How NoDirection works
We find the facts. You decide.
This page explains how our AI agents work, the main rules they follow and their limits. Stories are published automatically after AI review, without routine human review before publication.
1. We compare a broad range of sources
We monitor almost 100 news outlets across the political spectrum (newspapers, broadcasters, agencies, digital and magazine titles) and around 20 official sources, mostly UK government departments, and group reports about the same event. Events covered by five or more outlets get a full story; smaller ones get a short brief, which is rewritten as a full story if more outlets pick it up.
For each story we use the most recent article from each outlet, up to 10 outlets, with official sources first. Where we cannot read an article’s full text, we work from its headline and summary. See every source we monitor.
2. An AI writer drafts the story
Our main writer is an AI agent that also searches the web for primary sources: official statements, court and government documents, full quotes, data and comparable past cases. If it does not return a story by its deadline, a fallback AI writer takes over. The fallback does its own web research for full stories, but not for short briefs. Both follow the same written rules, the core of which is: use only the sources, and never present an unverified claim as an established fact. Both go through the same review before anything is published.
Key factual points in each story are labelled:
- Confirmed: directly supported by an official record, a primary document, or reports from two or more outlets that are not repeating the same original source. A label applies only to what the evidence establishes: an official statement or document can confirm that a claim was made without confirming that the claim is true.
- Reported: reported as fact by one outlet, or by several drawing on one origin.
- Claimed: asserted by a named person or group.
- Disputed: contradicted by another source. We say who disputes it and what each side relies on.
- Speculation: predictions, conjecture and possibilities that the evidence has not established.
- Opinion: value judgements, by anyone.
Two important safeguards guide this. Reports drawing on the same original source, such as one news agency report or one press release, count as one line of evidence, even when the wording differs. And when an official announces something, we can confirm that they said it; the substance stays a claim unless the evidence establishes it.
3. A separate reviewer checks it
A second AI agent, which did not write the story, is instructed to check the story’s statements, numbers, names, dates, quotes and context against the sources. It sends the story back if a statement is not supported by its sources, a claim is presented as fact, a detail does not match, the wording is loaded or adopts one side’s framing, or context that changes a quote’s meaning is missing.
The writer gets up to two rounds to fix the problems (one for short briefs written by the fallback writer). If the story still fails, it is not published.
We also apply extra restrictions to stories about criminal proceedings and serious allegations. The reviewer holds a story back if it could imply guilt of someone charged but not convicted, adds details about a live UK criminal case beyond what police and courts have said, repeats serious allegations against a named person from a single outlet with no response from them, or if the sources are too thin to say what happened. These stories are checked by our more capable model. These are safeguards, not a guarantee that AI will spot every legal issue.
The reviewer also rates how much each story adds compared with reading one outlet’s article. Stories that mostly restate the headlines, and short briefs, stay on the site but are kept out of search engines.
4. We assess spin
Once a story passes review, a third AI agent, the spin analyst, gives each outlet’s article a score from 0 (straight factual reporting) to 100 (heavily slanted). It compares the article’s headline and text with the confirmed facts and with how the other outlets covered the same event. The score summarises how the article frames the story. It is not a measure of how much of the article is false.
It looks for nine techniques: loaded language, leaving out key facts, speculation presented as fact, opinion presented as fact, emotional appeals, selective quotes, emphasis, clickbait headlines and reliance on unnamed sources. For each one it finds, it must quote the exact words (or, for an omission, name what was left out) and explain why that is framing. We show this on every story.
The analyst is told to:
- score by technique, never by which side an article helps;
- judge the article, not the outlet’s reputation, and be equally strict with every outlet;
- remember that a one-sided story that reports accurately can still score low, and a neutral-sounding story that leaves out key facts scores higher;
- judge only what it can read when we have just an article’s headline and summary, and keep those scores modest.
It also notes the political direction of that article’s framing (or “not political”), how sensational it is, and whether the headline matches the story.
On the site, 0 to 29 is low spin, 30 to 59 is some spin and 60 or more is high spin.
Spin scores are AI judgements. They can be wrong, and the same article could score somewhat differently if assessed again. We have not yet tested how consistent the scores are or compared them with human assessors; when we do, we will publish the results here.
5. The weekly spin league
The weekly league averages each outlet’s scored articles in stories we published or updated during the previous seven days. An outlet appears only when we have scored at least five of its articles in that period. It describes the articles we analysed on the events we covered, not everything the outlet published. Scores for articles we could only read in part (headline and summary) are currently included. Outlet profile pages show a separate 30-day average.
6. Who checks the spin scores
Spin scores appear as soon as a story is published. After that, a second, stronger AI model independently re-scores up to 20 stories a day: the 10 most heavily spun and 10 chosen at random. If the two scores for an article differ by more than 15 points, the article is marked as under review on its story and left out of the spin league until they agree. A human reviewer also checks scores at random on the site after they are published. We publish the excerpts and reasoning behind each score so that anyone can assess and challenge it. Outlets and readers can challenge a score by emailing [email protected]. A human reviewer looks at every challenge, and if a score is wrong we correct it.
7. Corrections and updates
Anyone can report an error with the button on each story. An AI reviewer checks the report against the sources. If it is upheld, the story is rewritten, reviewed again and the correction is noted on the story and on our corrections page. If the corrected version does not pass review, we take the story down until it is fixed. If you disagree with the outcome of a report, email [email protected] and a human reviewer will look at it. When new facts come out, we update the story and keep a record of every version.
8. Images
When we show a photo an outlet used, it is to comment on how that outlet used it to frame the story. We credit the outlet and link to the original. If we cannot say anything useful about how it was used, we do not show it. We never generate images of real people.
9. What AI can get wrong
Our agents run on OpenAI models. They can misread a source, miss context, give confident wrong answers or carry biases from their training. We cannot remove that risk, so we design around it: key points link to their sources, a separate agent reviews every story, spin scores show their reasoning, and errors are corrected in public. If you find a mistake, please report it.