Prepared with AI assistance from the linked primary sources. Company and government claims are attributed to their issuers. How we use AI.

An evidence-first AI newsroom begins with a simple obligation: readers should be able to distinguish what happened, who says it happened and what the publication concludes from it. AI coverage makes that obligation especially important because commercial announcements often combine present facts, future plans and ambitious expectations in the same document.

Authorship must be equally clear. When an article is written with AI, that role should be disclosed plainly. A publication should not imply interviews, site visits, independent testing or human editorial approval that did not occur. Disclosure does not replace verification; it tells readers how to interpret the work while the linked evidence supports its factual claims.

Build claims from the source that owns them

A company release can establish what the company announced. A government document can establish a policy’s published terms or procedural stage. A model card can describe the developer’s evaluations and declared limitations. Each source has a useful scope, and none automatically proves every broader conclusion a writer might draw.

The Associated Press’s published 2023 guidance treats generative AI output as unvetted source material and places responsibility for accuracy on journalists. AP also restricts AI-generated publishable content. An AI-authored publication should not suggest that it follows that entire policy; the relevant lesson here is to avoid treating generated text as evidence. AP’s generative AI standards.

NIST’s Generative AI Profile describes confabulation among the technology’s risks. Fluent prose, detailed numbers and plausible-looking references therefore require checking against the underlying material. A citation should be opened and read for the specific claim it is being asked to support. NIST’s Generative AI Profile.

Preserve the distinctions that headlines erase

Dates are a frequent source of confusion. Ubitus’s 9 September 2026 release described a groundbreaking held on 4 September. Announcement date and event date both belong in a precise account. Ubitus’s Maizuru release.

Status language deserves the same care. Planned capacity is not installed capacity. A financing agreement is not necessarily completed financing. A product launch does not establish adoption or revenue. These are ordinary reporting distinctions, but losing them can substantially change a reader’s understanding of an AI project.

Analysis should then explain the significance without disguising inference as reported fact. A writer can argue that an infrastructure programme may improve access, while making clear that the source does not yet establish utilisation or economic impact. Hypothetical examples should be labelled so that they cannot be mistaken for customer case studies.

Make the article easy to challenge

Place source links beside the claims they support. State important limitations where they affect interpretation. Preserve enough detail about dates and measurement definitions for another reader to reproduce the comparison. When an error is found, a visible correction should explain what changed.

The result need not be cumbersome. A concise article with a precise claim, a readable explanation and an inspectable source can be more useful than a long story built around unsupported certainty. Evidence-first publishing is ultimately a design choice about trust: it gives readers the means to assess the work themselves.

Sources

See something we should correct? Contact the newsroom.