Methodology

MediaBias News is an aggregating newsroom. We read what other outlets report about International news in English, wherever it is reported from, compared across the outlets that carried it — with particular attention to stories one part of the political spectrum is covering and another is not., assess the quality of that reporting, and write our own account from the facts it establishes — always linking back to the originals.

The three scores are the reason to read us rather than any single outlet. They are published on every story, with a one-line explanation of why each landed where it did.

The three scores

Trust and Craft are not opinions of a story rounded to a number. Each starts at a floor and earns points only for things that can be pointed at in the text — a source named on the record, a figure whose origin is stated, a right of reply offered. Where the coverage falls short in specific ways, it loses points the same way. The lists below are the actual checks.

Trust — 0 to 100, higher is better

How well corroborated and evidenced the original reporting is. Starts at 30 and earns:

• a source named on the record, with their role
• a document, official statement or dataset quoted directly
• two or more genuinely independent outlets — syndicated copies of one article do not count
• the central claim attributed to a named body rather than to “sources”
• figures given with their origin stated, and specific dates and places

It loses points where the central claim rests on anonymous sourcing alone, or where key facts are hedged throughout. A wire brief carrying one named official typically lands in the forties or fifties; the seventies require the kind of corroboration a story only gets once several newsrooms have worked it independently. A low Trust Score is not an accusation of dishonesty — it means there is less here that you or we can check.

Craft — 0 to 100, higher is better

The quality of the journalism itself, separately from how well sourced it is. Starts at 45 and earns:

• telling the reader why this matters, or what led to it
• quoting anyone affected, opposed or responsible, or offering a right of reply
• keeping fact clearly separate from comment
• specific figures, dates, streets and names rather than vague quantities
• saying what happens next, or naming what is still unknown

It loses points for a press release reproduced with no reporting added, for speculation presented as established fact, and for a headline that promises more than the body delivers.

Why the numbers changed in July 2026

An earlier revision described each band in words and asked the model to pick one. We checked what that produced across 393 stories and it barely discriminated at all: the Trust Score used six distinct values in total, and 262 of those stories sat in the same 70–79 band. Two very different pieces of reporting both scoring 75 tells you nothing, which hollows out the only thing this site offers.

So the scores are now built by adding up the checks above, and every story published before the change has been re-scored on the new basis. Numbers across the site are lower than they were and mean something different: they are a count of evidence present, not an impression. Eight older articles whose source text was never fully captured could not be re-scored, and rather than show you a number from a scale we have retired, we removed their scores entirely.

Hype — 0 to 100, lower is better

How far the presentation runs ahead of the substance. This is the one score where a high number is bad.

• 0–20 — measured, proportionate language.
• 41–60 — notable sensationalism, loaded adjectives, a thin story dressed up.
• 81–100 — fabricated urgency, or a headline contradicted by the article.

Where the coverage comes from

The machinery that turns a live event into a page here is this newsroom's own. Article text is pulled with a fallback ladder — a structured extractor first, Mozilla's Readability behind it, our own parse behind that — because no single extractor survives contact with forty different publishing platforms, and a feed advertising full text routinely carries a teaser instead. Copy that several mastheads have republished word for word is detected and counted once. The clustering, the three passes that score, write and headline a story, and the gate that discards a draft sitting too close to its sources are all set out below.

That extraction ladder is not an original idea and we are not going to present it as one. MuckScraper, an open-source aggregator doing comparable work, arrives at the same multi-step fallback for the same reasons, and credits AllSides for its ratings much as this site does.

What we do not build ourselves is the list of which outlets covered which event. Tracking that continuously, across the whole press, is not something a newsroom this size can do from a standing start, so we pay for access to a commercial news-aggregation service and use it for exactly that: which events exist, and which outlets covered them. None of its published text appears here and none of its analysis or ratings is presented as ours. Everything you read on a story page is written by this newsroom from the outlets' own reporting, and every outlet named is linked to its own article. What that service gives us is a starting point, not content.

Not every outlet in a comparison is read in full. Reading forty articles to write one story is not something this site can afford to do forty times a day, so the usual pattern is that two of the outlets are read in full and the rest contribute a short account of what they published. Every story page says which arrangement produced it, in those words, at the foot of the article.

Who decides an outlet leans left or right

Not us. Every leaning rating on this site is the published work of one of three organisations — AllSides, Ad Fontes Media and Media Bias/Fact Check — and each is named beside the rating it made. They rate factual reliability on a separate scale, which is shown separately, because a leaning and an accuracy record are different questions and running them together is the mistake this site exists to argue against.

They frequently disagree with each other, and we publish every rating we hold for an outlet rather than averaging them into one number that would conceal it. The three scores this newsroom does produce — Trust, Craft and Hype — assess a single piece of reporting and are described below. They are never a rating of a newsroom.

What it costs, and what it is not

Everything described on this page is free to read. There is no account to create, no metered article limit and no subscription tier: the comparison is the product rather than an upgrade to it, so nothing about how a story was covered sits behind a payment.

It is not a rating organisation. AllSides, Ad Fontes Media and Media Bias/Fact Check do that work, and this site publishes what they concluded rather than a verdict of its own — all three side by side, each named, including where they contradict one another. They disagree often enough that averaging them into one number would conceal more than it reported, which is why all three are shown separately.

It is also not independent of the services above, and says so on this page rather than leaving it to be discovered: story discovery is bought from a commercial aggregator, and the leaning ratings belong to the three organisations named. What is this newsroom's own work is the extraction, the deduplication, the clustering, the writing, the three scores, the outlet records computed from those ratings, and the blindspot findings.

How a story is produced

Three separate passes run in a fixed order, and the order matters.

Syndicated copy counts once

Outlets frequently republish each other's articles word for word. Three mastheads carrying the same piece is one piece of reporting, not three confirmations, and treating it otherwise would push Trust Scores up for no good reason — an audit found half our multi-source stories contained syndicated copy. Identical source text is now detected before scoring and the assessment is told how many genuinely independent reports it is looking at.

1. Scoring runs first

The scoring pass sees only the original source articles — never our own write-up, because at that point it does not exist. This is deliberate. If we scored our own rewrite, the score would measure our writing rather than the reporting it came from, and would drift upwards over time.

2. Rewriting

A second pass writes an original article from the factual claims the scoring pass isolated. It is instructed to write from the facts rather than the sentences, to attribute every claim to who said it and which outlet reported it, and to invent nothing.

Every draft is then checked mechanically against its sources. We compare ten-word runs, ignoring anything inside quotation marks, since reproducing what someone actually said with attribution is journalism rather than copying. If more than 12 per cent of our unquoted prose matches a source, the draft is discarded and no article is published. For calibration: a hand-checked, genuinely original rewrite measures around 4 per cent, almost entirely on unavoidable phrasing such as statutory wording and restated statistics.

3. Packaging

A third pass writes the headline and social copy. It is constrained by the Hype Score we just published: our own headline may not read as more hyped than the coverage we are assessing. Publishing a hype measure and then writing clickbait would make the measure worthless.

We check that we keep to it. An audit of every published headline found none carrying a figure the article did not support, and one that escalated its source's register — the BBC reported that a translator had written a scathing review and we published “slams”. That headline was changed and logged in the corrections list, and the packaging instructions now name the tabloid verbs that are not to be used and rank alternative headlines by how well they inform a reader who will never click, rather than by how tempting they are.

What runs the passes

Current prompt revision: 3.2.0-mediabias.2.0.0+light.r2. Each story records the revision that judged it, so a score can always be traced to the exact instructions that produced it.

• Scoring — google/gemini-2.5-flash-lite, temperature 0
• Rewriting — google/gemini-2.5-flash-lite
• Packaging — google/gemini-2.5-flash-lite

The The comparison desk profile

The network-wide accuracy, attribution and anti-fabrication rules are identical across every title. This publication then adds its own newsroom profile, revision 2.0.0, so its reporting can have a distinct audience, rhythm and coverage lens without weakening those safeguards.

Give every story one page that shows how left, centre and right covered it, what each side left out, and every outlet that ran it.

• State what happened before saying anything about how it was covered.
• Attribute every claim to the outlet that made it, by name.
• Where outlets disagree on a fact, say that they disagree rather than picking a winner.
• Treat an omission as a finding: which side did not report a detail is itself reportable.

Scoring runs at temperature zero, which is the setting that asks a model to be as consistent as it can be. It does not make it deterministic, and we would rather give you the measurement than the assurance. We took 15 stories and scored each one twice from identical input: the Trust Score came back unchanged 11 times out of 15 and the Craft Score 10 times out of 15, with an average difference of two to three points and a worst case of 16. Identical inputs can be routed to different machines, and the answer moves a little when they are.

So a score is repeatable to within about five points, not to the point. That is the honest resolution of the number, and it is why the bands below matter more than the digits.

Models are chosen for cost as well as quality: this site is free to read and carries no paywall, and running it sustainably matters.

Where this can go wrong

We would rather state the limitations than have readers discover them.

• The scores are judgements, not measurements. Two careful editors would disagree at the margins, and so would two runs of a model. Treat a score of 72 and a score of 78 as the same verdict.
• We can only assess what we can read. If an outlet has strong off-the-record sourcing it cannot describe, the Trust Score will understate the story.
• Corroboration is not truth. Several outlets running the same press release is not the same as several outlets checking it independently, and the scoring pass cannot always tell the difference.
• Language models make mistakes. They can misread a date or conflate two people. This is why every story links its sources, and why corrections are logged publicly.

The counts, worked through

Three standing pieces take the measurements this page describes and follow them somewhere:

• The media bias charts disagree with each other on most outlets — how far apart AllSides, Ad Fontes Media and Media Bias/Fact Check are on the same publishers, counted.
• The outlets every rater calls centre — and why the AP is not one — the closest thing to a checkable answer to “which source is unbiased”.
• Half the stories one side ignores, it runs a day later — what a blindspot is worth once coverage has had time to settle.

Getting it wrong

When we are wrong, we correct it and record it in the public corrections log, including any case where a published score was changed and why. If you think a story or a score is wrong, tell us at editor@mediabias.news.

Common questions