Let editors fix the floor of a recommender, and engineers tune the rest
Newsrooms and public broadcasters argue about recommenders as if the choice lay between editors and algorithms. The workable answer is a written floor, owned by editors, beneath which no model may rank, and genuine freedom above it.
Arguments about recommendation systems in newsrooms and broadcasters usually have two camps. One holds that algorithms flatten journalism into whatever keeps people scrolling, and that editors alone should decide what readers see. The other holds that readers now expect a service shaped to them, and that an edited front page is a relic of print. I have worked with both camps, first at a Norwegian public-service broadcaster and now at a Danish news publisher, and I think the argument is badly framed. The useful question is not whether to personalise. It is which decisions a model may never make, and who writes them down.
Engagement is a measure, not a mandate
A recommender learns from whatever it is told to maximise. In streaming and news that is usually some form of engagement: viewing time, articles read, return visits. Those are reasonable things to measure. They become dangerous when they are treated as the organisation's purpose, because each of them systematically undervalues work whose worth does not show up as time spent.
At Fjellvang Kringkasting, the Norwegian broadcaster where I led personalisation from 2016 to 2021, the first fully personalised test of the streaming front page raised viewing time per visit by about 9 per cent. It also nearly removed news, regional programmes and Sámi-language content from most users' screens, because those were watched in short sessions or by smaller groups. Children's programmes, which we deliberately did not track by individual profile, scarcely appeared on family accounts at all. Nothing in the model was malicious. It simply had no way of knowing that a public broadcaster exists partly to put those programmes in front of people.
Newspapers face a quieter version of the same problem. At Vesterlund Mediegruppe a model optimised for reading time favours long features and crime coverage. It undervalues short, essential local reporting: a council's budget vote, a school closure, a planning dispute. Readers say they subscribe for exactly that reporting, and our cancellation surveys bear them out, but they do not linger over it. A model chasing minutes would, in time, teach the newsroom that local accountability journalism does not perform.
Writing the floor down
The arrangement I have come to recommend has three parts.
First, editors own a written floor. It lists what the recommender may never override. At Fjellvang the floor was a third of front-page positions chosen each morning by an editorial desk, at least one news item on the first screen and no individual profiling of children's viewing. At Vesterlund the floor is that no model ever reorders the editors' front page, and that personalisation runs only in a labelled module beneath it, which readers can switch off. The floor is short, specific and approved by the editor-in-chief, not by the data team.
Second, engineers have real freedom above the floor. Within the personalised space the data team may test, tune and replace models without seeking permission for every experiment, provided each change is logged and its effect on a small set of agreed measures is reported. Editors should not have to approve feature weights, and data scientists should not have to defend every ranking choice in an editorial meeting. Separating the two spaces is what makes collaboration possible at all.
The floor also needs a review date. Ours at Vesterlund is reread every spring by the editor-in-chief, two section editors and my head of data science, with a year of evidence in front of them: what the module recommended, which stories readers reached through it, and where local or minority reporting lost ground. In 2025 the review added one rule, that a story corrected after publication may not be recommended until the correction is visible at the top, because the model had kept surfacing a popular article whose error we had corrected only in a note at the foot of the page. That kind of rule would never emerge from an engagement metric. It came from an editor reading the log.
Third, the arrangement is published. Readers should be able to find a plain description of what the editors decide, what the model decides and which signals it uses. The EU Digital Services Act requires online platforms to explain the main parameters of their recommender systems and, for the very largest, to offer an option not based on profiling. Whether those duties reach a news site is a matter for lawyers. The practices are sound, and we adopted them anyway.
What the floor costs and why we keep paying
A floor costs engagement, and it is important not to pretend otherwise. At Fjellvang the reserved share reduced viewing time per visit by 4.2 per cent in its first quarter compared with the fully personalised test. Better editorial selection narrowed the gap to 1.5 per cent after a year, but it never closed. At Vesterlund, keeping the edited front page intact and leaving personalisation off by default for new readers means our click-through on recommended articles is lower than vendors tell us it could be.
In return we get things engagement measures do not show. At Fjellvang the share of weekly users watching at least one news or documentary programme rose from 34 to 47 per cent, which is much closer to what a public broadcaster is for. At Vesterlund, readers who use the personalised module read more local stories, not fewer, because the floor keeps local reporting in view and the model then learns that those readers value it.
There is a governance benefit as well. When a recommender's rules are written down and owned, a board can hold someone accountable for them. When they live only in a model's weights, nobody owns the outcome, and the first anyone hears of a problem is a complaint from a reader or a story in a rival paper.
My advice to editors is to stop opposing personalisation in principle and start writing the floor. My advice to data teams is to welcome the floor, because it gives them a mandate they can defend. And my advice to boards is to ask to see the document. If it does not exist, the organisation's editorial values are being set by whatever its model was last told to maximise.