Why NewsHelper exists
Most news aggregators optimize for the fastest possible headline. NewsHelper does the opposite: each day it picks a small number of stories getting genuine attention across several outlets and tries to explain what's actually going on — plus point toward a book or article if a headline isn't enough.
How the stories get picked
Headlines are pulled from free, public RSS feeds — Google News, Google Trends, BBC, and NPR. A story only makes the day's list if multiple outlets are covering something close enough to the same thing; that clustering and ranking is plain, deterministic code, not an AI's judgment call, since deciding "are these four articles the same story" is a much easier problem to get right without a model than to verify after the fact.
How the summaries get written
Each story's plain-language summary is generated by a small AI model running on a home computer, not a cloud AI service — a deliberate choice to keep this a zero-cost hobby project rather than one with an ongoing bill. That also means summaries come from a smaller, less capable model than a commercial assistant, which is part of why the disclaimer below matters.
On book and article recommendations
The model only ever suggests a topic — never a title directly. Every book recommendation is checked against a real library database (Open Library, falling back to Google Books) before it's allowed to appear; if nothing matches, the suggestion is dropped rather than published unverified. Book links point to a plain Bookshop.org search, not an affiliate program — this project makes no money from anything you buy.
On satire and misinformation tags
NewsHelper uses two independent, narrow signals — never an AI verdict on what's true. A manually maintained list of known satire/parody outlets gets a SATIRE tag; a story is never hidden for this, only labeled, since a satirical piece trending enough to look like real news is more useful flagged than hidden. Separately, headlines are checked against Google's public database of already-published fact-checks; a match only appears when the fact-check text is similar enough to the actual headline, and the page always links to the original fact-check rather than asserting anything on its own.
Being honest about the limits
This is a hobby project built and run by one person, on a home computer, using free tools and a small AI model. It will sometimes get things wrong — a summary can miss nuance, a fact-check match can be a looser fit than it first appears, a source's own reporting can simply be incomplete. See the disclaimer on the digest page for the full version; the short version is: treat this as a starting point for understanding a story, not the final word on it.
Under the hood
NewsHelper is open about how it works. The full pipeline — RSS fetching, story ranking, AI enrichment, and rendering — along with the reasoning behind each design decision, is documented and public: github.com/billford/newshelper.