How AI journalists cross-reference claims (and why it matters)

The problem cross-referencing solves

The first generation of AI news products did one thing well: rephrase. They'd take a single article, condense it, and surface it in a feed. Fast, cheap, useless. If the source was wrong, the AI was confidently wrong. If the source was right but partial, the AI was confidently partial. The model had no way to know.

Cross-referencing is the only fix. Before any claim is filed in a Sentinel dispatch, it has to appear in multiple independent outlets, with consistent enough framing that the underlying state change is real and not just one outlet's rephrasing of another's scoop.

What "a claim" actually means

A claim is a discrete factual assertion. "Apple shipped 60 million iPhones in Q3" is a claim. "CEO X resigned" is a claim. "Fed cut rates by 25 basis points" is a claim. A whole article contains many claims, and they don't all have to clear the bar for the article to be worth reading. But the headline claim — the thing the dispatch is fundamentally about — has to.

An AI journalist extracts the load-bearing claims from a candidate dispatch, treats each one as a separate verification target, and runs the cross-reference pass on each independently.

What a cross-reference pass actually does

1. Source independence check.The pass looks for the same claim across outlets that don't share an upstream source. Two outlets republishing the same Reuters wire isn't two sources; it's one source twice.
2. Time alignment. The claim has to appear across outlets within a window short enough that all of them are reporting independently, not chasing each other.
3. Numerical consistency. If the claim contains numbers — earnings, casualties, percentages, margins — the numbers across outlets have to agree, or the dispatch flags the disagreement explicitly.
4. Primary-source check, when possible. For company filings, regulator releases, official statements — anything where the primary document is publicly accessible — the pass tries to match the claim to the document directly, not just to the outlets reporting on the document.
5. Contradiction check. If any independent outlet reports a contradicting version of the claim, the dispatch is held until either the contradiction resolves or the contradiction itself becomes the story.

What gets filed and what doesn't

Filed: a claim that clears the bar across multiple independent outlets, with numbers consistent or disagreements transparently noted, traced where possible to a primary source, and not contradicted within the time window.

Not filed: a single-source scoop, even from a reputable outlet, until other independent outlets have confirmed. Most rumors. Most algorithm-driven trending stories whose underlying claim is paraphrased rather than reported. Most social-media-originated "breaking" news whose source chain dead-ends in a screenshot.

The cost: AI journalists are not the first to file most stories. The benefit: they almost never file a wrong one, and the dispatches you get are dispatches you can act on without a second-source check of your own.

Why this is the only honest model for real-time AI news

In real-time news, speed and accuracy trade against each other. The honest framing of any AI news product is which side of the trade it lives on. A pure-rephrase product is fast and unsafe. A human-edit-every-piece product is safe and slow. Cross-referencing is the third option: somewhat slower than rephrase, much faster than full human review, and accurate enough to publish without a verification step by the reader.

Sentinel sits on the cross-reference side of that trade by design. The output is dispatches, not headlines. Every dispatch carries its source-count metadata so you can see the verification depth at a glance.

Where to see it in practice

Open a sample dispatch on the site to see the format — beat label, headline, cross-reference count, source chain. Each dispatch you receive in the app carries the same metadata.

The glossary defines cross-reference and the related terms in their canonical form.