Assisted writing at the desk

1. What “assisted” actually covers

Assistance is a spectrum, not a switch. At one end sit tools that have been in newsrooms for decades: spelling dictionaries, style linters, automatic transcription, headline length counters. In the middle sit summarisers that compress a long official document into bullet points, translation engines that produce a first pass for a subeditor, and search tools that surface earlier coverage. At the far end sit generative systems that produce whole paragraphs from a prompt. Treating these as one category makes newsroom policy impossible to write, because the verification burden differs enormously. A transcript still has an audio original to check against; a generated paragraph may have no source at all. Any workable desk rule starts by naming the tool class and the artefact it produces.

2. Where the tools sit in a copy flow

Map the copy flow before you map the tools. A typical online item passes through commissioning, reporting, drafting, subediting, headline and standfirst writing, picture selection, legal read where needed, publication, and post-publication amendment. Machine assistance behaves differently at each of these points. Used at the reporting stage it can widen the material a reporter reviews, which is mostly additive. Used at the drafting stage it introduces text that nobody has yet verified. Used at the subediting stage it can silently smooth a hedge into a flat assertion — “may have” becomes “has” — which is the single most common way an assisted edit changes the meaning of a sentence without anyone noticing.

3. Prompting is an editorial act

A prompt encodes assumptions: framing, tone, what counts as background, which side of a dispute is treated as the default. Ask a system to “explain why the scheme failed” and it will supply reasons, whether or not failure is established. Because prompts are usually typed and forgotten, this editorial decision escapes the record that the rest of the desk keeps. Newsrooms that take assistance seriously store prompts alongside drafts, in the same way that they store interview notes. It also helps to write prompts that ask for structure rather than conclusions: request a list of the claims made in a document and the page on which each appears, rather than a summary of what the document proves.

4. The failure modes that matter in news

Three failures cause most published errors. First, invented specifics: a plausible-sounding statistic, a case citation, a section number in an act, a date that does not exist in any source. Second, misattribution: a real quotation moved to the wrong speaker, or a paraphrase presented inside quotation marks. Third, confident compression: a nuanced official finding reduced to a headline claim that the document does not support. All three are hard to spot precisely because the output reads like competent copy. Errors that look like errors get caught; errors that look like journalism do not. This is why assisted drafts need claim-level checking rather than a general read-through.

A practical countermeasure is to strip the draft back to its factual skeleton before subediting. List every number, name, date, place, title and direct quotation, then locate each one in a primary record. If a claim cannot be traced, it is removed rather than softened. Softening — turning an unverifiable figure into “around” or “reportedly” — preserves the fabrication with a hedge attached.

5. Attribution and the paper trail

Assisted text tends to arrive without provenance. A human reporter can say where a sentence came from; a generated paragraph frequently cannot, and any citation it offers must be treated as a hypothesis to test rather than a reference to quote. Desks that manage this well keep a per-item note recording which stages used a tool, what the tool produced, and who verified the output. The note is internal, short and factual. Its value appears at correction time: when a reader disputes a figure, the desk can establish in minutes whether the figure came from an official release, a phone call or a generated draft, and can respond precisely instead of defensively.

6. Style, voice and quiet homogenisation

Generative systems converge on a median register: balanced sentence lengths, measured transitions, an even temperature. Applied across a whole desk, that convergence erases the differences that let readers tell one publication from another, and it flattens the voice of individual reporters who have spent years developing one. It also has a subtler effect on meaning. Median prose prefers symmetry, so a dispute in which the evidence is heavily one-sided can come out sounding evenly balanced. Style guides written for assisted work should therefore cover proportion as well as tone: how much space a claim earns, and whether the copy reflects the actual weight of the record.

7. Rights, licences and confidential material

Two rights questions recur. The first concerns input: pasting an embargoed release, an unpublished contract, a source’s message or personal data into a third-party tool may breach confidence, copyright or data-protection duties, depending on the tool’s terms and where processing happens. The second concerns output: text produced by a system trained on other people’s work can reproduce distinctive phrasing, and the ownership position varies by jurisdiction. Neither question has a single universal answer, which is why newsroom policies normally name specific approved tools, specify which categories of material may never be pasted into any of them, and require sign-off for anything outside the list. Check the current position with a qualified professional.

8. Telling the reader

Disclosure works when it is specific. “This article was produced with AI” tells a reader nothing they can act on. “The transcript of this interview was produced by automatic speech recognition and checked against the recording by the reporter” tells them exactly what was machine-made and what was verified. Useful disclosure names the stage, the artefact and the human check. It also has a threshold: most desks do not disclose spellcheck, and most do disclose generated prose or a synthetic voice. Publishing the threshold itself — a short, stable policy page that says what is disclosed and what is not — is more credible than a label whose meaning changes from item to item.

Assistance levels and the check they require

The right-hand column reflects common practice across European newsrooms rather than any legal requirement. Individual publishers set their own thresholds and publish them; the EU AI Act sets separate transparency duties for certain synthetic outputs.

Before an assisted draft is published

• Every number, date, name and place in the draft has been traced to a primary record.
• Every direct quotation has been matched to a recording, transcript or original text.
• Any citation supplied by the tool has been opened and read, not merely reproduced.
• Hedges in the source material survive into the published sentence.
• The prompt and the tool used are recorded in the item’s internal note.
• No confidential, embargoed or personal material was pasted into an unapproved tool.
• A named person on the desk has signed off the copy and can answer for it.
• The disclosure line, if the threshold is met, names the stage and the human check.

Checklists reduce omissions; they do not transfer responsibility. The person who publishes an assisted sentence answers for it exactly as they would for a typed one.

Questions readers ask