What newsroom AI policies actually say in 2026

Over the past nine months we collected every publicly available newsroom AI policy we could find — 27 in total, spanning seven languages and four continents. We were looking for a baseline: what do editors actually commit their newsrooms to, in writing, when AI enters the building?

The first surprise was tonal. Most policies are not prohibitions. They are workflow descriptions. They explain what tools may be used, by whom, at which stage of a story, and what disclosures attach to the output. Hard "no" lines exist — synthetic imagery of real people is the most common — but they sit inside a permissive frame.

The second surprise was convergence. Independent newsrooms, working without coordination, landed on a similar set of commitments: a named human editor on every story; disclosure when AI use is material to the published product; a prohibition on AI-generated quotes; and a workflow note that draft summarisation is permitted but draft writing is not.

The third surprise was what is missing. Very few policies say anything about training data, licensing, or compensation. Almost none address the question of which models are acceptable on which devices. Almost none mention the labour implications inside the newsroom itself.

We are publishing the full dataset, with a one-page comparison matrix, under CC BY-SA on the OER page. If your newsroom has a policy we missed, write to us — we will fold it into the next revision.