The story looks fine. The grammar holds, the sentences are tidy, and the headline says something you half-remember hearing on the radio. Then you reach the bottom and realise you cannot name the person who wrote it, the people they spoke to, or the document behind the first number in the opening line.
That is the shape of a quiet change in UK news. Software can now turn out passable copy faster than any newsroom can, and plenty of websites are filling their pages with it, sometimes with a label and more often without one. Add the local titles that have quietly shrunk their reporting staff, and the hybrid cases — a human byline, a machine first draft — and the question stops being theoretical. The reassuring part is that machine-written news leaves fingerprints. You do not need special tools to find them. You need ninety seconds and a mildly suspicious streak.
The prose is rarely the giveaway
Bad spelling and clumsy grammar are the wrong things to look for. Text-generating tools produce clean, confident English and rarely fumble a comma. If anything, the warning sign runs the other way: pages of smooth, frictionless prose that tell you nothing you could only learn by standing somewhere.
A reporter who sat through a meeting gives you the details that stick, such as the name of the hall, the vote that went the wrong way, the councillor who sighed. Those details are not decoration. They are the proof that somebody was in the room, and they are what makes the rest of the piece checkable. Copy assembled without a source gives you the shape of the meeting and none of its texture.
So ask a blunt question of every paragraph: what is in here that could only have come from a phone call, a document somebody had to go and fetch, or a room somebody had to sit in? Where the answer is nothing, keep your guard up.
Residents have expressed mixed feelings about the proposals. A local spokesperson said the council was committed to reviewing the matter thoroughly. The decision will be taken in due course.
Nothing in that is false. Nothing in it is much use either. Real quotes have edges — a half-finished sentence, an "er", a stray remark about the weather or the traffic. Polished, perfectly balanced sentences attributed to someone who is never given a name, a job title or a way of being contacted are the single most reliable signal in this whole business.
The sourcing is where it usually falls apart
Turn to the attributions. Human journalism, at its best, names people: an official, a document, a date, a figure. Copy produced without a source leans on ghosts instead.
- "Experts say" and "industry observers believe", with no expert and no industry named.
- "According to reports", where no report is identified, dated or findable.
- Quotes without edges. Real speech is uneven. Polished, perfectly balanced sentences attributed to someone who is never given a job title deserve a second look.
- No primary sources at all. Planning applications, council papers, court lists, company filings and regulator notices are the raw material of local news. A story about any of those that refers to none of them is running on air.
- Confident vagueness about time. "Recently", "earlier this year" and "in recent weeks" standing in for an actual date.
- Nobody named at all. A seven-hundred-word piece about a planning row, a fatal accident or a business collapse in which not one human being has a name is not reporting. It is a summary of reporting.
Six checks you can run in about a minute
- Search a distinctive phrase. Take an odd clause from the middle of the piece, not the headline, and search it in quotation marks. The same wording turning up across unrelated sites tells you what you need to know.
- Trace the most specific claim. Pick the one number, date or decision that would matter if it were wrong, then look for where it came from. If the only page carrying that figure is a page quoting the same figure without a source, you have found a loop rather than a fact.
- Click the byline. Real newsrooms have author pages, previous stories and a contact route. A byline that leads to a blank profile, a stock photograph or nothing at all is a byline doing decorative work.
- Look at the bottom of the page. An "About us" page with no masthead, no address and no named editor is not usually a newsroom. It is a domain.
- Reverse-search one image. Pictures lifted from press libraries or stock sites, or reused under a caption that does not quite match, are everywhere on machine-filled pages.
- Read the last paragraph. Human stories tend to end on a quote, a scene or a stated next step. Generated ones tend to end on a summary and a flourish about the importance of community engagement or the need for balance.
What the label means — and what it does not
Some UK publishers now disclose, and you will increasingly see lines such as "This article was written with the assistance of AI tools and reviewed by an editor". Read that as a statement about the process, not as a guarantee of quality. There is a large difference between a reporter who used a machine to tidy a transcript and a site that generated forty stories from one council press release overnight, and the same short label can cover both.
Three rough grades of disclosure
- Assisted. A human did the reporting; a tool helped with transcription, subbing or search. Usually harmless, and often invisible to the reader, which is the point.
- Generated and edited. A machine produced a draft from real source material and a person checked it. The quality depends entirely on how much checking happened, and on whether the underlying material was any good to begin with.
- Generated at scale. No reporting, no named source, no meaningful review. This is the category that fills search results with pages about things that never quite happened.
Absence of a label tells you nothing at all, which is the awkward part. Publishers who would never dream of running unchecked machine copy often say so; the ones pumping it out rarely do. So the label is useful when you see it and irrelevant when you do not. The fingerprints are still the thing to go on.
When the clues point somewhere else
Be fair before you accuse. Press releases read like press releases, and plenty are published almost untouched under a human byline. Syndicated copy repeats itself across dozens of titles because it was sold to dozens of titles. Writers working in a second language, translated copy, legal and academic prose, and the work of somebody having a very bad afternoon can all produce flat, source-light writing.
None of that is machine writing, and treating it as such is how honest journalists get caught in the crossfire of a moral panic. The distinction that survives all these exceptions is not about style at all. It is about traceability: whether there is a person, a document, a date or a phone number behind the claim. If there is, you can check the story, argue with it or complain about it, and either way somebody is answerable. If there is not, there is nothing to argue with.
The ninety-second version
Ignore the grammar. Ask who is speaking, where the number came from and when it happened, then try to find those answers somewhere other than the page you are reading. If you can, you are reading journalism, whoever or whatever typed it. If you cannot, you are reading filler — and it is worth saying so out loud, because the only thing keeping the good stuff alive is readers who notice the difference and reward it with their attention.
Photo: viarami / Pixabay
