AI Broadsheet — an Open Desk newsroom
We don't ask you to trust us. We show you the record.
Open Desk Journalism is a way of reporting in which any reader can follow who made an important claim, what evidence supports or contradicts it, what a human editor checked, and why the information has changed since.
Why we do this
When a company says it is cutting jobs “because of AI”, the line travels through the news in hours. Almost nobody checks later whether it was true: whether the work was really automated, moved somewhere else, or whether AI was a convenient explanation. And almost nobody comes back six months on.
Those announcements steer real decisions: what people study, where they apply, what governments regulate. We want them kept on the record, tested, and followed up.
Our promise: we follow not only what was said, but the evidence behind it and what happened next.
What we aim for
- 01
A public, permanent record of what powerful actors in AI say about work, kept before a claim goes viral, not after.
- 02
Readers who can tell at a glance whether a claim is only recorded, still without a conclusion, supported, contradicted by other evidence or refuted.
- 03
AI that makes a small desk fast, with people who stay accountable for every decision.
- 04
Creators and witnesses as partners: credited, paid, and tied to a record that can be checked.
How a claim moves through the desk
Seven steps, from the statement to the signed result, then the reopening. Go at your own pace.
The desk, step by step
Fictional exampleAI
Capture
A press release says AI is now doing 1,200 people's work. The desk saves it with an archived copy and a time stamp, and opens record OD-9999.
Tip: the ← → keys move between steps.
OD-9999Recorded, not yet tested
“AI is now doing this work. We are reducing our dispatch and back-office teams by 1,200 roles.”
History of the result
- Recorded, not yet tested15 Sep · recorded from the press release
AI brings speed. People decide.
Everyone works on the same record, and no one can rule outside their role.
AI: speed
Watches public statements, opens records, splits claims, matches earlier words and documents, flags contradictions. Never writes a result.
Reporters: evidence
Check documents, talk to former staff and experts, send the company its questions and wait for the answer.
The editor: the decision
The only person who writes a result, with the reason, signed by name. Every verdict is read first by a second person.
Creators: reach
Share signed records in their own voice, show the record number and update their post when the record changes. Witnesses are asked for consent, credited and paid.
Five results, not a score
No truth percentage. A result can move either way as new evidence arrives. Under the stamp, two small lines say where the review and the right of reply stand.
- Recorded, not yet testedWe recorded who said it, where and when. This is not a verdict: we have not tested the claim yet.
- No conclusion yetWe have looked, and the evidence so far is not enough for a conclusion, or the documents that will settle it are not out yet.
- Supported by the evidenceIndependent evidence supports the claim, within the scope and on the date stated. A match with the company's own material is not enough.
- Conflicting evidenceRelevant, reliable evidence points in different directions. This is not a finding that the claim is false.
- Refuted by the evidenceIndependent evidence shows the claim, as recorded, is false. It says nothing about intent.
Six commitments
- 1
We separate attribution from verification
Quoting a company correctly is not the same as accepting what it says. A company's announcement shows that it said something; on its own it never counts as independent support.
- 2
Every important claim gets a traceable record
Claims that could change what a reader understands or decides get a permanent number, the source passage, an archived copy and the evidence for and against. We also say what the evidence does not show.
- 3
A person is responsible for what we publish
Choosing the story, judging the evidence, giving the right of reply, approving and correcting stay with a named editor. AI is a tool; its output is never proof of its own accuracy.
- 4
Change stays visible
A new development, a correction, a clarification and a retraction are named as what they are. Every change of result keeps its date, reason and signature, and the earlier result stays on the page.
- 5
We listen to readers and protect sources
Readers can suggest claims, bring documents and challenge a record. Accuracy is not a vote: the editor weighs the evidence and explains the decision. Whistleblowers, children and people under pressure can stay anonymous, for stated reasons.
- 6
Evidence does not bend to tone or funders
Hope, solutions and harm appear as far as the evidence shows them, no further. No sponsor or supporter can choose a story, set a result, see it before publication or stop it, and we take no money from companies on the ledger.
Our targets for the first season
These are goals, not results. We will publish what we reach, and what we miss.
50
statements screened in the first 10 weeks; we publish how many became records, and why
10
active records at most at any time, so each one gets real testing
5
business days of right of reply before any verdict
80%
of readers in our tests tell “recorded” from “supported by the evidence”
Our first beat: AI and work. Others (the economy, rights) follow once this desk has proved itself.
What we don't claim
- We did not invent citing sources, public corrections or human responsibility. What we add is a test written before the result, a verdict that waits for the right of reply, and return dates set when the record opens.
- A result is not a truth score for a company: it is about one claim, at one date. A statement that is not in the ledger has not been assessed.
- A method is no guarantee against error. When we get it wrong, the correction is dated and public.