Reporting, Writing, and Which One Is Actually the Job

Ask someone what a journalist does and they say writing. Ask a journalist where their time goes and writing is rarely the largest share.

Reporting. Finding out what happened. Calling people, reading documents, filing records requests, sitting through hearings, checking claims, cultivating sources who will tell you things over years. This is the work, and it is mostly not at a keyboard.

Judgement. Deciding what matters, what is a story, what is true enough to publish, whose account to believe when accounts conflict, and what the audience needs to understand it.

Composition. Turning what you know into prose that a reader will follow: the structure, the opening, the pace.

Production. Headlines, standfirsts, captions, social copy, newsletter versions, search optimisation, the metadata nobody enjoys.

The compression follows the same shape as every profession in this catalogue. Production compresses heavily. Composition compresses substantially, with a caveat about voice that lesson two covers. Judgement does not compress. And reporting barely compresses at all, because it consists of getting information that is not written down anywhere, which is precisely the definition of what a model trained on written material does not have.

That last point is the whole cursus in one line. A model can only work with what has been written. Journalism exists to produce what has not been written yet.

Every profession in this catalogue warns about invented output. Here the stakes are different in kind, and it is worth being precise about why.

In most fields a fabricated detail is an error to be corrected. In published journalism it is the one thing the profession treats as disqualifying. Fabrication ends careers in a way that being wrong does not, because the entire proposition of the work is that what is printed was checked.

And the exposure is not only reputational. A published false statement of fact about an identifiable person can be defamatory, and liability attaches to the publisher and often the writer. A generated quotation attributed to a real person is a false statement about them in their own voice, which is close to the worst available case.

There is now a public record of this going wrong. Newsrooms and content operations have published AI-assisted material containing invented facts, quotes and sources, and the resulting corrections and retractions have been substantial enough that several outlets have issued explicit policies.

What makes it dangerous specifically here. A generated quote reads exactly like a real one. There is no linguistic signal. A quote is the single highest-trust element in a piece, because readers treat quotation marks as a promise that a person said those words.

So one rule sits above everything else in this cursus, and it does not have exceptions. Nothing inside quotation marks comes from a model. Quotes come from a recording, a transcript, or your notes, and if you cannot point to the source, it does not run.

Reporting resists compression, but several of its supporting tasks do not, and they are where the hours quietly go.

Transcription. This is the clearest win in the profession. Interview transcription used to cost roughly the length of the interview again, or a budget line. It is now fast and cheap, and it changes behaviour: you record more, you go back to what was actually said rather than your notes, and you quote more accurately.

Document sets. Reporters increasingly receive large volumes: a records request return, a court filing, a corporate disclosure, a leaked cache. The task is finding what matters in material nobody has indexed. Assisted search over a supplied document set genuinely changes what a small team can review.

Background assembly. Getting up to speed on an unfamiliar area quickly, so your first question to an expert is not a basic one.

Structured extraction. Turning a list of filings, incidents or records into a table you can sort, which is how many data-driven stories begin.

The caution that applies to all of these. They accelerate finding, not verifying. A passage surfaced in a document set still has to be read in context. A background summary is orientation, not a citable fact. And the most dangerous output remains an answer about something no supplied document covers, where the model fills the gap.

The reporting standard does not change: you verify what you publish against a primary source. What changes is how fast you get to the thing worth verifying.

Sorting the elements of a published piece by whether generated text may reach them.

Never. Anything inside quotation marks, because a quote is a promise about what a person said. Any factual assertion not verified against a source. Names, dates, figures, titles and affiliations taken from a model rather than checked. Sources and citations, which are generated fluently and are frequently invented. And any characterisation of what a real person thinks or intends.

With verification. Background and context, where the model drafts and you confirm each claim. Explanatory passages about how something works. Structural suggestions.

Freely, because nothing factual is asserted. Headline and standfirst variants for you to choose between. Social and newsletter versions of a piece you wrote. Alternative structures for material you already have. Editing your own prose for length, clarity and rhythm. Interrogating your own transcripts and documents.

The organising idea is the direction of information flow. Where the model is reshaping material you brought, it cannot invent, because everything it works with came from you. Where it is supplying material, it can, and the output looks identical either way.

So the safe uses are the ones where you supplied the facts, and the dangerous ones are where you asked for them.

A risk specific to this profession, and one where the consequences fall on someone other than the writer.

Journalists hold material whose disclosure can harm people badly: the identity of a confidential source, the existence of an investigation before publication, a whistleblower's account, documents provided on condition, and the fact that a particular person is talking to you at all.

Pasting any of that into a general consumer tool is a disclosure to a third party. Under many consumer terms the content may be retained and may be used to improve the service. It leaves your control, and it becomes something that could in principle be sought by a party who wants to know who talked to you.

The protections journalists rely on, shield laws and source protection principles, were built around notebooks, phones and newsroom systems. They were not designed with a commercial third party holding a copy of your working material.

The practical positions, in order.

For genuinely sensitive material, use tools that run locally or under a contract your organisation has negotiated, with retention and training use addressed explicitly.

Separate the tool from the identity. Transcribing an interview where the speaker is named in the audio is different from analysing a passage with names removed.

And hold a hard category that never goes into any external tool: the identity of confidential sources, and anything that would reveal it by inference.

That last one is easy to breach by accident. A document that does not name a source may still identify them by what only a handful of people could know.

A distinct risk from fabrication, and one writers underestimate because it does not feel like copying.

A model trained on published text can reproduce phrasing, structure and occasionally near-verbatim passages from its training material. In most output this is unremarkable, because common phrasings are common. But on a distinctive turn of phrase, an unusual formulation, or a summary of a specific piece, generated text can land close enough to a source to constitute plagiarism in the way any editor would recognise it.

The writer did not intend to copy and cannot see the source, which makes this harder to defend than ordinary plagiarism, not easier. The usual defence, that you arrived at the phrasing independently, is exactly what happened and is exactly what nobody can verify.

There is a second version specific to reporting. Ask a model to summarise what is known about a topic and it will produce, in effect, a synthesis of other outlets' reporting, without attribution. Publishing that is taking other newsrooms' work, which is both an ethical problem and, where a piece follows another outlet's original reporting closely, a commercial and legal one.

The practical protections. Run generated passages through the plagiarism checking your organisation already uses; the tooling predates this problem and works on it. Do not use models to summarise coverage as a substitute for reporting. And where your piece rests on another outlet's original work, credit them, which was always the standard.

The underlying principle is unchanged: attribution is owed for reporting you did not do, and the fact that a tool laundered the provenance does not remove the obligation.

Worth stating explicitly, because the editorial function is what a lot of AI writing tooling implicitly claims to replace.

An editor does several distinct jobs that get bundled under one title.

Line editing. Making the prose better sentence by sentence. This is the part a model does genuinely well, and writers who use it as a line editor on their own drafts generally improve.

Structural editing. Seeing that the piece begins in the wrong place, that the argument does not hold, that the fourth section is the actual story. Partly assistable, because structural suggestions are useful even when wrong, and they prompt a rethink.

Fact checking. Verifying claims against sources. Not assistable in the sense that matters, because the checking is the point and a model checking a model is not verification.

Judgement about publication. Is this fair to the people in it? Is it strong enough to run? What is the legal exposure? Have we given the subject a proper chance to respond? Is the framing right?

And protection. A good editor protects the writer from their own enthusiasm, and the publication from both.

That last cluster is not a language task at all. It is institutional judgement about risk, fairness and standing, exercised by someone accountable for the outcome.

So the honest reading. Tooling substitutes reasonably for line editing, assists with structure, and does not touch the part where an editor decides whether a piece should exist in this form. In an operation that has cut editors, the tool covers the layer that was least important and leaves the gap where it was most.

An order that follows exposure rather than visibility.

First, transcription. The largest unambiguous win in the profession, it operates on your own recording, and it improves accuracy rather than threatening it because you quote from what was said rather than from your notes.

Second, interrogating your own material. Your transcripts, your document set, your notes. The model can only work with what you supplied, which is the property that makes it safe.

Third, production work. Headlines, standfirsts, social copy, newsletter versions of a piece you already wrote and verified. Nothing factual is being asserted that was not already in the piece.

Fourth, line editing your own drafts. Real improvement, and the input is your prose.

Fifth, with verification: background and explanatory passages, where every claim gets checked against a source before it runs.

And not at all: quotes, any fact you have not verified, citations, characterisations of what people think, and anything about a confidential source going into an external tool.

The pattern is the one from the diagram. Uses where you supply the material are safe and are most of the value. Uses where you ask the model to supply material carry the profession's one disqualifying risk.

Most writers start at the fifth item because it looks like the biggest saving. It is the one that requires the checking discipline to already exist.