Transparency
How a Daily Pol piece gets made.
Daily Pol uses AI assistance in drafting, checking and - for standard-tier pieces - release. Reporting that names a person or makes an allegation is released only by a named human, who is accountable for everything published. This page describes the process exactly, including its limits, and every piece states which release path it took.
The Record is now publishing, and its pieces are AI-assisted as described below. They are marked “The Record” and carry a desk byline - see who writes Daily Pol for what that byline means. Everything else readable here is the recovered 2021–2022 archive, written by the site’s previous operator and republished unedited with its provenance.
The pipeline
A piece moves through four stages. The first three are automated. The fourth - release - split into two paths on 7 September 2026: a piece that makes a negative claim about a named person, or alleges a crime, still cannot publish without the editor of record's personal approval, while a standard-tier piece may be released by a second adversarial machine review, described below. Both paths are gates in the code rather than intentions, and which path released a piece is recorded on the piece itself - so this description is checkable against any piece released through the gate; pieces from before it - including six published earlier on 4 September 2026 itself - and the recovered archive carry no release record. From the gate's arrival on 4 September until 7 September 2026 every release required a human; see about for what happened before that.
1. Research
An AI research stage searches and reads primary sources - government documents, filings, datasets, transcripts, official statements - and builds an evidence base. Every page it reads is saved verbatim at that moment, so the evidence a claim rests on is the exact text the system saw, not a later version of the page.
2. Drafting
A separate AI stage writes the piece using only that evidence base. It produces the article together with a claims ledger: a row for every factual assertion, naming the source, the exact supporting passage, and a confidence level. A claim with no ledger row does not exist, and cannot appear in the article.
3. Adversarial checking
A third stage is instructed to refute the draft, not confirm it. It re-reads every saved source and attacks each claim: does the quotation appear verbatim; does the source in context mean what the claim uses it for; do the numbers match digit for digit; is anything asserted that the ledger does not cover; is an allegation being stated as fact.
An automated gate then runs mechanically before any person sees the draft. It verifies that every quotation appears in the saved source text, that crime allegations carry a primary court or police document, that the headline breaks none of the standards, that banned vocabulary appears nowhere outside quotation marks, and that the piece is not too close a paraphrase of any source.
4. Release - two paths, recorded on the piece
Standard-tier pieces - no negative claims about named people, no crime allegations - may be released by a second machine review, independent of everything that drafted the piece. Two separate model passes run: a reviewer that must dispose of every one of the checker's objections against the captured sources and repair what it can, and a refuter whose only instruction is to find a reason the piece must not publish, with uncertainty counting as no. Any objection left standing, any repair that does not apply exactly as written, any disagreement between the two passes, or any failure in the final automated gate holds the piece for a human - the machine cannot override any gate a human can. A machine-released piece says so on the page, in so many words: no human read it before publication. That sentence is the cost of publishing at machine cadence, and we would rather print it than soften it.
High-risk pieces - negative claims about named people, allegations of crime - are released only by a named editor, who reviews the draft against a written checklist. The checklist deliberately contains only judgement the machine cannot do: opening the saved sources and reading the quotations in context to catch a passage used against its meaning; swapping the party affiliations of everyone named to test whether the piece would have been written the same way; confirming that what the draft claims is new actually is. These pieces additionally require a documented request for comment before publication. An editor can also pull or release any piece on either path; the machine can only release the standard tier, and only through its own gates.
What we will and will not do
Every published piece carries a desk byline and a disclosure that AI assistance was used in drafting, and a named human - the editor of record - is responsible for its contents. We do not present a desk as a person: the bylines are listed on who writes Daily Pol with a plain statement of what they are, and no biography, photograph or career history is claimed for any of them. We do not publish unverified generated claims as reporting, and no sentence reaches review without a ledger row behind it, regardless of what drafted it.
Where this can fail
Automated checking proves a quotation is present in a source. It cannot prove the source is right, that the framing is fair, or that something important is missing. Those are human judgements, and human judgement is fallible - which is why the corrections page exists and why corrections are published rather than quietly edited in.
If you find an error, tell us. Correction requests are answered before anything else.