AI and automation

There is a version of a motoring site where a machine writes the copy, nobody reads it, and the traffic holds up for about a year. This is not that, and this page exists so you do not have to take our word for it.

The short version

InAutoNews is written and edited in house. Reporters decide what is worth covering, read the material, write the pieces and sign them. We use software, some of it built on language models, for the parts of the job that are fetching, searching, sorting and checking. It is a research and production tool. It is not the author, and it has no route to a published page.

Where any of it touched the text of an article, the article says so. If you find something on this site that reads like it was written by a machine and not marked as such, that is a failure of ours and we want to hear about it.

What people do

Deciding what is a story. Judging whether a source is any good. Reading the filing. Calling the press office. Choosing the angle, writing the sentences, cutting the two thirds that turned out not to matter. Deciding a piece is not ready. Every one of those is a person's job here, and none of it is delegated.

A second person reads every article before it is published. On anything where a tool contributed to the text, that person is named at the foot of the article by name and role.

What the software does

The useful, unglamorous half of the work. In practice that means:

• Watching. Monitoring manufacturer newsrooms, regulators and filings so an editor sees a release at eight rather than at noon, and grouping the twenty versions of the same announcement into one item to look at.
• Fetching. Retrieving the documents an editor has asked for and keeping a copy, so what an article was written from is still there a year later when somebody queries it.
• Summarising for us, not for you. Reducing a sixty page technical release to the paragraphs a reporter needs to read properly. The summary is a reading aid inside the office. It is not the article.
• Checking. Testing each factual statement in a finished draft against the documents it was written from, and flagging anything the documents do not support so an editor has to deal with it before publication.
• Housekeeping. Tagging, categorising, sizing pictures, spotting broken links, matching a restored article to the historical address it belongs at.

On routine, high volume items with no interpretation in them, a tool may assemble a first draft from sources an editor supplied. That draft is a starting point and is treated as one. It is rewritten and edited by a person before publication, and the published piece says machine assistance was involved.

What it is never allowed to do

• Write a quotation. Not a paraphrase in quotation marks, not a reconstruction of what somebody probably said. Nothing on this site is attributed to a person who did not say it.
• Score a car. A verdict is a claim about having driven something.
• Invent a source, a figure, a price or a date. Every one of those has to come from a document that exists and was retrieved.
• Find its own sources. Sources come from an editor. Models asked to supply citations produce ones that look convincing and frequently do not exist, so we never ask.
• Publish. There is no path from a model's output to a live page that does not pass through a named person.
• Impersonate a person. No synthetic byline, no invented reporter, no made up expert.

When we tell you

Any article where a tool contributed to the published text carries a note saying so and names the editor who approved it. Research assistance that never reaches the page is not disclosed on the article, for the same reason we do not tell you which search engine a reporter used, though it is described here.

Illustrations produced by a generator are labelled on the picture itself, in the caption, on every page they appear.

Who is accountable

A person, always, by name. An article produced with machine assistance cannot be published without an approver recorded against it, and that is a rule the database refuses to break rather than a habit the newsroom is asked to keep. It cannot be skipped from the admin screens because it is not a checkbox there.

If a piece is wrong, the answer is not that a tool got it wrong. The editor named on it read it and let it go out, and that is the whole point of naming them.

Why it cannot work from memory

A language model asked about a car it has read about will answer confidently from recall, and the answer will be right often enough to be dangerous. Ours is never asked. Documents are supplied and retrieved first, and a request with nothing to work from is refused rather than answered.

After a draft exists, each claim in it is checked back against those documents. Anything unsupported is flagged for an editor and blocks the article until it is resolved. That check is the reason this arrangement is safer than a reporter working alone at speed, and it is the part we would keep if we had to give up everything else.

Photographs and illustrations

Where an article carries a photograph, it is a real photograph, used under a licence that permits it, credited to the photographer beneath the picture and linked to the original file. We do not generate images of real cars, real people or real events, and we do not retouch news photographs.

Where no usable photograph exists, an article may instead carry a drawn illustration. It is labelled as an illustration wherever it appears. It carries no badges, no registration plates and no identifiable people, and it is constrained so that it cannot be mistaken for a record of something that happened. An illustration on a piece about a 2013 crash is decoration. It is not evidence, and it is never captioned as though it were.

The rebuilt archive

Articles at addresses this publication used between 2008 and 2018 have been newly written for the relaunch. They are not the original text, they say so on the page, and they carry the date the story originally ran. Buying the domain did not buy the copyright in the original articles, and nothing here reproduces them.

Those pieces are researched from archived copies of the original pages and from contemporary reporting elsewhere, and they list what they were written from. Where the record is thin, the article says the record is thin.

What we send to model providers

The tools described here run against commercial model APIs, currently OpenAI's. What goes to them is editorial working material: press releases, filings, published articles, our own notes and drafts.

What does not go to them, at all: messages sent to us through the contact form, mailing list addresses, anything a reader has told us in confidence, and any of the traffic data described on the privacy page. Reader data and editorial tooling are kept apart deliberately, and no reader's personal information is used to train anything.

AI crawlers reading this site

We do not currently block crawlers that gather text for AI systems, and our robots.txt is the authoritative record of what is and is not welcome at any given moment. There is also an llms.txt describing what this publication holds, because a system summarising us should at least summarise us accurately.

Reading the site is not the same as reusing it. What may be quoted, and what needs asking about first, is set out on the terms page. If our position on crawlers changes, it changes here and in robots.txt on the same day.

Reviewing this policy

This page is reviewed at least twice a year and whenever the way we work changes. The date at the foot of it is the date it was last looked at rather than the date it was last edited, so an unchanged date means somebody checked and found nothing to change.

Disclosure costs us nothing we want to keep. If you think something here is wrong, or that the page and the site have come apart, tell us. That is the part of this that actually matters.