About

The Research tools lets you pick the records and the dates yourself and leave with a permalink, a CSV, and a citation that still resolves next year. If you are quoting a number from this site in something you publish, build it there first.

Open Research tools

Every row is a document: a headline, a record, a filing, an episode. It carries the outlet or organization, the date, and a link that opens the source in a new tab. A bare date is the day the document was published. A date the row calls collected is the moment this site first saw the item, shown because the feed it came from carried no publication date of its own, or carried one in the future. The record keeps which of the two it has rather than printing one as the other. Two small labels come from published rules that anyone can re-run. Source type is a lookup of the domain in a public registry: primary document, organization statement, government record, wire service, news outlet, aggregator, or commentary. Wording reports which kind of verb the headline uses: reports what happened, reports an assessment, uses conditional wording, or mixes them. In headline comparison, a headline is set beside the passage it rests on and compared on 6 things a claim can differ by: how certain it is, what quantity it names, how wide it is, who it is attributed to, whether it describes an attempt or an outcome, and whether it happened at all. The result says which of the 6 differs and shows the words on both sides. A figure is only ever compared with a figure counting the same thing. None of these labels is a judgment about accuracy. Everything they rest on is printed so you can check it against the two texts yourself.

The Monitor does not report news. It measures how AI is reported, against the documents that reporting cites. It breaks no stories and runs no correspondents. Beyond counts and the record of what was said, the only writing of its own it publishes is a small number of labelled editorial summaries, 0 of them today: each is marked on its story page as the Monitor's own words, names who wrote it, and is checkable against the sources it names. The Monitor never republishes an article's own text. Everything below is a limit on those measurements.

It does not independently verify third party reporting. It does not say why anyone said anything. Headline comparison finds only what its published rules cover: certainty, scope, quantity, polarity, attribution, outcome. It is silent on everything else, including whether a headline is about the right subject at all, and silence from the rules is not a finding that a headline and its source agree. The wording rule is a word list, and a document can use conditional words for a confirmed event. Volume counts depend on the topic query and on how well each language is covered upstream. Machine translation of headlines can be wrong and is always labeled. Market forecasts are estimates by the firms that sell them.

The archive is being filled in backwards. Live collection starts on the day the collector was switched on (Sep 14, 2026, this snapshot's own totals.collection_started) and grows forward from there. Nothing published before 1 January 2026 is stored by any path. A headline already more than thirty days old when the collector first sees it is kept, but marked as filled in rather than watched, so it never moves a freshness figure or the index. Some items on the coverage chart are dated before collection started, because a feed carries its recent back catalogue and those items arrived with the first fetch. Those bars are drawn hollow and say so, and they measure how deep a feed's back catalogue runs, not how much was published that day. One lane now fills the record backwards towards Jan 1, 2026: the GDELT Global Knowledge Graph file archive, the same file server and the same parser the live GKG lane already runs, walked one fifteen minute file at a time. Items it recovers keep the date the archive gives them and are marked as filled in later, and the days they land on are drawn apart from days this site watched. That population is not this wire's own. GDELT indexes a much wider set of outlets at several times this site's daily volume, so the lane keeps a bounded sample from each file, stores it under the wide topic, shows it on the world and topic pages, and never counts it in the Coverage Intensity Index. Read those days as a wider record of what was published, not as what this site would have collected. Until the record is old enough to hold a year, no reading here should be taken as a trend.

A correction is dated, states what was published and what is correct, and is shown on the item it corrects and on this page. Corrections are never silent. Anyone can request one at [email protected].

Any named person or organization can send a reply. The editor verifies it, then it is published verbatim beside the item it answers, with a dated correction note if the record changes. A contact address is optional; it goes to the editor to confirm identity, is never published, and is not stored on this site.

The site sets one cookie, a random identifier that gives you a pseudonym for notes. It holds no address and no name. Signing in is optional and never required to read anything; the account itself stores a one way code derived from your Google account and a session id, and nothing else. No password ever reaches this site, and the account page can delete all of it. Creating an account gives the Monitor your name and your email address, which the card says plainly before you press anything. They, the date, the page you came from, and the sentence you agreed to are written to one private spreadsheet the editors keep, which this site cannot read back. They are used to tell you when the Monitor starts sending something and for nothing else, they are never sold, rented, or handed to any other party. Deleting your account from the account page erases the name and the address in the same action, leaving only a dated line with nobody attached to it; a note to [email protected] does the same within 72 hours. Nothing is sent from this site today. Google Tag Manager loads only if you allow measurement in the cookie notice, and reports page views, read depth, section views, video plays, and outbound clicks; those events carry no name, email, or note text. Choose essential only and the container is never fetched. The footer link reopens that choice on any page. To limit abuse, posting, rating, and reporting are counted per pseudonym and per a salted hash of the connecting address for the current day; those counters are deleted within two days and are never shown or shared. Note and reply forms use Cloudflare Turnstile to block automated posting; Turnstile receives the request details it needs to make that decision. Outlet and organization marks load from Google's favicon service, s2/favicons, keyed only by the publisher's domain, the same way a flag loads from Flagpedia or a video thumbnail loads from YouTube; that request carries no name, session id, or page content. Where that service has no icon for a domain, or the request fails, the mark falls back to a letter mark drawn in the site's own type, fetched from nowhere. Where a person's portrait is shown, your browser loads that image directly from Wikimedia Commons or Wikipedia when you open the page, the same way it already does for a country flag or a video thumbnail, and that photo's author and licence are credited on the photo credits list. A video loads from YouTube's privacy enhanced player only when you press play. A reply's optional contact address goes to the editor's private channel and is not stored on this site. The capture form at the foot of the home page stores a name and an address the same way the sign in screen does, in the same spreadsheet, under the same terms. Requests to remove personal data from a note or a reply go to [email protected] and are handled within 72 hours. Paciva's data handling, security practices, and policies are published at trust.paciva.ai.

One snapshot per day in JSON, byte identical on re-fetch, at a dated address that never changes, so a citation pins to a fixed artifact. This site's own layer is CC BY 4.0: what the Monitor wrote, the artwork it made, and the structure of the record. That licence covers only those things. It does not extend to a publisher's headline, to a passage quoted from a document, or to any third-party asset shown here, all of which stay with their own rights holders and are not licensed by this notice. Upstream sources keep their own terms, listed on the technique page, and what is shown, withheld and still under review is on the rights page. Cite as: AI News Monitor (2026), Paciva. Collected from publisher feeds, Google News, Hacker News, YouTube and the GDELT Project's Global Knowledge Graph.. Current snapshot: 2026-11-11.json. Live state: state.json.

Every collected item carries a stable citation id: AICM- followed by the item id, itself a hash of the canonical URL, so a citation written today resolves to the same document later. The ledger endpoint at /api/ledger returns items as JSON with filters for since, until, topic (a published rule slug), language, country, domain, and cursor pagination, capped at 200 rows per call. Pin a citation to a fixed artifact with the immutable daily snapshot at /api/state/snapshots/<date>.json. Site layer CC BY 4.0; upstream terms unchanged.