In short

The news on AI.info is found, written and checked by advanced AI systems, in an automated newsroom conceived and designed by Gabriele Masetti, the founder of AI.info.

Before it can go out, a story passes through eleven steps. Code looks for every quotation and every figure on the pages the story cites, and two AI checkers, from different families of models, read it against those pages, each on its own. When they hold a story, an arbiter can lift the hold only with lines of the pages that code finds there. The newsroom publishes nothing on its own: before publication, an AI reading agent reads every story against its sources once more, and decides whether it goes out.

How a news story is made

Every story goes through these eleven steps, always in the same order. Beside each step is who does it.

  1. Before the writing

    1. 01

      The story

      AI

      First, an AI system traces the story back to its original source: a press release, an official blog, a regulator’s or a court’s record, a scientific paper. Stories are found in the official channels of companies, agencies and institutions, or in other outlets’ reports, but the article is always based on the original source, and cites it.

    2. 02

      The checks before writing

      Code

      Before the writer is called, code compares the story with what the site has already published: a story told from the same source as a recent one stops here. So does an event more than three days old, and a source page that cannot be opened or holds too little text to read.

  2. The writing

    1. 03

      The writing

      AI

      An advanced AI model writes the article from the source, searching the web for further information. Every page it takes information from is cited at the end of the article; anything that cannot be found on one of those pages is stopped by the checks that follow.

      The pages it read are recorded with the draft, and a private copy of their text is kept with the day it was read, so that what a page said can be shown even after the page changes.

  3. The checks

    1. 04

      The machine checks

      Code and AI

      Code looks for every quotation and every figure of the draft on the pages it cites. A quotation or a figure that is on none of them holds the story, whatever any model says. Code also measures the house standard: the length, the phrases that mark machine-made prose, any sentence that talks about the checking instead of the news. And a model compares the draft with the recent stories, so that the same news is not published twice.

    2. 05

      The first checker

      AI

      An AI checker, whose only task is to find what is wrong, reads the draft against the pages it cites. Our own reader fetches those pages and, when a site turns it away, tries other routes, among them the publisher’s own copy, the public records of papers and of company filings, and an archived copy. The checker compares every name, figure, date and quotation, and the person each quotation is credited to. It lists every sentence that states a fact no page states, a figure that lost the page’s “up to” or “about”, a link that does not lead to the story, a source that is not the original. It searches the web only when a cited page could not be read.

    3. 06

      The second checker

      AI

      A second AI checker, from a different family of models, reads the same draft against the same pages on its own: it does not see what the first one found. Their results are added together. A story is held when either checker holds it, and the second can only add findings, never clear what the first one found.

    4. 07

      The arbiter

      Code and AI

      Two checkers raise more false alarms than one. So when a story is held by the checkers’ findings alone, an arbiter, another AI reading, is asked about each finding, with the sentence it names and the passages of the pages nearest to it, and must answer with a line of the page. Code then looks for that line on the page. The hold is lifted only when every finding is shown to be a false alarm by a line that code finds there; otherwise the story stays held. What code found itself, a quotation or a figure on no page, is never put to the arbiter.

    5. 08

      The correction

      Code and AI

      A held story is not thrown away. It goes back to a correcting model with each finding, the sentence it names and the passages of the pages nearest to it, and the instruction to fix the findings and nothing else. Code then cuts any sentence the correction added about the source or the checking, compares the corrected story with the first one sentence by sentence, and records every change no finding asked for. The corrected story goes through the checks again from the start, up to three times.

  4. Before publication

    1. 09

      The reading before publication

      Reading agent

      The newsroom publishes nothing on its own. About every two hours, an AI reading agent takes the waiting stories one by one, opens every page each one cites, and reads the story against them: every name, figure, date, quotation and attribution, the headline, the age of the event, whether the story belongs on the site. The checkers’ findings are hints to it, not verdicts. Then it decides: to publish the story, to correct it in the source’s own words, to keep it out of search engines if it has too little substance, or to take it down. A page it cannot open is “not read”, never proof that something is missing.

      Before a correction is made, a second reading checks the finding on the page and must quote the line that contradicts the story; a finding nobody can quote is not corrected. Each correction is recorded with the story: the words before and after, and the page they rest on.

    2. 10

      Publication

      The article goes out signed AI.info Team. Beside the byline are the European Union’s icon for content made with AI and our check mark with the word “fact-checked”, together with the number of sources consulted. The list of sources is at the end of the article.

  5. After publication

    1. 11

      After publication

      Code and AI

      A published story can be read again at any time. If it turns out to be wrong it is corrected, and the change is recorded; a story that should not be on the site is taken down. The Italian translation is checked by code (figures, links, quotations and who said them) and read again against the English by another AI model, not the one that translated it; a translation that fails is not shown.

The news is filed under these desks: Breakthrough, AI Research, Industry News, Tools & Products and Ethics & Policy.

What the founder does

The system described on this page was conceived and designed by Gabriele Masetti, the founder of AI.info. The steps, the order in which they run, the rules on sources and the way errors are corrected are the founder’s decisions, and so is the rule that runs through all of them: an error is called an error only with the line of the page that shows it. New checks are tried on drafts with known errors before they are switched on.

The founder reads every story, but we cannot guarantee that this happens before publication; when an error turns up, the article is corrected or withdrawn. The founder is responsible for what we publish, and can be reached at hello@ai.info.

Designed by

What we do not guarantee

  • That every article has been read by a person before publication.
  • That the checks catch every error. When one gets through, the article is corrected and the correction is shown.

Sources

Every article names its original source (a press release, an official document, a record, a study) and lists every other page it took information from. Outlets that only repeated the news help us cross-check, but we do not cite them as the source.

If the original source cannot be found, the story is not published.

Corrections

When we find an error in a published article, we correct it in the text and record the change. An article that should not be on the site is withdrawn.

Report an error

Found an error? Write to us from the contact page, choosing “Report an Error”, or email hello@ai.info. Tell us the article’s address and the sentence that is wrong; if you can, add a link to a page that gives the correct information.