Our Approach to News Curation

This document describes how Tailor My News selects, ranks, and presents articles in your daily briefing. It is written for users, researchers, and grant reviewers who need to understand our algorithmic decision-making process.

Selection Criteria

Every article in your briefing passes through a multi-factor selection pipeline. No single criterion determines inclusion or exclusion. The system evaluates candidates along four orthogonal dimensions, each producing a normalized score between 0 and 1. The final ranking is a weighted composite of these scores, with weights adjusted by the user's diversity preference setting.

Topic Relevance

Semantic similarity between article content and the user's declared interests. We use embedding-based matching rather than keyword lookup, which allows the system to surface articles that are conceptually related to your interests even when they do not share exact terminology. A user interested in "machine learning" will also receive articles about neural architecture search or model distillation, without needing to enumerate those subtopics manually.

Source Authority

An assessment of the publishing source's editorial standards, factual accuracy track record, and institutional credibility. This is not a political alignment score. A source receives high authority marks if it maintains editorial independence, publishes corrections, and attributes claims to identifiable sources. Wire services and established newsrooms with public editorial standards score highest.

Recency Decay

Articles lose relevance over time according to a logarithmic decay function. Breaking news from the last 6 hours receives full recency weight. Articles older than 48 hours receive substantially reduced weight, though they can still appear if their topic relevance score is high enough. This prevents the briefing from becoming a stale archive while preserving important stories that develop slowly.

Diversity Balancing

After initial ranking by the three factors above, a diversity rebalancing pass ensures that the final selection is not dominated by a single source, geographic region, or viewpoint cluster. This pass may promote lower-ranked articles if doing so materially improves the diversity score of the overall briefing. The aggressiveness of this pass is controlled by the user's diversity preference.

Source Evaluation

Source evaluation is a persistent process, not a one-time classification. We assess sources along three axes:

Reliability. Does the source consistently publish factually accurate reporting? We draw on third-party media reliability assessments and monitor correction frequency. Sources that routinely fail fact-check evaluations are excluded from the candidate pool entirely, regardless of topic relevance.

Editorial Standards. Does the source maintain a separation between news reporting and opinion? Does it attribute claims to named sources? Does it publish corrections and retractions when warranted? We weight sources that adhere to recognized journalistic standards (SPJ Code of Ethics, IPSO Editors' Code) more heavily than those that do not.

Geographic Coverage. We track the geographic focus of each source to ensure that your briefing is not inadvertently confined to a single national perspective. A user interested in "climate policy" should receive coverage from European, Asian, African, and Latin American outlets, not exclusively from domestic publications. Geographic diversity is measured at the source level and the article level independently.

Diversity Balancing

The diversity balancing algorithm operates as a post-ranking reweighting step. It measures three distinct types of diversity and optimizes for a composite score:

Viewpoint diversity. Articles are classified along a simplified editorial orientation spectrum based on the source's historical editorial positioning. The algorithm monitors the distribution of orientations in the candidate set and introduces corrective weight to prevent over-representation of any single orientation cluster. This is not about "balance" in the false-equivalence sense. It is about ensuring that the user is exposed to the range of serious analysis on a given topic.

Geographic diversity. Articles are tagged with their primary geographic focus (the region they are about, not necessarily where the publisher is located). The diversity pass ensures representation across at least three geographic regions when sufficient relevant content exists.

Topic diversity. Even within a user's declared interests, the algorithm prevents any single subtopic from consuming the entire briefing. If a user follows "technology," they will not receive ten articles about a single product launch unless that event is of exceptional significance.

Users control the intensity of diversity balancing through three preference levels: Focused (minimal rebalancing, highest topic relevance), Balanced (moderate rebalancing, recommended default), and Diverse (aggressive rebalancing, maximum source variety). The diversity score visible in the transparency dashboard reflects the composite outcome of these adjustments.

What We Do Not Do

  • No engagement optimization. We do not select articles based on predicted click-through rates, time-on-page, or any behavioral engagement metric. The curation algorithm has no access to engagement data and cannot use it as a ranking signal.
  • No outrage amplification. Emotional valence is not a ranking factor. Articles are not promoted because they are more likely to provoke anger, fear, or indignation. Sensationalized headlines receive no preference.
  • No filter bubble reinforcement. The diversity balancing algorithm is specifically designed to counteract the natural tendency of relevance-based systems to narrow a user's information exposure over time. If your diversity score drops below 40, the system surfaces a warning and recommends adjustments.
  • No advertiser influence. Article selection is not influenced by advertising relationships. Tailor My News does not accept sponsored content placement within briefings. There is no mechanism by which a third party can pay to have an article included in or excluded from any user's briefing.
  • No social pressure signals. We do not use "trending" in the viral sense. Articles are not promoted because other users on the platform are reading them. Each user's briefing is independently curated from their own interest profile.
  • No opaque personalization. Every curation decision is logged and visible to the user in the transparency dashboard. There are no hidden ranking factors. If an article appears in your briefing, you can see exactly why.

User Control

The system provides multiple points of user intervention in the curation process. Users can add, remove, or modify their interest topics at any time. Source preferences allow whitelisting and blacklisting of specific publications. Topic muting temporarily suppresses content about subjects the user does not wish to see. The diversity level preference adjusts how aggressively the algorithm diversifies source selection.

These controls are not decorative. Each one directly modifies the parameters used in the selection pipeline. Whitelisting a source increases its authority weight for that user. Blacklisting a source removes it from the candidate pool entirely. Muting a topic sets the topic relevance score to zero for matching articles during the mute period.

The transparency dashboard provides real-time visibility into the aggregate effects of these choices. Users can see their source distribution, topic distribution, geographic spread, and overall diversity score. This feedback loop is designed to support informed decision-making about information consumption, not to gamify engagement.

Transparency Commitment

Every algorithmic decision made by Tailor My News is logged and made available to the user who receives it. The curation log records the selection reasons for each article, including which criteria contributed to its inclusion, what score it received on each dimension, and whether it was promoted or demoted by the diversity balancing pass.

This is not selective transparency. We do not show users a simplified version of the decision while hiding the real logic elsewhere. The data in the transparency dashboard is the same data the algorithm uses. There is no secondary, hidden model operating behind what we disclose.

We believe algorithmic transparency is a prerequisite for epistemic autonomy. If a system decides what information you see, you have a right to understand the basis for those decisions. This commitment extends beyond the interface: we are prepared to explain our methodology to any researcher, regulator, or grant reviewer who requests it.

Research Foundations

The design of this curation system draws on several bodies of research in information science, media studies, and cognitive science. The following are foundational works that inform our approach.

  • The Filter Bubble: What the Internet Is Hiding from You Eli Pariser (2011). The foundational text on algorithmic personalization and its consequences for democratic discourse. Our diversity balancing algorithm is a direct response to the filter bubble problem Pariser identifies.
  • Republic: Divided Democracy in the Age of Social Media Cass Sunstein (2017). Analysis of how information filtering creates echo chambers and undermines deliberative democracy. Informs our emphasis on viewpoint diversity.
  • Algorithmic Accountability Reporting Nicholas Diakopoulos (2015). Framework for transparency and accountability in automated decision-making systems. Directly informs our curation logging and transparency dashboard design.
  • The Information Diet Clay Johnson (2012). Argument for conscious information consumption analogous to nutritional awareness. The "nutrition label" metaphor in our transparency dashboard originates from this work.
  • Epistemic Autonomy and the Architecture of Information Systems Theoretical framework for understanding how information system design affects an individual's capacity for independent judgment. Our commitment to user control and transparency derives from the principle that information systems should support, not supplant, autonomous reasoning.
  • Measuring Media Diversity Philip Napoli (2011). Empirical approaches to quantifying source diversity, content diversity, and exposure diversity. Our diversity score calculation draws on the Shannon entropy-based methods Napoli describes.

For Researchers and Reviewers

If you are conducting research on algorithmic curation, media diversity, or information system transparency, we welcome inquiries. We are prepared to provide additional technical detail, answer methodological questions, or discuss our approach in an academic context.

Contact Us