How GreyNews works
Every score, badge, and verdict on this platform comes from a documented, repeatable process. Here is exactly what we do — and what we do not do.
Claim Verification
Every factual statement in an article is extracted and cross-referenced against authoritative data sources. A claim is only marked Accurate if corroborating evidence exists from at least two independent sources.
Our AI reads the article and identifies every discrete factual claim — statements that can be tested against data.
Each claim is matched against government statistics (BLS, Fed, World Bank, IMF), peer-reviewed research, newswire corrections, and official statements.
Accurate = supported by ≥2 independent sources. Inaccurate = directly contradicted by authoritative data. Unverifiable = insufficient public data to confirm or deny.
A Bayesian confidence interval (e.g. 92–97%) reflects how consistent the supporting evidence is. Wider ranges mean thinner evidence.
Source Scoring
The Source score reflects the outlet's historical accuracy — not the quality of this specific article. It is calculated from the outlet's track record across all articles we have processed.
How often the outlet has issued corrections or retractions relative to total article volume
Percentage of verifiable claims in past articles that checked out against authoritative sources
How many distinct source categories (government, academic, industry, NGO) are cited per article on average
Manipulation Detection
The Manipulation score (0–100, lower is better) measures how much an article relies on emotional triggers rather than facts. It is a composite of three signals.
Counts emotionally charged words per 100 words, benchmarked against a 10,000-term connotation lexicon. Includes fear, outrage, and urgency triggers.
Measures the sentiment gap between the headline and the article body. A 90% negative headline paired with a 50% negative body scores high.
Detects passive voice used to obscure agency, selective quoting of only one side, and adjective-to-noun ratios above editorial norms.
Bias Detection
Bias detection does not label outlets as “left” or “right.” It identifies specific structural patterns in how a story is told. Four categories are tracked.
How an event is characterized — word choices that assign blame, credit, or urgency without stating facts. Example: "regime" vs "government", "crisis" vs "situation".
Factual information present in other sources covering the same event that is absent from this article. Detected by cross-referencing coverage.
Exaggeration of stakes, certainty, or urgency beyond what the evidence supports. Overlaps with manipulation detection.
Over-reliance on a single category of sources (e.g. all US government, all financial media). Diversity of perspective is tracked per article.
What we don't do
We do not investigate or publish original fact-checks. We cross-reference claims against existing authoritative data and publicly available sources.
Source scores are statistical facts about historical accuracy, not editorial judgments. A 78% score means 78% of verified claims checked out — nothing more.
Claim extraction and stance detection are probabilistic. Every verdict should be treated as an analytical signal, not a definitive ruling. Source links are always provided so you can check yourself.
The same methodology runs on every source with identical thresholds. No outlet is excluded. Any outlet can contact us to dispute a specific verdict with evidence.
This system is in active development. Scores are indicative signals, not definitive verdicts. The methodology is peer-reviewed annually and updated as our models improve.