Editorial policy

One article a week, covering the Monday to Sunday that has just finished, published on the Monday evening after it, UK time. It is written for people putting AI to work in the UK: what happened, what it means for that work, and a link to the source behind each claim.

It is not a wire service and it does not try to cover everything. The brief asks for the three or four things that change what a practitioner does, with the most important one given the most room.

The draft is written by a large language model, Anthropic’s Claude, working from a fixed brief that lives in this site’s repository. The brief covers the structure, the sourcing rules, the words it may not use and how the piece should read. The model is told never to write from memory: every date, figure, version number and price is looked up with a live web search before it goes in.

A second, separate request then fact-checks the draft as somebody else’s work, with web search again. It walks a named list of failure modes: the right number attached to the wrong event, a percentage whose definition has drifted, a stale figure presented as current, two parties’ numbers merged into one sentence, a “first ever” that is not, a benchmark that is no longer a valid measure. Claims it cannot trace to a source are removed rather than softened.

Nobody at AISAT writes the prose, and the model is not asked to invent experience. First person is allowed for a judgement or an opinion. Claimed experiences, conversations and tests are not.

Before anything is published, the finished article is measured against the same contract the brief describes. Among other things it has to be:

• between 557 and 3500 words of body text;
• supported by at least 3 sources, of which at least 2 are primary: the announcement, the filing, the paper or the regulator’s own page rather than a write-up of it;
• carrying 5 to 8 questions answered at the end, in plain text;
• structured properly: headings in order, tables whose rows match their columns, a definition of the key idea near the top, and a headline short enough to survive a search result.

An article that fails any of those is saved as a draft and is not published. So is one where the fact-checking pass did not run at all. An organiser then reads the failures and decides what to do with it.

A roundup that passes those checks publishes on its own, which means it is on the site before a person has read it. Rather than hide that, every article states it. Until an organiser has been through it, the article carries a notice reading “AI-generated. Not yet reviewed by an editor”, and the card for it on /news says the same.

When an organiser has read it, the notice is replaced by their name: “Drafted with AI, edited by” them. That is a person saying they have checked it, so it is only ever set by one of them pressing the button on the article itself, never by the pipeline.

Every article lists its sources at the end, with the publisher and the date of each one, and marks which are primary. A primary source is the thing itself; anything else is reporting about the thing, and the list says which is which so you can judge a claim by what stands behind it.

Source links open in a new tab and are marked so that they pass no search ranking on from us: on an article no person has read yet, a link is a reference and not an endorsement. A source address that does not pass the site’s own URL checks is shown as plain text instead of as a link, and says so.

Corrections are made by editing the article in place, at the URL it was published at. We do not quietly replace a figure and move on: the byline carries the date it was last updated whenever that is a different day from the one it was published on.

If you find something wrong, email hello@aisat.uk with the link and what it should say. Corrections to facts are welcome from anybody, member or not.

Some roundups have a spoken version, which an organiser generates by hand. It is read by a synthetic voice, and the player says so. The article on the page is the transcript: nothing is said in the recording that is not written above it.

AISAT publishes the roundup. AISAT was founded in May 2026 by John Bowman, its Founder and AI and Automation Lead, and is run by the volunteer organisers on the team page. The roundups are free to read, need no account, and are listed in full at /news.