do not invent the experience of having been somewhere, met someone, or seen something first-hand.
- No fake humans. Our AI-assisted bylines are openly disclosed as AI collaborators here and on the Masthead. We do not impersonate real people; we do not claim our writers are human; we do not file freelance pitches under their names.
- No unsourced claims. Every factual claim has a citable origin. Where something is rumoured or unconfirmed, we say so.
- No neural-tone laundering. We do not use AI to rewrite copy from another publication and pass it off as original reporting. When a story originates with another outlet, we credit them as "first reported by".
Quality governance
Articles are scored against the Pub.cat QVR framework (Quality, Value, Relevance): a three-dimensional rubric covering writing quality, reader-value signal (dwell time, return-in-session, revenue per session) and topical relevance. Articles that score below threshold are flagged for editor review and either fixed or pulled. The framework is owned by Pub.cat and re-tuned weekly.
Models we use
Drafting runs primarily on Gemini 2.5 Flash (Google) for its long-context news handling and tone consistency. Image generation, where used, runs on Flux-schnell (Replicate) and is always labelled. Editorial gates run on a mix of small and frontier models depending on the gate. The model mix evolves; we will update this page when it changes materially.
Errors and corrections
If you spot an error (factual, attributional, accessibility-related, or a privacy issue) see our Corrections Policy. AI-assisted journalism makes mistakes; we want to hear about them and fix them transparently.
Contact
For questions about how AI is used at Barna.News, email [email protected].