Across 2,693 AI industry events, 16.9% were covered by more than one article, at an average of 1.32 articles per event. That average hides most of the story. Broken out by category the rate runs from 6.0% (RESEARCH_AND_PAPERS) up to 33.8% (MODEL_RELEASE), a spread of roughly 6x. Some kinds of AI news get written about by everyone at once, and some get written about once.
Counted 2026-09-08 from Scoopkit's own /v1/events endpoint across all 10 categories, 2,693 events in the current archive window. "Covered more than once" means the event's article_count is 2 or higher, which is Scoopkit's clustering deciding that two or more separate articles describe the same story. This page regenerates automatically rather than being published once.
Duplicate coverage by category
| Category | Events | Covered more than once | Rate | Articles per event |
|---|---|---|---|---|
| MODEL_RELEASE | 139 | 47 | 33.8% | 2.04 |
| HARDWARE_AND_INFRA | 117 | 33 | 28.2% | 1.45 |
| LEGAL_AND_IP | 82 | 21 | 25.6% | 1.51 |
| ORG_AND_PEOPLE | 92 | 22 | 23.9% | 1.34 |
| SAFETY_AND_POLICY | 325 | 74 | 22.8% | 1.59 |
| FUNDING_AND_DEALS | 333 | 65 | 19.5% | 1.31 |
| AGENTIC_AND_PROTOCOL * | 23 | 4 | 17.4% | 1.39 |
| PRODUCT_AND_API | 573 | 99 | 17.3% | 1.25 |
| GENERAL_INDUSTRY | 541 | 61 | 11.3% | 1.17 |
| RESEARCH_AND_PAPERS | 468 | 28 | 6.0% | 1.09 |
| All categories | 2,693 | 454 | 16.9% | 1.32 |
Categories marked with an asterisk have fewer than 40 events in the window, so treat their rate as directional rather than solid.
Why the spread is this wide
MODEL_RELEASE sits at the top at 33.8%, averaging 2.04 articles per event and peaking at 17 for a single story. That fits how those stories break: one announcement, then simultaneous coverage from every outlet that follows the space, plus analysis and reaction pieces within the same day.
RESEARCH_AND_PAPERS sits at the bottom at 6.0%, averaging 1.09. Coverage there is thinner per item and spread across far more items, so duplicate pileups are rarer even though the raw volume is high.
The practical read: deduplication earns its keep unevenly. If you only ever query RESEARCH_AND_PAPERS, clustering is doing comparatively little for you, because that coverage arrives pre-deduplicated by the nature of how papers get written about. If you track model releases or safety incidents, it is doing most of the work of making the feed readable.
The most piled-on stories in the window
| Articles merged | Category | Event |
|---|---|---|
| 20 | SAFETY_AND_POLICY | OpenAI and Hugging Face partner to address security incident during model evaluation |
| 17 | MODEL_RELEASE | Path to Astra: critical capabilities and frontier safeguards |
| 14 | MODEL_RELEASE | Anthropic launches Opus 5 |
| 13 | SAFETY_AND_POLICY | OpenAI releases official report on Hugging Face breach involving rogue AI model |
| 12 | MODEL_RELEASE | Anthropic says Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads and up… |
These are the events where a raw article feed would have handed you the same story many times over. The top entry alone would be that many rows in a keyword search, each with a slightly different headline and timestamp, and each requiring you to work out that they are one story.
What this measurement cannot see
Under-merging looks identical to a story nobody covered twice. If clustering fails to group two articles about one event, the result is two events with article_count of 1 each, which is indistinguishable here from two genuinely separate stories. That would understate these rates. The clustering thresholds are deliberately conservative for this reason: a wrong merge destroys information, a missed merge merely fails to save you a row.
Article counts are capped by what the pipeline sees. These numbers reflect coverage from the configured source set, not every article published on earth. A story covered by 40 outlets where 6 are in the source list shows up as 6.
The window moves. This is a rolling archive, not a fixed historical sample, so a single unusually heavy news week shifts the per-category rates noticeably.
You can reproduce every number here with a free API key: article_count is a free-tier field, and so is primary_category.
FAQ
What counts as a duplicate?
Two or more separate articles that Scoopkit's clustering assigned to the same event. The mechanism is embedding similarity with an entity-overlap check and a model tie-break on borderline pairs, described in full in how deduplication works.
Why do research papers barely duplicate?
A preprint is a single artifact with a canonical URL, and coverage of it tends to be one write-up per outlet that bothers at all. Compare that with a major model release, which produces a launch post, several outlet write-ups, an analysis piece, and a benchmark reaction, all within a day.
Does a high rate mean the category is noisier?
It means the raw article stream is noisier there. The event output is one row per story either way, which is the point.
How often does this page update?
Automatically, on a recurring basis. The figures reflect the window current on 2026-09-08. Query the API directly if you need live numbers rather than a dated snapshot.