How Prism works

Three times a day an automated job reads the public RSS feeds of about 175 news outlets chosen to span the spectrum — The Guardian, Jacobin, and Common Dreams on the left; the BBC, Al Jazeera, and DW in the centre; the Washington Examiner, National Review, and Breitbart on the right — plus a few TV channels’ YouTube feeds. Roughly a hundred of them are outside the United States: Canada, the UK and Europe, the Middle East, Asia, Africa, Latin America and Oceania, from the Daily Maverick to the Japan Times to the Toronto Sun. Every feed is verified live before it joins the list, and every run draws from outlets in turn, so a high-volume feed cannot crowd the others out.

Only what a feed openly provides is used: the headline, a short description, a link, a timestamp, and where present a lead image or video id. Full article text is never fetched, copied, or stored. Prism doesn’t crawl article pages, use anyone’s API, or get past a paywall, and its reader identifies itself as PrismBot so any publisher can see in its own logs exactly what it takes.

Articles about the same event are grouped, and an event becomes a story only when at least two different outlets covered it. Only political and civic news is kept — elections, legislation, courts, government actions, foreign policy, the economy, civil rights. Sports, entertainment, and celebrity news are filtered out.

Prism publishes two different ratings, and the difference matters.

The article rating is the one that counts. Every piece in a story is placed on the five-point spectrum, and the words from the piece that put it there are shown beside it in “How each outlet covered it”. A newsroom is not one voice — the same masthead runs plain wire copy one hour and a sharp column the next — and rating the piece keeps the claim checkable: you can read the quoted words and disagree. Every quotation is verified to be a real substring of that article’s own headline or summary; a placement whose quotation can’t be found is discarded, never shown.

The outlet rating is internal. Every outlet Prism reads carries one of six working labels — left, lean-left, centre, lean-right, right, or libertarian — but you won’t see them as labels on the site. They exist to keep the source list spread across the spectrum and to match a video to the slider (videos aren’t rated piece by piece). Nothing on a story is coloured or captioned by an outlet’s general lean: a piece Prism hasn’t placed from its own words simply shows as not rated. That is a deliberate choice — across the articles Prism has placed so far, the piece sits on the same side as its outlet’s general lean only about six times in ten, so an outlet label would misdescribe the specific article roughly two times in five. The working labels are based on the outlet’s stated editorial stance, the direction of its opinion pages, and where the established public raters place it — used as a cross-check, not copied, and none of them has reviewed or endorsed anything here. For a second opinion on any outlet, go to the people who do this at scale, with published methods:

• AllSides — blind surveys, editorial review, and reader feedback
• Ad Fontes Media — analyst panels scoring individual articles for bias and reliability
• Media Bias/Fact Check — editorial review against a published rubric

Both ratings are relative and contestable. Where the centre sits depends on the country and the decade, and some outlet ratings are probably wrong. The source list isn’t perfectly balanced either: of about 175 outlets, 63 carry a left-of-centre working label, 62 centre, and 53 right-of-centre — and for the hundred or so outside the United States, that label is relative to the outlet’s own country’s politics, which is a rougher judgment still. That is why Prism publishes no score for how “balanced” a story’s coverage was — most stories here are carried by two or three outlets, and a tally of a leaning list on a sample that small would measure Prism’s reading list, not the world’s coverage. Prism shows who reported a story and where each piece sits, and stops there. The conclusion is yours.

If a rating looks wrong, say so — use the feedback link at the bottom of any page, name the outlet, and say why. Every message is read, and ratings do change.

Which photograph an outlet runs is one of the strongest framing choices in journalism — the rubble or the podium, the crowd or the arrest. So Prism doesn’t pick one picture and call it the story’s image. Where several outlets supplied a photo, moving the spectrum slider on a story card swaps between them, and “How each outlet covered it” shows them side by side, each credited to the outlet that chose it. If no outlet at the selected position supplied one, the nearest available photo is shown.

Every photograph loads directly from the publisher’s own server. Prism does not copy, re-host, or alter anyone’s photography, and it never generates images with AI.

Each story carries two kinds of writing, kept deliberately apart.

“What happened” is the factual floor — a short account of only what the coverage agrees occurred, with contested figures attributed rather than asserted (“Lebanon’s health ministry puts the toll at 12” rather than “12 people were killed”) and unknowns named as unknowns. It belongs to no position on the spectrum. It is the account you should be able to accept whatever you think.

Then five framings of those same facts, as a thoughtful person on the left, centre-left, centre, centre-right, and right would each tell them. Centre is a real position here — someone who distrusts both poles, prefers incremental fixes to sweeping ones, and thinks each side has hold of part of the truth — not “the neutral one”. The facts sit above all five and belong to none of them.

The model is told to write in its own words rather than copy any outlet, never to invent a statistic or a quote, and to work only from the supplied material. If the grouped headlines turn out to cover more than one event, the group is rejected and discarded rather than written up.

Left through right are relative markers, not fixed ideologies; where the centre sits depends on the country and the era, and the whole scale shifts over time. All five are mainstream positions. The outer two are not “far left” and “far right” — terms that describe something outside the mainstream and are mostly used as accusations — but the strongest version of a committed partisan argument.

What separates left from centre-left is depth of critique rather than volume: the left asks whether the system itself produced the outcome; the centre-left accepts the system and argues for reform inside it. The same distinction separates right from centre-right. The Backgrounders page covers the traditions behind each position, with links to outside sources.

Beneath an article in “How each outlet covered it”, a short note may point at a specific difference in wording — a loaded word, an omission, an active-versus-passive construction — compared with how other outlets described the same event. The model is instructed to draw only on that outlet’s own headline and description, and each note sits beside the link to the original so you can check it.

Notes describe words, never motives. A note may say what wording emphasises or omits, or how it lands on a reader. It may not say why an outlet chose it — Prism cannot know what a newsroom intended and does not claim to. That rule is in the model’s instructions and enforced in code: a note that asserts a purpose is discarded rather than published.

Because every feed is re-read three times a day, Prism notices when an outlet rewrites a headline after publishing. The before and after are recorded — beside the article on its story page, and together on the Headline Tracker. Live blogs, which retitle one address all day, are left out.

When a new, datable development of an event already covered arrives — a vote after a proposal, a ruling after a filing — the new story is linked to the earlier one. The story page then shows the chapters in order, with the previous chapter’s five framings one click away so you can see what moved.

“Talk it through” on a story opens a chat with an AI moderator that steel-mans the strongest argument on each side and ends with a question that makes you weigh the trade-off. It never declares a winner.

“Talk to a viewpoint” in the header lets you debate an AI persona built to argue one political tradition in good faith, from the inside — instructed to name its tradition’s internal disagreements, correct misconceptions about its own side, and concede what is genuinely hard to answer. It is a synthesis: not a real person, not modelled on one, and not an official voice for any party or movement.

“Debate two viewpoints” on a story page lets you pick two personas and read a six-turn exchange about that story, followed by a neutral note on where they actually agree and what really divides them. Each pairing is written once, when a reader first asks for it, then saved for everyone.

Spot the framing shows a real headline from the past few days and asks whether it reads from the left, the centre, or the right. The answer is Prism’s placement of that article — made from the piece’s own words, with the quotation behind it shown on the reveal — and never the outlet’s reputation. The set changes daily, streaks are kept in your own browser, and every headline links to the original.

Missing one isn’t a failure. Plenty of headlines are written straight; the useful part is noticing which word tipped you off — or fooled you.

The account, framings, placements, notes, and chat replies come from an AI model and are published without human review. They can be wrong, can miss context, and carry the limits and biases of the material they were given and of the model itself. Grouping articles by event is imperfect, and the political-relevance filter sometimes lets something odd through or drops something it shouldn’t.

Prism has no newsroom, no editors, and no reporters. It publishes no original reporting and is not a news outlet. It is a lens for looking at other people’s reporting — and a lens can distort. Use it to notice how framing works and to find the primary sources, then read those yourself. The outlet links on every story go straight to the originals.

If something here is wrong — a framing, a note, where a piece was placed, a story that shouldn’t exist — use the “Give feedback” link at the bottom of any page or the “Flag” link on a story. It goes straight to the person running the site, and it is read.

If you are a publisher and want your material removed, say so and it will be removed. You don’t need a lawyer, a formal notice, or a reason. Write to usetheprism@gmail.com and it will be dealt with the same day wherever possible.