How NewsAI works

NewsAI starts from configured primary channels — official blogs and research feeds (RSS), GitHub releases, and Hacker News — rather than open web scrape-and-invent. Items are stored with their source URL and publisher family before any summary is written.

Near-duplicate coverage of the same event is clustered so the feed shows one card per event when possible, with evidence drawn from the contributing sources.

Every summary is checked before it becomes a feed card. The primary link and every attributed quote URL must exactly match a URL supplied by the source items for that story. Unvalidated links are rejected — we never invent a source URL for display.

On story pages, the sources list and quote links are the same validated set. Source labels are derived from known hosts or hostnames, not editorial trust claims beyond that.

By default, the public feed only includes eligible stories whose event time (or summary time when event time is unknown) falls within the last 14 days. This keeps the main surface timely without silently rewriting history. Older material is not the default browsing experience.

Ranking combines quality signals (evidence, credibility of source families, engagement from likes), freshness, and novelty, with penalties for weak solo coverage. Eligible stories are ordered into a short-lived snapshot so pagination stays stable while you scroll.

Diversity rules limit back-to-back stacking of the same category or source family. A freshness lane keeps recent coverage visible. Modes such as For you, Trending, and Latest change ordering — they do not invent stories outside the eligible pool.

Signed-in readers can choose preferred categories and like or save stories. Those signals may add a bounded boost to ranking order on For you. They never become an implicit topic filter: the All feed still mixes topics, and non-preferred categories remain visible.

Preferences are opt-in. Skipping onboarding does not hide news; it only leaves ranking unboosted by category choice.

NewsAI does not add click, scroll-depth, read-time, or third-party behavioural tracking for personalisation. The only personalisation signals are category preferences, likes, and saves that you explicitly create. Anonymous likes use a local device id; they are not sold or used to profile readers across the web.

On each story page you can report an incorrect summary, wrong source, duplicate, wrong category, or other issue. Reports are stored for operators with rate limits and plain-text details only. Submitting a report never automatically removes, rewrites, or re-summarises a story.

When an operator applies a correction, the story may show a public correction note and timestamp. Provenance rules still apply: source URLs are not silently swapped for unvalidated links.