Where AI news actually comes from
We grouped 2,612 AI news articles into 539 distinct stories and looked at what each story’s central claim actually rests on. The most common origin of an AI news story is a company press release: 58% trace back to one. Independent benchmarks account for 3%. Peer-reviewed or preprint papers account for 5%.
Where does an AI story come from?
| Origin of the underlying claim | Stories | Share |
|---|---|---|
| Company press release | 311 | 58% |
| Rumour or anonymous source | 81 | 15% |
| A shipped, usable product | 69 | 13% |
| Peer-reviewed or preprint paper | 27 | 5% |
| Speculation | 19 | 4% |
| Independent benchmark | 17 | 3% |
| Demo | 14 | 3% |
Roughly three quarters of what arrives as “AI news” is a press release or a rumour. Around one story in thirty involves anyone outside the company measuring anything.
Do outlets disagree? Mostly not — and that is the problem
Of 539 stories, 383 had no disagreement at all. Every outlet covering the event reported it the same way, with the same framing and the same figures. Only 156 stories — 29% — contained a genuine divergence.
That is not reassuring. It is what you would expect if most coverage of an event is a rewrite of the same source, because it is. Unanimity across forty outlets is a weaker signal than disagreement across four, not a stronger one.
How much of it is the same article twice?
282 of the 2,612 articles were syndicated or affiliate copies — the same piece republished under a different masthead. That is 11% of the corpus. We group them under the originating outlet so a story’s outlet count is not inflated by wire copy.
If you read AI news in an ordinary feed, roughly one in nine items you scroll past is something you have already read.
How thin is the average story?
306 of 539 stories — 57% — were covered by exactly two articles. The distribution has a very short head and a long, thin tail: 42 stories drew ten or more articles, and three drew more than a hundred.
| Story | Articles | Outlets |
|---|---|---|
| Gemini breached three companies during a security test | 148 | 138 |
| Antitrust suit against Anthropic, OpenAI, Google and SpaceXAI | 139 | 133 |
| Trump’s proposed “AI Force” and “AI Czar” | 108 | 103 |
What we take from this
Two things worth carrying around:
- Before acting on an AI capability claim, ask who measured it. If the answer is the company that built it, you have an announcement, not a result. That is the majority case.
- Treat unanimous coverage as a weaker signal than divided coverage. Forty outlets agreeing often means forty rewrites of one source.
This is why every story on this site labels what the claim rests on, and why the disagreement between outlets is shown rather than smoothed away.
What this does not show
- The corpus is not the whole internet. It is what our source list reaches — around 127 publication feeds plus a set of news searches. Paywalled reporting is under-represented, because we often cannot read it.
- Evidence grade describes the claim, not the publication. A press release carried by an excellent newspaper is still a press release.
- The grading is automated and not yet hand-audited. A language model assigns it from the article text. We have not checked a large enough sample to publish an error rate, so treat these percentages as indicative rather than precise. Auditing a sample is the next thing we intend to do to this number.
- One snapshot. Whether 58% is normal or unusually high for this industry, we cannot yet say. Ask again in six months.
Figures are from the AI Coverage corpus as of 22 September 2026. The method — including where it fails — is described on the method page.