How It Works

Taxonomy v1, field schema frozen Sep 14, 2026. Every entry below is dated from this repository's own history, not written from memory.

Two hourly component series: coverage volume (publisher feeds and Google News, all languages, wire duplicates merged) and discussion volume (Hacker News). Each is divided by its own 90-day median and multiplied by 100, so 100 means the median day. Default weights: 0.67 and 0.33, the ratio originally published for these two. The index is the weighted mean of the two. It was defined with four components until 18 September 2026. Broadcast minutes came from GDELT’s television index, which carries nothing dated after 8 October 2024, and front-page prominence came from GDELT’s Frontpage Graph, which has served empty files for every hour tested in 2026. Both were removed rather than carried as permanent blanks, and neither ever contributed a reading. The Internet Archive television collection that GDELT indexes is still current, so broadcast attention may return on that source directly; it will be announced here if it does. It is undefined until 30 days of baseline exist. The components are always shown beside it and Research tools lets you set any weights. The sensitivity table fills in once a baseline exists.

Fourteen topics, one published keyword rule per topic, applied case-insensitively to the headline and, for non-English items, to its machine translation; anyone can re-run it by hand, and the match itself is the reason an item carries the label. A second, separate pass runs the same headline past a text model (Meta Llama 3.1 8B, run on Cloudflare Workers AI at temperature zero, falling back to two smaller and larger Llama models) given the same fourteen topics with a one-line definition each and asked which apply, at most three; its answer is stored on the item with the model name and shown for review, but a topic count and a topic label are built from the rule alone, because a model guess is not, on its own, a headline-based reason. Audited 2026-09-21: counting either labeler put Apple's Siri settlement and waitlist headlines under Power & Datacenters on the model's word alone, corrected here by no longer counting it. Every topic page states both figures beneath its totals, rule and model, so the gap stays visible. An item can carry several topics; an item the rule does not match still appears on the wire. Neither labeler reads past the headline, and neither says anything about whether the story is right. The Worker also runs one per-topic collection query per five minute cycle against GDELT, rotating, so every topic's coverage refreshes about hourly; those query strings are collection vocabulary and live in the repository beside these rules. The map on the Power and Datacenters page is a separate curated record: every site entry is dated and linked to the report it is drawn from, status and scale are stated as reported by that source, and dots sit at the named city, county, or region, not the parcel.

Two items are paired only when the same topic rule matched both and the record places them on the two opposing catalogued positions for that specific question; a third position that answers a different question, such as on regulation, is never paired against either. Only warning and pushback, the record's two catalogued positions on whether a catastrophic-risk claim holds, are treated as opposed on that question; the record's third catalogued position, regulation, answers a different question (whether AI should be regulated) and is never paired against either, so a topic where the record holds no genuine opposing pair shows fewer pairs, or none, rather than a manufactured one. Each side of a pair is labelled by what kind of evidence it is: a headline is the title a publisher or a platform gave the piece, never the subject's own spoken words; a post is shown quoted because it is the person's own published text; a statement is the record's own paraphrase and is labelled as one. The three catalogued positions are named plainly (Warning, Pushback, Regulation), never by an assumption about intent.

Two ways in, and the page marks which one applied. The first is a record entry: an editor writes the person into the record with a role, an affiliation, and dated links to what they said or did. Those cards carry statements and actions. The second is the coverage itself. Every cycle the collector reads the headlines it has stored over the last 7 days and pulls the names out of them, and a name that clears the threshold gets a card marked Found in the coverage. Such a card claims nothing except what produced it: the name, the headlines that carried it, and the outlets that ran them. No role, no employer, and no photograph is attached to a name found this way, because none of that is in a headline.

The rule, in full. Discovery reads only sentence case English headlines. About half of publishers capitalize every word of a headline, and in those a capital letter carries no information, so a name is never introduced from one; they are still counted. Machine translated headlines are counted and never used to introduce a name either, because translation does not preserve names: one engine returned a Korean deputy prime minister under a name that belongs to no one. Inside an eligible headline, a run of two or three capitalized words is a candidate unless a word in it is a known place, country, demonym, organization, month, day, role or ordinary headline vocabulary; unless a comma, a conjunction, a quote or a possessive runs through it; or unless the headline carries no person signal at all, meaning no speech verb, role, honorific, personal pronoun, or something only a person does. A candidate is published once it appears in at least 3 headlines from at least 2 different outlets inside the window. Counting then runs over every headline in the window, matching the full name or a surname long enough and uncommon enough that no one else on the roster carries that word in their name. Both numbers on a card come from that one pass, so the count under a card and the order of the panel always agree.

What this buys, and what it costs. It buys a page that follows the news instead of following an editor: a person who turns up in this morning's coverage is on the page this afternoon, and their name joins the rotation the collector queries by name, so the record keeps finding them. It costs recall. The person signal requirement and the sentence case requirement both throw away real names, on purpose, because a wrong name on this page costs the whole record its credibility while a missed one comes back on the next headline. A name that should not be here is a correction, and corrections are published.

A scheduled job runs every 5 minutes and enqueues jobs by cadence: per cycle sources every run, hourly sources in the first slot of each hour, daily sources in the first slot of 06:00 UTC. A consumer fetches each source, deduplicates by canonical URL and title cluster, derives the fields above, and writes to a database. Publisher feeds and the GKG lane carry general coverage, so a headline enters the record only when it clears the topic rule, version 3, published 3 October 2026. That rule has four parts. Phrases are matched without regard to case: artificial intelligence, machine learning, large language model, chatbot, superintelligence, frontier model, AI safety, AI risk, existential risk, AI regulation, and a named frontier lab or model, with equivalents in Chinese, Japanese, Korean, Arabic, Hindi, Russian, German, French, Spanish, Turkish, Greek, Hebrew, Indonesian, Vietnamese and other listed languages. Acronyms are matched with case: AI, A.I., LLM, AGI, KI, xAI and GPT followed by a digit. Case matters here because lowercase ai is Italian and French for to the, and lowercase ki is an ordinary Turkish conjunction, so matching them loosely admits a jewellery advertisement and a traffic report as AI coverage. Product names that are also given names or ordinary words, Claude, Gemini, Grok, Llama, Copilot, Codex, Qwen and Sora, need a second AI term beside them or a second such name before they count. Robotics and physical AI count too: humanoid, robotaxi, self driving, autonomous driving, embodied AI, world model and named robotics and autonomous driving companies such as XPeng, Unitree and Waymo, with equivalents in the listed languages; the bare word robot counts unless the headline is about a household or consumer robot. A headline admitted by this part alone is stored under the wide topic and never counted in the index. Version 3 added the robotics part after version 2 was found to refuse every headline about XPeng's humanoid robot production line that did not also say AI. Measured against every GKG file for 8 September 2026, version 3 admits 13 per cent more headlines than version 2 and removes none. Version 1 matched everything without case in one list. It was replaced after the three part rule was measured against 4,856 real GKG rows, where it removed 32 items that were not about AI, and against the 200 most recent items already on this wire, where it removed none. Items from Google News are already selected by the published topic query. Nothing published before 1 January 2026 is stored by any path. A headline more than 30 days old at first fetch is kept on the live path but marked as filled in rather than watched, the same mark the archive lanes below give their rows, so it never counts toward the index or a freshness figure. Collecting now: publisher feeds from the OPML, split across four queue messages so every feed is polled every cycle; a topic discovery lane that walks fourteen topics across ten locales, four queries a cycle, resolving each hit to its publisher before anything is stored and skipping headlines whose title cluster is already on file; a per person discovery query that rotates through every name on the watchlist about once an hour; Hacker News counts; curated YouTube channels; curated podcast show feeds; the prediction market lane (fetches Polymarket and Kalshi, but a venue is fetched only once its terms are read and cleared, so it stores nothing until one is: see rights); the GDELT GKG raw file lane, described below; and the money ledger (FEC committee totals and large itemized receipts, and SEC EDGAR filings, for a roster of AI companies, people and committees, a few roster entries a cycle, each row linked to its filing); and the editorial Sheet. Listed but not collecting, so nothing on the site depends on them: Congress.gov, LegiScan, the Federal Register, the Senate LDA lobbying lane (built, and it collects once the Senate's key is set), arXiv, OpenAlex, Wikipedia pageviews, the original Podcast Index API (replaced by the shows' own feeds above), the GDELT Web NGrams fallback (built, collecting), Metaculus (reputation points, not a price), fact check feeds. Every GDELT lane was removed on 18 September 2026 after each was tested directly. The document API refuses datacenter address space with a 429 and has no key, account or paid tier, so there was nothing to request; the same query answered normally from a residential connection the same day. The television API answers but its index carries nothing dated after 8 October 2024. The Frontpage Graph still publishes an hourly pointer file and every hour tested in 2026 decompresses to nothing. Over this site's whole run those eight lanes contributed no rows at all, so their removal changed no figure published here. GDELT's raw file archive is a separate service and it answers every network tested, so it is now collected as its own lane. Four files an hour, two English and two translingual, each capped, each parsed for the headline GDELT stores in the last column of its Global Knowledge Graph. Those rows are stored under the wide topic, appear on the world page, and are never counted in the index, because that corpus is a different publisher population: about one per cent of it overlaps the outlets listed on this page and it runs several times their daily volume. It is additional coverage under its own label. The same file server is also being walked backwards to fill the record in to Jan 1, 2026, one fifteen minute file at a time, from the day collection started. Those rows carry the date the archive gives them, are marked as filled in later rather than watched, land under the same wide topic, and are counted in the index no more than the live GKG rows are. More lanes fill the record in to the same date through the collection methods above: each curated channel's full list of uploads; each podcast show's whole feed; the Hacker News count for each past hour; each Kalshi market's daily closing price from before this site first saw it; and each publisher feed's archived copies in the Internet Archive's Wayback Machine. Every row they file is marked as filled in and never counts as a day the collector watched. The filled Hacker News counts give the index's discourse component its full ninety day baseline, route hn_backfill. Headlines in languages other than English are translated by Cloudflare Workers AI (m2m100) and always shown beside the original, labeled machine translation. The site's own interface text can be read in other languages through the language control; those strings go through the same model, are cached so each is translated once per language, and quoted material is never translated. Ask Pax answers reader questions from the Monitor's own search results, Llama on Workers AI, and every answer links the results it cites. After each cycle the current state is written into the HTML and the site is redeployed, so a crawler sees a complete page; open pages also poll the current state every minute. One immutable snapshot per day is served at a dated URL. The editorial tables (source registry, primary source records, headline pairs, corrections, replies) live in a spreadsheet the editor maintains and are re-read every cycle.

24 entries in the source registry, 17 collecting today. Generated straight from worker/src/sources.js and worker/src/consumer.js, so this table cannot say something the paragraph above does not, or drift from it unnoticed.

A story groups every article the Monitor holds on one event under one address that outlives any single run. Assembly s1, two passes. The first clusters a rolling window by title similarity, tuned to avoid joining different events. The second merges clusters that are the same event told twice, at a weighted centroid similarity threshold of 0.25 measured against a window of real headline pairs rather than chosen by hand; that pass looks only within a 24 hour window and needs at least 2 outlets before two clusters count as one event. A story stays open for 168 hours of rolling coverage and gains items as they arrive; an editor can merge or flag one for review by hand, and every such change is dated on that story's own corrections list. Merging too eagerly destroys a distinction a reader needs; merging too little leaves two pages an editor can combine in one action, so the threshold favors leaving events apart.

A story lists the companies, organizations and people it involves. A name is listed when the headline of one article in the story names it, when the lists of names GDELT takes from the text of at least two of the story's articles include it, or when one article's list names it at least three times. Only names, numbers and dates are kept from article text, never its wording. A name the record cannot match to exactly one known company, organization or person is left off, not guessed.

This site's layer (the derived fields, the index, the snapshots) is CC BY 4.0. A permissive license on this layer does not change upstream terms, so each source's terms are listed. Cite as: AI News Monitor (2026), Paciva. Collected from publisher feeds, Google News, Hacker News, YouTube and the GDELT Project's Global Knowledge Graph.. Excluded on purpose: Reddit, NewsAPI.org free tier, Perigon free, MediaStack free, Brave default tier, Bing, and any Google News scraper, each for a licensing or storage reason recorded in the repository.