Table of contents
Since 2 August 2026, the EU AI Act has required publishers to label news text generated by AI, unless a person has reviewed it and someone holds editorial responsibility for it (Marko explains the rule here). So every newsroom using AI faces the same question: should we label, and will readers trust us less if we do?
Surveys suggest they might. Across 28 markets, only 36% of people are comfortable with news made by humans with the help of AI, and just 19% with news made mostly by AI with human oversight (Reuters Institute, 2024). But a survey question is not an article on a news site, so we tested the real thing.
This post is a short tour of the results. The full research paper, with every number, model and robustness check, is free to read and download.
One article, three bylines
We wrote one short, routine local story about a city's new bus timetable, the kind of service news newsrooms hand to AI first, and put it on a realistic mobile news page.
679 adults in seven countries read it: five bound by the AI Act (Croatia, Italy, Germany, France and Spain) and two outside it (the UK and the US), each in their own language and set in a real local city.
Everything was identical except the byline, which software picked at random for each reader.
Because the version was random, the three groups are alike in every way, as in a drug trial, so any difference in how they judged the article comes from the byline alone. Straight after reading, people rated the article's quality, credibility and professionalism from 1 to 7, and told us whether they would read it to the end.
What we found
The label did not make the article look less credible
The journalist-only version averaged 5.78 out of 7. Both AI versions came in about five hundredths of a point lower, well within chance. We can rule out any penalty larger than about 0.2 points, and “with AI” and “created by AI” scored the same.
Readers were just as willing to read it
Would read an article like this to the end
If a label hurt, more readers would skip the story or leave the site. They didn't.
Most readers already assume AI is involved
With no label at all, 62% thought AI had a hand in the story, and only 18% believed a human wrote it alone. For most readers, a label does not reveal a secret. It confirms what they already suspect.
What this means for newsrooms
For routine news, transparency did not cost trust. Neither the voluntary “with AI” label nor the mandatory “created by AI” one lowered credibility or readership, and no country showed a significant penalty. The caveats: we tested one calm, factual story, and our readers came from an online research panel, so they are younger, more educated and more at home with AI than the average adult.
Check our work, or repeat it
The AI named in the byline is ours, so we want this study to be easy to verify. We fixed the main outcome before the analysis and report every outcome we collected. And everything you need is public: the anonymized data with a codebook, the analysis code, the full questionnaire and the article itself. Anyone can reproduce every table and figure in the paper, or run the same experiment with their own story and readers.
Next in the series
- Did readers even notice the label? Fewer than you would think, and that cuts both ways for regulators.
- Beliefs versus ratings. The labels changed what readers thought about how the article was made. So why not how they rated it?
- Does it depend on the reader? We know each reader's age, gender, education, news habits and how often they use AI at work. General attitudes to AI-made news split sharply along these lines. Did reactions to the label?
- The surprise in version C. Readers who saw “created by AI” behaved differently in one way we did not expect.
Read the full paper
The Credibility Cost of Mandatory AI Disclosure: Experimental Evidence from Online News in Seven Countries. Ivan Mužić and Paulo Šanković, October 2026.
Download the PDF→
Written by
COO and co-founder of NewsLabs. Former central bank researcher, now focused on how people judge and use AI output. This connects my research background with what we’re building at NewsLabs.
LinkedIn
