Skip to content
PRISM

How it works

How Prism checks a story.

Six steps, from the link you paste to the report you keep. Each one shows its sources, so you can check Prism too.

Beside each step: what it found in one sample story, a city council's vote on a harbor rail line.

  1. 01

    Read

    Paste a link, a YouTube link or the text itself. Prism pulls out the article's text, title, author and date, or reads a video's captions. It never works around a paywall, and if it finds only part of an article, it stops rather than judge a fragment.

    citydesk.example/metro/harbor-rail-extension

    9 sentences, 104 words read

  2. 02

    Sort

    An AI model sorts every sentence into a claim, an opinion, rhetoric or neutral. Naming a source doesn't make a fact any less checkable, so a quoted figure is still a claim. Longer sentences weigh more in the percentages, and a sentence the model can't place is left out rather than guessed.

    Fact-led: 46% claim

  3. 03

    Trace

    Prism reads the website's public registration record and where it is hosted, adds the owner for outlets it knows, and warns when an address imitates one of them. It searches for the outlet's funding and reputation, and for the author when one is named. A label such as political leaning appears only when it cites a source. An outlet's research is reused for 30 days and kept, never the author's, for the owner to review for the Library's outlet pages. An author note counts only sources that also name the outlet, because people share names.

    Owned by Harbor Media Group

    Domain registered March 2011

  4. 04

    Check

    Prism picks up to 10 checkable claims, works out the questions that would settle each one and searches for answers, with one search aimed at rebuttals. A verdict needs two independent sources that agree, an official record or a published fact check. Otherwise the claim stays Unverified, which means unconfirmed, not false.

    Supported

    2 independent sources

  5. 05

    Compare

    Prism gathers up to 24 other reports of the same event and lists the facts they establish before it reads the article, so the article can't steer what counts. Then it marks each fact covered, softened or missing, and never calls a fact missing from a part of the article it couldn't read.

    2 covered1 softened1 missing

    4 facts from 4 other reports

  6. 06

    Report

    Pick the sections you want and save them as one PDF, put together in your browser, with every citation kept.

    One PDF, 6 sections

    Made in your browser

Where it can be wrong

  • AI makes mistakes. The sorting, summaries and verdicts come from AI models, which is why every verdict shows its sources.
  • Search misses things. Very recent events, pages behind logins and sources in other languages can be out of reach.
  • English only. Prism is built and tested for articles written in English.
  • Video means YouTube. Prism reads YouTube captions, so a video without them can't be read.
  • Results can shift. A second run can differ by a few points, and a verdict reflects the evidence found on the day it was checked.
  • Ownership data has gaps. It is strongest for Indian outlets and the major international groups.

What runs it

Google Gemini is the main AI model and Groq the backup. Tavily runs the web searches. rdap.org, Cloudflare and ipwho.is look up who registered a website and where it is hosted. What each one receives is set out in your data. Reading is open to everyone. Refracting a story needs a Google sign-in and has a small daily allowance per person, so the free services stay available.

Refract a story