AI in publishing: what 655 readers taught one newsroom

Newsrooms have spent two years testing generative AI, and the FIPP World Media Congress in Madrid showed what happens when it becomes part of the daily production cycle. A study of 655 readers across seven countries found that transparency labels do not dent credibility, and more than 60% never noticed them. Meanwhile UK editors say reader revenue changes how stories are packaged, not what gets commissioned. This analysis sets out where automation earns its place, how to keep brand voice under editorial control, and what to measure in the first 90 days.
Key Takeaway What it means for your publication
Reader trust survives the label In a study of 655 readers across seven countries, a transparency label did not change how credible an article looked. Publishers can disclose AI involvement without fearing a backlash.
Most readers never saw the label More than 60% of participants could not recall seeing a disclosure after reading a few articles, which suggests the argument over labelling matters more inside newsrooms than outside them.
Automation is strongest on routine output Transcribing a full parliamentary session or parsing a spreadsheet of tourism statistics frees reporters for interviews and original reporting rather than replacing them.
Brand voice is the control point Enterprise tools are being configured to match a title’s tone and style, which keeps editors in charge of what published copy sounds like.
Paid readers want analysis, not news Editors at three UK national titles report that reader revenue changes what gets packaged as premium, not what gets commissioned.
Measurement makes the case internally Connecting a workflow change to time saved per reporter is what wins over sceptical colleagues and finance teams alike.

AI in publishing has spent two years in the testing phase, and the interesting question is no longer whether newsrooms will use it. The question is what happens when it becomes part of the ordinary working day, embedded in the way a reporter files, an editor commissions and a publisher decides what to pay for.

At the FIPP World Media Congress in Madrid, Ivan Šarić, co-founder and chief executive of NewsLabs, made the case that this transition is already underway in smaller markets. His company began with three people two years ago and now employs 14, building a publishing platform aimed at media houses in Central and Eastern Europe. The team’s argument is straightforward. Organisations without a dedicated AI workflow are losing ground to competitors producing routine news in a fraction of the traditional time.

Two pieces of evidence from that session deserve attention from anyone running an editorial operation. The first is a reader study that tested whether audiences punish publishers for using AI. The second is a live demonstration of what the tooling actually does, which is far less about writing prose than most commentary assumes. Together they sketch a practical path for editors who want the time savings without the reputational gamble.

This article unpacks both, looks at what UK editors are saying about how paying readers change commissioning, and sets out how to introduce automation into a production cycle without handing over editorial judgement. If you are weighing up where to start, the sequence matters more than the tool.

Why AI in publishing has moved past the experiment stage

The first phase of generative AI in newsrooms looked a lot like a series of pilots. A team would test a tool for a few weeks, produce a handful of stories, write up the findings and then quietly stop. The pattern repeated across large and small publishers because the tools were built for general use, not for the specific rhythm of a news desk.

What changed, according to the FIPP Congress discussion, is the arrival of purpose-built enterprise tools. Šarić described a shift from consumer chat interfaces towards systems that sit inside the publishing platform, connected to the sources a newsroom already uses. That distinction matters. A general assistant waits for you to paste something in. A connected system watches the wires, the feeds and the public record.

The workforce argument is worth noting too. NewsLabs grew from three to 14 people in two years, which is a small operation by any measure. The claim is not that such a team can replace a national newsroom, but that a regional publisher can now buy capability that used to require an engineering department. You can read the full FIPP report on AI in publishing for the session detail.

What NewsLabs learned across two years

The most useful part of the company’s account is how little of it concerns generating text from a prompt. The work went into ingestion: getting video, social posts and statistical releases into a form the newsroom can act on. Most publishers underestimate this step, and most stalled pilots stall here rather than at the writing stage.

  • Ingestion beats generation. The value sits in handling hours of source material, not in producing a first draft from nothing.
  • Regional focus pays. A platform built for Central and Eastern European publishers handles local data sources that global tools ignore.
  • Small teams move first. With fewer people, a routine-work saving of a few hours a day is immediately visible in the output.
  • Integration is the product. Sitting inside the existing publishing platform matters more than any single feature.

The transparency question publishers keep asking

Every editorial board meeting about AI reaches the same fork in the road. If we use it, do we tell readers, and will they care? The instinct to worry is understandable. Trust is the asset a publisher cannot rebuild quickly, and a visible error caused by automation would be costly.

Rather than rely on instinct, NewsLabs ran a study with 655 readers across seven countries: the UK, the United States, France, Germany, Italy, Spain and Croatia. Participants read articles carrying three different bylines. One was written entirely by a journalist. One was written by a journalist with AI assistance. One was generated entirely by AI.

The result cuts against most of the industry’s assumptions. The transparency label did not change how credible the article appeared to the audience. Readers were also just as willing to finish AI-assisted pieces as human-written ones. Quality and relevance did the work that publishers feared only a human byline could do.

What 655 readers actually noticed

The detail that should change how you think about disclosure is recall. More than 60% of participants did not remember seeing a label at all after reading several articles. That is not evidence of deception. It is evidence that labels are a compliance tool rather than an editorial signal.

Put simply, the readers in the study judged the copy in front of them. If it was useful, they read it. If the byline disclosure was visible, most of them moved past it without registering it. The anxiety about disclosure is largely a newsroom conversation, and a reasonable one, but it is not a reader conversation.

There is a regulatory layer to keep in mind. The European Union requires labels on AI-drafted content that lacks human oversight, so disclosure is a legal matter in many markets regardless of audience reaction. The study simply removes the fear that complying will cost you readers. For a wider view of how audiences actually behave around generative tools, the Reuters Institute’s 2026 findings on AI habits are a useful companion, showing use growing fast while news remains a niche application.

Why trust is decided by quality, not disclosure

Šarić’s conclusion was that readers are largely indifferent to AI involvement provided the journalism holds up. That places the burden squarely on editorial standards, which is where it already sat. A weak story assembled by a careful human is still weak. A well-sourced piece with automated transcription behind it is still strong.

The practical implication is that your AI policy should be shorter than your style guide, and your style guide should do more work than ever. Rules about sourcing, attribution, verification and correction carry more weight than any paragraph explaining how a draft came together.

Where automation earns its place in the editorial workflow

The live demonstration at the Congress is the part worth showing to a sceptical editor, because it addresses the tasks nobody joined journalism to do. Two examples stood out, and both involve material that already exists in a publisher’s day but is usually abandoned for lack of time.

The first is video. The platform can ingest hours of footage, such as a lengthy parliamentary session, then transcribe, summarise and draft the relevant news stories within minutes. Reporters who once sat through an eight-hour meeting to extract two usable stories can now review a structured draft and spend the saved hours on calls and original angles.

The second is structured data. Šarić showed the system downloading and analysing a large spreadsheet of monthly tourism statistics from the Croatian Bureau of Statistics. Work that would occupy an economics reporter for hours became a draft in a fraction of the time, with the journalist free to interrogate the numbers rather than type them.

From eight-hour meetings to two stories

Consider what that change does to a small newsroom’s week. A council meeting, a committee hearing, a regulator’s public session: all produce hours of speech for a handful of publishable facts. The judgement about which facts matter stays with a human. The transcription, the timestamps and the first pass at a summary do not have to.

There is a discipline required, though. Long meetings contain material that is legally sensitive, defamatory if handled carelessly, or simply wrong when a speaker corrects themselves three minutes later. Automated summaries need a human read before publication, and that read should be a proper one rather than a skim.

Structured data and the economics beat

Statistical releases are the other obvious fit. Inflation figures, tourism numbers, housing completions, court listings and planning applications all arrive as spreadsheets. Parsing them is skilled work, but much of it is mechanical, and the mechanical portion is where mistakes creep in after a long day.

Automation here does something subtler than speed. It creates consistency, because the same checks run every time. That helps smaller titles most of all, since they rarely have a dedicated data journalist and often skip the story entirely. If you want to see how a connected workflow handles this on the publishing side, Publishrs.com’s editorial workflow tools are built around the same principle of routine tasks running in the background.

Building a workflow that keeps editors in control

Every tool in this category makes the same promise about control. The difference between a workflow that holds and one that drifts is whether the control is configured in advance or asserted after something goes wrong. Configuration is the part most publishers skip.

NewsLabs configures its platform to match the specific tone and style of each media brand, which is how editors keep authority over published copy while gaining the time saving. That is the right model, and it is available to any publisher with a clear style guide and a willingness to enforce it.

Style matching and brand voice

Before any draft is generated, decide what your title sounds like on a normal day. Sentence length, preferred terms, house spellings, the way you handle titles and honorifics, your approach to quotes and attribution. Write it down in a form a system can follow, and make sure it matches what your best reporters already do.

Then test it against published work. Take ten articles from the past month, run the same source material through the configured system, and compare. Where the output drifts, the fault is almost always in the brief rather than the technology, and fixing the brief improves your human onboarding at the same time.

What to measure in the first 90 days

Workflow changes live or die on evidence. Pick measures you can actually collect, and collect them before you start so you have a baseline. A short list is better than a dashboard nobody opens.

  1. Hours saved per reporter each week. Ask people directly and keep the number honest, including the review time automation adds.
  2. Output of original reporting. Track interviews, FOI requests and off-diary stories, which is where the saved hours should show up.
  3. Correction rate on automated drafts. A rising rate early is normal; a rising rate at month three is a configuration problem.
  4. Time from source material to publication. This is the measure your subscribers feel, particularly on breaking stories.
  5. Premium conversion by format. Which pieces persuade readers to pay, and whether the mix shifts as routine output grows.

What reader revenue teaches us about routine content

There is a useful counterweight to all the enthusiasm about output. Editors from three UK national titles, speaking at the IPSO Conference, were asked whether reader revenue had changed what they commission. The answer was consistent and slightly deflating: not much.

What reader revenue changed is packaging, measurement and the value exchange. Amanda Williams of the Daily Mail described the shift as one of presentation rather than substance: the news story, the story behind the story and the comment follow-up were always commissioned, but the later pieces are now the ones readers pay for. Chloe Hubbard of The Mirror made the same point from the other direction, keeping breaking news open while putting analysis and opinion behind the paid tier.

That division of labour is exactly where automation belongs. Routine news becomes cheaper to produce and stays free, which protects reach. The saved capacity goes into the analysis, interviews and off-diary work that persuade readers to subscribe. Publishers who automate the premium layer instead of the routine layer will find the economics move against them. The full discussion on reader revenue and editorial decisions is worth reading alongside your AI plan, as is the argument that structural problems rarely yield to hiring alone.

The same logic applies to audience data. Rachel Shields of The Sun described eight years of traffic data that told her how many people clicked without saying who they were. Logged-in subscribers change that, because the cohort becomes visible and coverage can be judged against retention rather than page views. A workflow that produces more of the right content is worth more than one that simply produces more content, and keeping audience data in the same place as your production cycle is how you tell the difference. Publishrs.com connects the two, so retention sits beside the content that earned it.

FAQ

Will readers punish us for using AI in our journalism?

The evidence so far says no, provided the journalism is good. In the NewsLabs study of 655 readers, a transparency label did not change perceived credibility, and most participants did not recall seeing one. Quality and relevance drove their judgement.

Do we have to label AI-assisted content?

It depends on your market. The European Union requires labels on AI-drafted content that lacks human oversight, and several national regulators are moving in the same direction. Treat disclosure as a compliance question and let your legal team set the wording.

Where should a small newsroom start?

Start with transcription and meeting coverage. The material already exists, the time saving is immediate, and the editorial risk is low because a reporter reviews the output before publication. Spreadsheet parsing is the natural second step.

Does automation replace reporters?

Not in the pattern described at FIPP. The reporters gain hours they previously spent on routine extraction and spend them on interviews, verification and original angles. What changes is the mix of work, not the number of people needed to judge a story.

How do we keep the brand voice consistent?

Write the voice down in a form a system can follow, including house spellings and preferred terms. Then test generated drafts against published articles monthly. Drift usually points to an out-of-date brief rather than a broken tool.

What is the biggest risk to manage?

Publishing without a proper human read. Long meetings contain corrections, sensitive material and statements that need verification. Build the review step into the workflow so it cannot be skipped when the desk is busy.

How quickly will we see a return?

Measure hours saved per reporter from week one, then watch where those hours actually go. In small teams the difference shows up within a month. In larger operations, expect the benefit to concentrate in a few desks before spreading.

Which content should stay entirely human?

Investigations, anything involving vulnerable people, legal risk or contested facts, and the analysis readers pay for. Routine output is where automation earns its keep, so protect the premium layer rather than filling it.

The next step is yours.

The pattern emerging from Madrid is less dramatic than the headlines and more useful. Readers are not waiting to catch you out. Reporters gain back hours that were previously consumed by material they did not choose. And the publishers who gain most are the ones who automate routine coverage while protecting the work that persuades people to pay.

That balance is easier to hold when the workflow and the publishing platform sit in the same place. Publishrs.com is built for exactly that arrangement, handling production and distribution so your editorial team can concentrate on the journalism only they can do.

Source: www.fipp.com

Publishrs.com

The official blog for Publishrs.com – the all in one digital publishing platform

Read More

Sign up for our Newsletter

Get the latest publishing news straight to your inbox