How source scores work

For the articles ChamberLight collects, a large language model reads each article and rates the organization that published it from 0 to 100: its transparency, editorial standards, corrections practice and record of original reporting. The model is told which outlet published the article. A source’s credibility score combines one rating per story across the source’s own stories.

Which articles count

• The source’s own articles, including its opinion pieces: an outlet answers for everything it publishes.
• Articles it reposted from another publisher when only the AI says so. Without firmer evidence the article stays with the site that ran it (see who wrote an article).
• Reposts with firm evidence count toward no one’s credibility. The site that ran them did not write them, and a publisher’s credibility is never built from copies of its work on other sites.

What credibility is not

• Not a fact-check. No one checks the claims in each article against the record to produce this number.
• Not a review of the outlet’s policies. The model does not visit the outlet’s corrections page, ethics code or ownership records. It draws on what it already knows about the outlet, so the score largely reflects the outlet’s reputation as the model understands it, more than the article in hand.
• Not an endorsement or a verdict on one article. A high score does not make an article true, and a low score does not make it false.
• Not based on votes. Votes on articles do not feed any source score.

For every article, the model also rates where the article sits politically: 0 is strongly left, 100 is strongly right, and 50 means it found no left or right tilt. A source’s lean combines one rating per story across its news stories.

Lean is not balance

A lean near 50 means no detectable left–right tilt. It does not mean the coverage is fair or complete: an article can be one-sided about something that is not a left–right question and still score 50. Balance is a separate rating of each article from 0 (entirely one-sided) to 100 (presents the sides fairly). The “Most balanced” list on the sources page ranks outlets by balance, using the same method and the same 10-story minimum; it is recalculated nightly.

Which articles count

• The source’s own news articles, and reposts on its site that only the AI attributes to someone else.
• Its stories reposted on other sites, when the evidence is firm. One story reposted by several sites counts once.
• Not opinion pieces. ChamberLight marks an article as opinion from signs such as an /opinion/ or /op-ed/ web address or the page’s own opinion label. An outlet’s lean describes its reporting; its credibility still covers its opinion pages.
• Not articles known to be in a language other than English.

Who has no lean

Lean is a US left–right scale. Outlets based outside the United States, outlets that publish mainly in another language and state-controlled media are not placed on it. Individual writers and online communities are not rated on credibility or lean, because both are ratings of news organizations. The full list of reasons is under sample size.

Originality is the share of a source’s hosted stories (the stories ChamberLight collected from the source’s own site), with curated links left out, that are not reproductions confirmed by verified or corroborated evidence (see who wrote an article). It is published once a source hosts at least 10 stories, and is shown as a percentage with the number of stories behind it. It has no range.

• Original stories, stories whose origin has not been checked yet, stories that only the AI flags as copies and reposts whose original publisher is not identified all count as original, because an unconfirmed copying flag is not evidence either way. Beside the percentage, a source’s page shows how many of its stories carry such a flag.
• Curated links, where a site points readers to another publisher’s article, are left out: a link is neither original nor copied content.
• Originality does not change credibility or lean. Every source is scored on its own articles, however much else it reposts.

Source pages also show a writing-quality figure. It is separate from credibility and is not one of the three scores. For each story, the model rates the journalism in the article itself, such as its sourcing, evidence and balance, from 0 to 100. A source’s figure is the plain average of those ratings over the stories that count toward its credibility. It is shown once at least 10 of those stories have a rating, and only for sources that can have a credibility score.

• It has no range, and unlike credibility and lean it is not pulled toward a starting point.
• It does not feed credibility or any other score.
• Analyses made with older versions of the instructions have no writing-quality rating, so the figure covers only stories analysed with newer ones. The rating itself is due to be revised (see known limits).

Many articles are copies: a local TV site running a wire story, a portal reposting a newspaper. ChamberLight works out who wrote each article because that decides which source’s scores it counts toward. It uses:

• the page’s canonical link, when it points to another publisher’s article;
• credits on the page: the byline, the listed author or publisher, a copyright line, structured data naming the publisher, or a note such as “This article first appeared on …”;
• a curated list of known syndicators, aggregators and archives;
• the AI analysis, which reads the article and says whether it looks copied and from whom.

Evidence levels

• Verified: the page itself shows it (a canonical link or page credits), or a ChamberLight editor confirmed it.
• Corroborated: the AI named the original publisher and an independent signal agrees: the site that ran it is a known syndicator, aggregator or archive in the curated list or by editor review, or the article’s author or byline names that publisher.
• AI only: the AI says so and nothing independent confirms it. The article stays with the site that ran it, which is shown as citing the other publisher.

Labels on a source

Each source page says what kind of site it is:

• Original publisher: Publishes its own reporting.
• Syndicator: Republishes articles from other outlets under licence, credited to them.
• Aggregator: Links to other publishers' articles rather than writing its own.
• Republishes without licence: Reproduces other publishers’ articles without a licence.
• Mixed: Publishes a mix of its own reporting and reproduced articles.
• Archive or proxy: Stores or relays copies of other publishers' pages; not a publisher itself.
• Provenance not yet reviewed: ChamberLight has not yet assessed whether this source publishes its own reporting or republishes others’.

A label comes from the curated list, from an editor’s review, or from the source’s own articles. A label from the articles needs at least 10 hosted stories. “Original publisher” is given only when ChamberLight can rule out, with 95% confidence, that 20% or more of the source’s stories are unconfirmed copies. When 10% or more of its stories carry an unconfirmed copying claim, the source goes to an editor instead. The automated check never labels a site as republishing without a licence; only the curated list or an editor can.

A label from the automated check never changes a score. A label from the curated list or an editor can: when it marks a site as a syndicator, aggregator or archive, the site’s reposts whose original publisher the AI named count as corroborated (see evidence levels above). They then leave the site’s credibility and lean, count as copied in its originality, and add to the original publisher’s lean.

What counts as a story

A story is one article. When ChamberLight shows the same article on the pages of several members of Congress, it still counts once, and a story reposted by several sites counts once for its publisher. Each story contributes its newest analysis.

Story–official pairs

The “Who they cover” panel on a source page counts story–official pairs: a story that names three members of Congress counts once for each of them, so a source can have more pairs than stories. Each official’s share is the share of the source’s stories that name them.

The 10-story minimum

With a handful of stories, one unusual article can swing an average. A score is therefore published only when a source has at least 10 stories. Below that the page says “Not enough stories yet” with the count so far, for example “4 of 10”.

Analyses that are left out

• Analyses the model itself marked as low confidence (below 0.5 on a 0–1 scale).
• Articles shorter than 150 words, when ChamberLight stored the article’s text.
• Articles whose origin has not been checked yet. They count once the check runs.
• No version filter is applied today: analyses made with every version of the instructions count.

Pulling small samples toward the typical value

A score is not a plain average. Each source’s average is pulled slightly toward a starting point, the typical score of comparable sources, and the pull weakens as the stories add up and agree with each other. For credibility the starting point is the average of sources of the same type (for example local news or government) when at least 5 sources of that type have 3 or more stories, and the average of all sources otherwise. Sources with no type yet, or typed only as “other”, always use the average of all sources. For lean it is the average of all sources. At 10 stories a source’s own average carries about 99% of the weight for credibility and 98% for lean, so the adjustment is small.

The range

Next to a credibility or lean score is a 95% range. It shows how precisely the stories pin the score down: it narrows as a source gets more stories and as its stories agree more. When a lean’s range crosses the line between two labels, the source could fit either, so ChamberLight names both, for example “Center-Left to Center”.

What you may see instead of a number

• Not enough stories yet: fewer than 10 stories, with the count so far.
• Not rated yet: ChamberLight has no scored stories from the source.
• Set by ChamberLight editors: an editor corrected the score with a recorded reason (usually after a challenge). It shows no story count because it does not come from the article analyses.
• Not rated: the source is deliberately left out of a score, for one of these reasons:

How often scores change

Scores are recalculated every 15 minutes. The date shown with a score is the day it last changed; a score that stays the same keeps its date.

Government offices, members’ offices and press-release services (46 sources today) are scored like any other source and keep their scores on their own pages. They are not ranked against news outlets: they appear on a separate “Official sources” list ordered by how many stories they have, never by credibility, and they are not placed on the left–right spectrum of news outlets.

An article page shows the analysis of that one article: the credibility rating of its publisher, its lean, its balance and, for a repost, the quality of the journalism in the article itself. Most come with the model’s reasoning; where it cannot be shown, a plain description of the measure appears instead. A source’s score is the combination of these ratings across its stories, as described above. When an article is a repost with firm evidence, the page credits the original publisher and shows that publisher’s credibility on the provenance card; the reposting site’s own score still appears, labelled “Host credibility”.

• Credibility leans on reputation. Because the model knows which outlet published each article, the score says more about how the outlet is regarded than about the article in hand.
• Scores combine analyses made by more than one AI model and more than one version of the instructions. The model and the version are recorded with every analysis.
• The scores describe the articles ChamberLight collected about members of Congress, not everything an outlet publishes. Articles come from links readers submit, which moderators review, and from ChamberLight’s collection of news feeds, filtered for coverage of members of Congress.
• ChamberLight’s collection has gaps, such as outages. Publishing charts shade months in which ChamberLight collected no stories, or almost none, from any outlet: a quiet month there says nothing about the outlet.
• Some reposts are not detected. A site that runs wire copy without a visible credit can be scored on the wire’s articles.
• Opinion pieces without an opinion marker in their address or page are treated as news.
• Many sources have no type yet (“Not yet classified”). This affects only which starting point their credibility is pulled toward.
• The writing-quality rating is due to be revised. On a source’s own articles it tends to run higher than credibility, and analyses made with older instructions have no rating at all.

Anyone can challenge a score or a label, with or without an account. Use “Challenge this score” on the source’s page and choose what you are challenging:

• Credibility score
• Lean
• Source type
• Original publisher / reposting
• Name
• Something else

Describe the problem in 20 to 2,000 characters. You can add a link to evidence (https) and an email address; editors may use the address to ask about your challenge, and it is never published.

What happens next

• A ChamberLight editor reviews every challenge. The score stays visible while it is open.
• The source’s page lists its challenges and their status: open, upheld or rejected. For upheld and rejected challenges it also publishes the editor’s note. The text of your challenge and your email address are not published.
• When a challenge is upheld, the editor records the correction with a written reason. A corrected score reads “Set by ChamberLight editors”; a corrected type, name or label changes at once.

How many you can send

To keep the queue usable, signed-in readers can send 5 challenges an hour; without an account the limit is 3 an hour and 10 a day. If you send a second challenge on the same source and topic within 24 hours of your first, while your first is still open, it is turned away.