AI in the newsroom: practical guidelines for content teams

If more than one person on your team writes content, someone is already using AI: for a draft, to summarise an interview, to translate a paragraph on the fly. The problem with AI in a newsroom is not the tool itself, it is the absence of shared guidelines: without written rules every writer decides on their own what is acceptable, and the quality of what you publish becomes unpredictable. In this guide you will find the concrete points to put in an editorial AI policy, with examples taken from how we work.

Why written guidelines are needed

An unwritten policy does not exist. As long as the rules stay in the manager's head, everyone interprets them their own way and the arguments always come after the damage, never before.

Written guidelines serve three purposes:

• Standardising quality. If one writer publishes barely touched-up AI drafts and another rewrites everything by hand, the reader notices the inconsistency before you do.
• Clarifying responsibility. Whoever bylines the article is accountable for every fact in it, even if a model suggested it. It needs to be put in writing.
• Protecting the company. Confidential material uploaded to consumer tools, invented quotes, images with copyright problems: these are legal risks, not just editorial ones.

We advise our clients to keep policies short: one page, with concrete examples of what is allowed and what is not. Twenty-page documents stay shut in a shared drawer.

Permitted uses and forbidden uses: the heart of the policy

The most useful part of the policy is an explicit list, split into two columns.

Uses that make sense to allow in most newsrooms:

• brainstorming headlines, angles and outlines;
• summaries of your own material (transcripts, minutes, internal documents);
• rephrasing paragraphs already written by a person;
• first drafts of repetitive formats, such as newsletters or descriptions, always destined for review;
• checks on grammar, typos and terminology consistency.

Uses to forbid explicitly:

• publishing generated text without documented human review;
• including data, statistics or quotes proposed by the model without verifying them against a primary source;
• uploading personal data, contracts or confidential client material to tools you have not assessed from a privacy standpoint;
• generating reviews, testimonials or experiences that never happened.

The last item sounds obvious, but it is the one we have seen violated most often, usually in good faith: a sample text that ends up online because of a rush.

Transparency with readers and clients

You must decide beforehand, and in writing, when to disclose the use of AI. The criterion we suggest is simple: transparency is owed when AI determined the substance of the content, not when it helped with the form. An article whose structure and content come from a model deserves a note; a paragraph rephrased to flow better does not.

If you produce content on behalf of clients, the picture changes: there, transparency belongs in the contract. Specify whether and how you use AI in the process, who does the review and who answers for errors. A client who finds out on their own, perhaps from a text that looks too much like a competitor's, is a lost client.

Bias and fact-checking: the checks stay human

Models generate plausible text, not true text. The difference shows in the details: dates, numbers, names, attributions. The policy must make a verification step mandatory, with precise rules:

• every quantitative claim must be traced back to a primary source consulted by a person;
• every quote must be verified against the original text, because models tend to reconstruct them from memory;
• on topics where the model may reflect stereotypes (professions, origins, gender) a dedicated re-read is needed, especially in content with examples and characters.

An operational trick that works: ask the writer to highlight in the draft every factual claim suggested by the AI. What is not highlighted is their own work; what is highlighted must be verified before publication.

The workflow matters more than the tools

The choice of model is the least important part. What determines quality is the workflow: who writes the brief, who generates, who reviews, who approves. A minimal workflow that holds up well:

1. brief written by a person, with angle, audience and sources;
2. draft, with or without AI depending on the format;
3. writer's review for accuracy and tone;
4. final check by an editor other than the person who wrote it.

The weak point of many newsrooms is dispersion: prompts saved in ten different documents, versions circulating over chat, no trace of who verified what. When volume grows, it pays to centralise: an internal panel with approved prompts, the status of each piece and the review history. It is the kind of tool we build as custom software for editorial and marketing teams, integrating AI models directly into the existing workflow instead of adding yet another platform to learn.

Want to bring order to your editorial workflow?

If your team produces content and you want to integrate AI with clear rules and tools built around the way you work, we can help: we design custom software for newsrooms and marketing teams, from editorial panels to model integrations. Tell us how you work today in a free call: together we will figure out where AI saves you time and where a person is still needed.