Method
This page describes what the system actually does, including the parts that work poorly. If something on the site looks wrong, the explanation is probably here.
Where articles come from
We poll the feeds of around 127 publications directly — trade press, general news, research labs, policy bodies, vendor blogs and community sources — and supplement them with a set of news searches covering the major labs, products and recurring beats. Vendor and research feeds get a longer collection window than newsrooms, because a lab blog publishes when it ships something rather than hourly.
How articles become stories
Each article is read once and reduced to a fingerprint: who the event involves, what kind of event it is, when it happened, and one plain sentence describing it with the outlet’s framing stripped out. Articles whose fingerprints agree are grouped into a story.
A candidate has to resemble the story as a whole, not merely one article inside it. That is deliberate. Matching on a single article lets one ambiguous piece weld two unrelated events together — a failure we hit and corrected.
Camps
AI coverage splits into recognisable camps. We label five: optimist, pragmatist, skeptic, safety and ethics.
The camp is judged from what the article says, not from who published it. A skeptical piece in an enthusiastic outlet counts as skeptical. This matters: measured against outlet-level labels, roughly 45% of articles read differently from their publisher’s general reputation.
Opinion columns are counted separately and kept out of the camp bar. A column is not coverage of an event.
Evidence grades
Each story carries a grade for what backs its central claim: a paper, a benchmark, a shipped product, a demo, a press release, a rumour, or speculation. This grades the sourcing of the claim. It is never a judgement about the outlet reporting it.
Syndicated copies
One wire story can run in fifty outlets under the same headline. Counting those as fifty reports would inflate every number on the site, so copies are grouped under the article they duplicate and left out of the counts. They are still listed and still linked — nothing is hidden, it simply isn’t counted twice.
Known limits
- Paywalls. Roughly a quarter of articles won’t yield their text. Those are classified from headline and summary alone, which is less reliable.
- Grouping errors. Two stories about one event sometimes stay apart; occasionally unrelated events merge. We bias toward over-splitting, because a split is visible and a false merge is not.
- Absence is not a verdict. When we note a camp produced nothing on a story, that is a fact about what we collected. It is not a claim about anyone’s motives, and our source list is not a census of AI writing.
- Summaries are generated. The neutral summaries and the agree/disagree analysis are produced automatically from the articles listed on the page. The original reporting is always one click away, and that link is the authority — not our summary of it.
Found something wrong? hello@aicoverage.org