On a Monday morning, a foundation that helps people facing violence publishes a short guide to finding support. It commissioned the guide from a supplier that uses a text-generation assistant; the piece was reviewed quickly because it was meant to accompany a campaign. One recommendation includes a telephone number for immediate help. A reader calls from a station, carrying a suitcase and running low on credit. A recorded message tells her that the number does not exist. This is not a harmless typo: the guide offered a concrete action, and it failed at the moment someone tried to take it.
When the coordinator asks what happened, the supplier says it only delivered a draft generated from prompts. The foundation says it relied on a specialist company. Both statements may be true, yet neither answers the question that matters to the person seeking help: who answers for material she received as published information? The answer cannot be “the system.” A tool can suggest sentences, combine documents, or invent a reference with persuasive confidence; it cannot accept a duty, hear a complaint, or repair harm. The party that decides to place a text before an audience keeps that responsibility.
Publishing means owning a decision#
Editorial responsibility does not require pretending that every organisation writes everything in-house, nor does it make technology forbidden. A foundation may commission translation, design, research, editing, or generation tools. A first draft may also help organise a long guide. The boundary is crossed when that division of labour is used to dissolve the final decision. The audience does not receive a chain of subcontractors. It receives a guide carrying the name, logo, and distribution channels of the organisation that released it. That organisation has adopted the content as safe enough to shape a reader’s conduct.
Editorial responsibility does not require pretending that every organisation writes everything in-house, nor does it make technology forbidden.
It helps to separate assistance with writing from permission to act. A sentence such as “it may help to book an appointment” needs context, but it does not carry the same risk as “call this number now,” “submit this document by Friday,” or “this service is free.” Actionable claims alter what a person can do today, so they deserve checking that matches their consequences. It is not enough that they sound likely or sit inside fluent prose. The publisher must be able to show where they came from, when they were checked, and what their real scope is.
In the guide’s case, a useful editorial review would have marked the telephone number as high-impact information. Before publication, a named person with authority to decide should have tested the call, confirmed the hours, asked which area the service covers, and checked whether it serves the people addressed by the campaign. If the resource could not be verified, the guide could remove the instruction, offer a confirmed institutional channel, or say plainly that it was still looking for a local contact. None of those options fills a list as quickly, but each is more responsible than filling a gap with an automated answer.
Give different claims different levels of scrutiny#
A sound editorial policy does not treat every sentence as if it required a court investigation. It does need an explicit way to classify risk. Information that urges urgent action or affects rights, health, safety, money, housing, or reputation should receive strengthened review. So should numbers describing a problem, names of organisations, attributed quotations, and legal or administrative instructions. A metaphor or a background summary calls for another kind of reading: fidelity of meaning, proper attribution, and consistency with the piece’s purpose.
This classification prevents two opposite mistakes. One is freezing a publication because every word cannot be proved to the same degree of certainty. The other is calling a text a “draft” even though it contains decisive directions, then using that label to lower the standard. The operational question is simple: if this sentence were false, stale, or incomplete, what might someone do because of it? The greater the consequence and the harder it is for the reader to detect the error, the more direct the checking should be.
Checking must lead to sources that can support the claim, not to a chain of repetitions. For a telephone number, the strongest source is the service itself: its official site, documented confirmation, or a test call. For a deadline, it may be the current notice. For a statistic, it is the original report and its methodology. A generated answer that cites an article, and an article that cites an old release, do not replace the origin. Editorial work consists precisely in walking that path backwards before asking other people to move forwards on the basis of the information.
Name the person who can stop publication#
“The team reviewed it” is not an assignment of responsibility. It is a phrase that, when something goes wrong, allows everyone to have participated while leaving no one with authority. Each guide needs an identifiable editorial lead, as well as a backup for absences. This is not about manufacturing a default scapegoat. It is about defining a role with time, access to sources, and the real ability to request changes, delay release, or withdraw a piece. If the person who approves cannot see the source material or pause the campaign, the title is decorative.
The lead does not work alone. The person who knows the programme can confirm whether a recommendation represents the service. The person who answers enquiries knows the obstacles people encounter. The person who manages the website can replace a page or add a warning. Those contributions still need to be organised. A very short internal table can record which claims require evidence, who checked them, where the evidence is kept, and when they must be reviewed again. The aim is not to manufacture paperwork for a distant audit. It is to stop an important decision from disappearing into scattered messages or someone’s memory of a meeting.
Clarity is equally important in the relationship with a supplier. A contract can require the company to deliver links, documents, versions, and uncertainty notices that support its proposals. It can also forbid the presentation of unverified material as fact. No clause, however, makes the supplier a substitute for responsibility to the public. The foundation may pursue a contractual breach; the reader needs an understandable correction and a useful alternative now. Those are separate paths, and the second should not wait for the first to conclude.
Versions are part of the evidence#
When a guide changes, keeping the previous version is not an admission of defeat. It is a condition for knowing what was said, when it was available, and what was corrected. Without a history, the foundation cannot establish whether the nonexistent number reached a download, a newsletter, or a shared screenshot. It cannot accurately answer a person who says she followed the instruction. Erasing the trail may look like a fast repair, but it leaves affected people without an explanation and the team without a lesson.
A proportionate record does not need elaborate tools. It can hold the publication date and time, a version identifier, the sources for actionable claims, the name of the person who checked them, and a note on material changes. When a fact is replaced, the record should say what changed, why, and how widely the earlier version was available. This material must be protected if it contains personal data or sensitive contacts; preserving evidence does not mean exposing case files. It means retaining enough memory to investigate and repair.
It also matters to distinguish source, interpretation, and decision. A source may say that a service operates Monday to Friday. The editorial interpretation may be that it is relevant to the guide’s audience. The decision may be to place it first among the recommendations. When those layers are blended, a dispute over opening hours looks like an attack on the whole campaign, and a reasonable correction becomes defensive. Separating them allows precise repair: the hours were wrong, the recommendation is suspended, an alternative is being checked. Readers can see what changed and what remains valid.
Correction reopens the channel#
A correction protocol has to exist before a failure, because urgency in a crisis favours incomplete responses. When a telephone number does not exist, the first task is to protect people who may try to use it: remove or disable the false instruction, place a visible warning where the guide appeared, and offer a confirmed resource if one is available. If there is no safe alternative yet, it is better to say so plainly than to improvise another. The correction should be readable without requiring anyone to know the organisation’s internal history.
Then comes reconstruction. The team needs to establish which version contained the number, how long it was accessible, which channels carried it, and what review took place before it was released. This is not to turn an investigation into a spectacle. It is to determine who must be notified and which safeguard was absent. If the guide went by email, the correction goes by email. If it travelled on social media, a discoverable correction belongs there. If partner organisations distributed a PDF, they need a replacement version and a short explanation. Correcting only the main webpage is like repairing one door while leaving every copy of the map open.
A responsible correction notice says what was wrong, what the current information is, when the problem was found, and what action has been taken. It need not dramatise the incident or blame a tool. It should avoid language that displaces the harm, such as “a technical inconsistency,” when people were given a route to help that did not exist. Direct language lets an affected reader decide whether she needs to revisit an earlier action. It also opens a contact route for further reports and identifies who will answer them.
Withdrawal is different from correction and needs its own criteria. It may be necessary when the safety of a guide cannot be verified, when potential harm is immediate, or when its documentary basis is so weak that a partial amendment would mislead. Withdrawing material does not mean pretending it never existed. A page can say that the material has been removed, explain why in comprehensible terms, and direct readers to confirmed resources. The organisation can then decide whether to publish a corrected version, commission independent assessment, or abandon that format.
Turn the failure into a practice of care#
The station scene shows why editorial responsibility is not an abstract debate about artificial intelligence. The problem did not begin when an assistant produced an unsupported string of digits. It began when that string passed through several human decisions without anyone treating it as a promise capable of directing a vulnerable person. It is reasonable to require the supplier to improve its safeguards. Yet a foundation does not protect its audience by waiting for the next system to make fewer mistakes. It protects them by designing a chain of decisions that catches an error before it becomes public instruction.
The next guide does not need a burdensome ceremony. It needs five actions that someone can check: appoint the editorial lead, flag actionable claims, verify them with a direct source, preserve sources and versions, and rehearse how the material will be corrected or withdrawn. The final question for every publication is concrete: if someone follows this instruction this afternoon and it goes wrong, will we know what happened, how to warn them, and who can act? If the answer is no, it is not yet time to publish. Delegating writing may save time; delegating judgment leaves the audience alone with the error.
