PaperlayerView exact revision on arXiv

Source-linked AI summary

Building Human Values into Recommender Systems: An Interdisciplinary Synthesis

Jonathan Stray, Alon Halevy, Parisa Assar, Dylan Hadfield-Menell, Craig Boutilier, Amar Ashar, Lex Beattie, Michael Ekstrand, Claire Leibowicz, Connie Moon Sehat, Sara Johansen, Lianne Kerlin, David Vickrey, Spandana Singh, Sanne Vrijenhoek, Amy Zhang, McKane Andrus, Natali Helberger, Polina Proutskova, Tanushree Mitra, Nina Vasan

arXiv:2207.10192v1cs.IRcs.SI

TL;DR

Recommender systems affect people and societies, raising the question of how to make them enact the values they serve and what objectives they should pursue. This paper synthesizes current thinking with an emphasis on industrial applicability. It shows how social-science measurement techniques apply to evaluating values-based outcomes in recommender systems.

  • Problem

    How recommender systems can enact the values of the individuals and societies they serve remains a central question, since avoiding harmful outcomes alone does not specify what systems should optimize.

    Source: §1 · pp. 2, 4

  • Method

    The paper synthesizes current thinking with an emphasis on industrial applicability, including social-science measurement techniques for evaluating values-based outcomes.

    Source: §1 · p. 4

  • Results

    The paper shows that social-science measurement techniques apply to evaluating values-based outcomes in recommender systems.

    Source: §1 · p. 4

  • Takeaways & Limitations

    A recommender policy can aim to maximize expected social welfare by combining stakeholders’ preferences through a social welfare function.

    Source: §5.3.1 · p. 30

    Value measurements are usually proxies, and using them for consequential decisions can trigger feedback that widens the gap between a metric and the underlying value.

    Source: §4.2 · p. 20

Abstract

from arXiv · show

Recommender systems are the algorithms which select, filter, and personalize content across many of the worlds largest platforms and apps. As such, their positive and negative effects on individuals and on societies have been extensively theorized and studied. Our overarching question is how to ensure that recommender systems enact the values of the individuals and societies that they serve. Addressing this question in a principled fashion requires technical knowledge of recommender design and operation, and also critically depends on insights from diverse fields including social science, ethics, economics, psychology, policy and law. This paper is a multidisciplinary effort to synthesize theory and practice from different perspectives, with the goal of providing a shared language, articulating current design approaches, and identifying open problems. It is not a comprehensive survey of this large space, but a set of highlights identified by our diverse author cohort. We collect a set of values that seem most relevant to recommender systems operating across different domains, then examine them from the perspectives of current industry practice, measurement, product design, and policy approaches. Important open problems include multi-stakeholder processes for defining values and resolving trade-offs, better values-driven measurements, recommender controls that people use, non-behavioral algorithmic feedback, optimization for long-term outcomes, causal inference of recommender effects, academic-industry research collaborations, and interdisciplinary policy-making.

1. Introduction

Recommender systems shape content across many domains, raising the question of how they can enact the values of the people and societies they serve. The paper offers an interdisciplinary synthesis of value-sensitive recommender design and practice, rather than a general theory.

1. Introduction

  • Recommenders affect individuals and societies across domains, making it important to ask how they can enact the values of those they serve.The cited examples include social media, news, hiring, shopping, and entertainment.Source: §1 · p. 2
  • The paper draws on technical and interdisciplinary perspectives to synthesize current thinking and establish shared language for value-sensitive recommender design.It frames the work as a mix of conceptual, empirical, and technical effort.Source: §1 · pp. 2–4
  • Recommenders serve multiple stakeholders, including users, content producers, platforms, and non-users, whose values may need to be elicited.Possible methods include participatory design, surveys, interviews, and deliberative elicitation.Source: §1 · p. 3
  • Avoiding harmful outcomes alone does not specify what a recommender should promote or optimize.The paper notes that preventing false information does not determine what news should be selected or how journalism’s societal goals should be supported.Source: §1 · p. 4
  • Its contributions span a compilation of relevant values and research, a production-news example of operationalizing political diversity, social-science measurement, design approaches, policy gaps, and open problems.The paper also reviews empirical effects and current practice around the values it identifies.Source: §1 · pp. 4–5
  • The authors characterize the work as a non-comprehensive synthesis based on their collective experience and interdisciplinary discussion.They present it as a view of current thinking and practice and suggest future research directions.Source: §1 · p. 5

2. Values for Recommender Systems

The section assembles a context-sensitive set of values for recommender systems and discusses their tensions, evidence, and measurement challenges. It emphasizes that effects and priorities vary across stakeholders and settings.

Building a values framework

  • The value list combines prior compilations, practitioner judgment, and multi-stakeholder input, then refines definitions through feedback.It is intended as a useful, broad grounding for discussion, not a universal or exhaustive catalogue.Source: §2 · pp. 5–6

Value tensions

  • Values can conflict across people, society, and time horizons, so operationalizing one value can involve trade-offs with others.For example, privacy may conflict with usefulness, while short-term entertainment may differ from longer-term informational benefits.Source: §2 · p. 6

Usefulness and control

  • Control concerns users’ ability to choose content and personalization and understand recommendation processes, but perceived control may differ from actual control.The section also notes that user choices can conflict with other values, including safety.Source: Usefulness · p. 7

Well-being

  • Evidence on recommender-related well-being is mixed, and many studies do not isolate algorithmic selection from other social-media activity.The section identifies defining, measuring, and optimizing for long-term well-being as open problems.Source: Usefulness · pp. 8–9

Rights and fairness

  • Privacy has multiple forms and can trade off against usefulness, transparency, and fairness, while fairness itself involves stakeholders with differing needs.Those interests can conflict, and the section notes that people outside a platform may also experience recommender externalities.Source: Legal and Human Rights · pp. 9–10

Public discourse and information

  • The section identifies uncertainty about recommender contributions to misinformation-related harms and notes that editorial principles for personalized news remain difficult to translate into algorithms.Evidence about misinformation’s spread does not by itself establish the role of recommendation algorithms.Source: Public Discourse · pp. 10–11

Public discourse and diversity

  • Diversity can serve users, creators, and society, but its meaning and metrics vary; evidence about personalization, exposure, and polarization is mixed.The section notes that diversity metrics may not capture the social-science complexity of media diversity.Source: Public Discourse · pp. 11–12

Safety

  • The section distinguishes polarization from radicalization and finds that systematic evidence for recommender-driven radicalization remains weak, leaving causal understanding of user trajectories a major challenge.Reports of radicalization involving recommendations also mention other factors, while systematic studies have limitations in experimental power.Source: Public Discourse · pp. 12–13

3. The State of Practice

The paper reviews recommender practice through a general architecture and an illustrative product-change process, drawing on practitioner experience.

3. The State of Practice

  • The practice overview describes common commercial design patterns and a hypothetical process for engineering a specific value into a recommender, while stressing that people and teams shape system objectives and evaluation.These sections draw substantially on the experience of authors who work on large contemporary recommenders.Source: §3 · p. 13

3.1 How Recommenders Work

The section presents a simplified recommender architecture, from available content and user context through candidate selection, ranking, and possible re-ranking. It also defines engagement as behavioral signals used as indicators of stakeholder value, while noting they may not align fully with specific values.

3.1 How Recommenders Work

  • A simplified architecture begins with available items and moderation, then uses item, user, and contextual data to produce a recommended feed or slate.The design is illustrative rather than a description of any one system.Source: §3.1 · p. 13
  • Candidate generation efficiently narrows a potentially huge corpus to hundreds or thousands of items, which ranking scores for relevance and re-ranking may adjust for feed-level goals such as diversity.Ranking often predicts engagement, while re-ranking can consider properties of the selected items together.Source: §3.1 · pp. 13–14, Figure 1 · p. 14
  • Ranking combines predicted user responses with other signals, so optimizing for engagement may not optimize value for every stakeholder.Signals can include source credibility, playlist diversity, or whether an item tends to be inspiring.Source: §3.1 · p. 14
  • Engagement is defined as ordinary interaction behaviors thought to correlate with value to users, platforms, or other stakeholders, not as a complete measure of specific values.Some value signals may require non-behavioral data.Source: §3.1 · pp. 14–15

3.2 Implementing a Product Change

The section uses a political-diversity example to illustrate how teams might translate a value into a product change. The process spans risk research, measurement, design, implementation, evaluation, safeguards, and continued monitoring.

3.2 Implementing a Product Change

  • The example prioritizes political diversity to support awareness of multiple perspectives, while noting that both the assumed lack of diversity and the presumed tolerance benefit have mixed evidence.The process is illustrative and not a general theory or a universal account of company practice.Source: §3.2 · p. 15

3.2.1 Research unintended consequences

  • Teams may investigate how exposure diversity affects knowledge and attitudes, including possible backfire effects and impacts on other values.Proposed changes may be constrained by thresholds for declines in other metrics.Source: §3.2.1 · pp. 16–17

3.2.2 Develop a metric

  • Choosing a diversity metric requires an operational definition and may require multiple measures; external stakeholders have sometimes been involved to increase legitimacy.One described approach trains human raters and uses their judgments to create data for a predictive model of productive diversity.Source: §3.2.2 · p. 17

3.2.3 Consider different product designs

  • Product teams may compare interface changes with ranking changes before selecting an approach to increase diversity.Examples include showing related articles from other viewpoints or nudging consumption through ranking.Source: §3.2.3 · p. 17

3.2.4 Implement the product change

  • Adding diversity prediction to ranking is challenging because diversity applies to item sets, and scoring lists can be more complex and expensive than scoring items individually.The example may therefore require more efficient algorithms for ranking item sets.Source: §3.2.4 · p. 17

3.2.5 Evaluate and advocate

  • Teams may test a change offline and with small A/B tests before deployment, then weigh diversity gains against engagement, quality, safety, or stakeholder costs.The example describes internal negotiation when a diversity increase accompanies declines in other outcomes.Source: §3.2.5 · pp. 17–18

3.2.6 Establish guardrails to prevent reversion

  • Review processes and numeric guardrails can limit later product or algorithmic changes that would reduce the diversity improvement.Source: §3.2.6 · p. 18

3.2.7 Monitor outputs and outcomes

  • Ongoing monitoring can detect operational failures and shifts in users or behavior, while surveys and ethnographic research can assess whether metrics track the intended outcomes.The example includes checking for drift between measured diversity and the broader goal.Source: Figure 2 · p. 16, §3.2.7 · p. 18

4. Value Measurement

Value measurement connects broad human values to metrics, recommender choices, and outcomes, but metrics are imperfect proxies that require ongoing scrutiny.

4. Value Measurement

  • Figure 3 depicts how metric selection shapes item selection and engagement, which in turn shapes outcomes over time.The section frames measurement as connecting values to outcomes through recommender choices and user interactions.Source: §4 · p. 18, Figure 3 · p. 19

4.1 From Value to Metric

  • Operationalizing a value means defining metrics for it, and those choices influence a recommender’s function, making stakeholder involvement important.Safety, well-being, and public interest illustrate how difficult it is to translate broad values into concrete metrics and how choices create trade-offs.Source: §4.1 · pp. 19–20

4.2 Characteristics of a Good Value Measurement

  • A metric rarely captures a value fully, and its meaning can shift when consequential decisions depend on it.The authors recommend identifying measurement gaps, monitoring them over time, and using ongoing qualitative user research as values evolve.Source: §4.2 · p. 20
  • Metrics can misrepresent groups when measurement signals vary across demographics, and stakeholders may also question a metric’s legitimacy despite its accuracy.Transparency, external audits, or participatory processes are discussed as ways to support legitimacy.Source: §4.2 · pp. 20–21

4.3 Data Sources for Measuring Values

  • Surveys and human annotations provide other value signals, but surveys are limited by response frequency while annotation is costly and can be noisy or biased.Behavior can also be used to predict survey responses for users who were not surveyed, but those predictions remain proxies requiring monitoring against survey responses.Source: §4.3 · pp. 21–22
  • Behavioral signals are plentiful but only imperfect proxies for values, varying with users, demographics, context, and interaction recency.Dwell time, for example, can differ depending on an item’s position in a results list, and collecting behavioral data raises privacy concerns.Source: §4.3 · p. 22

5. Design Approaches

The paper groups value-oriented recommender design affordances into interventions on item ranking, user controls and feedback, and system trade-off optimization.

5. Design Approaches

  • Figure 4 identifies three design options: intervene on item ranking, provide user controls and feedback, or optimize system trade-offs across stakeholders.The options address item selection, user experience modification, and multi-stakeholder perspectives, respectively.Source: Figure 4 · p. 23

5.1 Item Selection

Item selection approaches use item and user information to rank content, while trajectory-oriented approaches consider sequences of recommendations and their possible longer-term effects.

5.1.2 User Trajectories

  • Recommenders often predict immediate behavioral responses, while sequence models represent latent user states that may inform future recommendations.Interpreting hidden states is challenging and tied to the engagement metrics being predicted and optimized.Source: §5.1 · p. 23, §5.1.2 · p. 25

5.1.1 Item ranking signals

  • Ranking signals include item topics and content-policy labels, but topics do not directly identify user needs or values.Inferring how a particular user will relate to an item remains an open challenge; human labels may also be biased or disputed, with implications for freedom of expression.Source: §5.1.1 · pp. 24–25

5.1.2 User Trajectories

  • Reinforcement learning can plan recommendation sequences over extended periods, including sequences intended to support diversity or adaptive learning.With suitable metrics and objectives, it could help align trajectories with well-being; current ranking techniques do not directly optimize future trajectories.Source: §5.1.2 · pp. 25–26
  • RL-based value alignment faces offline-data bias, large action spaces, slate interference, difficult latent-state modeling, and potentially long adaptation horizons.User adaptation to recommender changes may take 3–6 months to fully materialize.Source: §5.1.2 · p. 26
  • A safe reinforcement-learning approach improved both worst-case and average-case outcomes for the fraction of unhealthy items recommended to any one user.The approach regularized the mean proportion of unhealthy content among trajectories in the top αth percentile.Source: §5.1.2 · p. 26
  • Trajectory concerns include possible preference shifts associated with manipulation and exposure to excessive amounts of otherwise acceptable content.The empirical evidence for filter-bubble, echo-chamber, or polarization effects is mixed, while simulations have demonstrated polarizing preference shifts under multiple specifications.Source: §5.1.2 · pp. 26–27
  • Whether recommenders causally change user preferences and outcomes remains a major open problem, despite varied studies of content trajectories and recommender effects.The evidence described includes network analyses without personalization, bot studies modeling users as unchanging, and a randomized Twitter feed study.Source: §5.1.2 · p. 27

5.2 Controls and Feedback

Recommender communication ranges from explicit controls to passive feedback, but people may not understand or use controls, and their effects are not always predictable. Explainability and newer interfaces offer ways to communicate recommendations, while their usefulness depends on the goals and language involved.

5.2 Controls and Feedback

  • Greater control is associated with higher reported satisfaction and engagement, while apparent controls can raise satisfaction even when their effects are random (Jin et al., 2017; Jannach et al., 2017; Vaccaro et al., 2018).Control use does not always translate into better control, and users’ perceived diversity may not match actual diversity.Source: §5.2 · pp. 27–28
  • Most users remain passive because controls may be unknown, difficult, or not worth using; feedback serving social purposes as well as preference input is one proposed approach.Examples include item ratings, pairwise comparisons, and social-signal buttons.Source: §5.2 · p. 28
  • Control effects can be hard to predict in magnitude or duration, even when the affected content is clear; showing which items would change is one proposed design.The input-to-outcome link may be direct for exclusions but less obvious for requests such as finding similar videos.Source: §5.2 · p. 28
  • Channels can organize recommendations around user goals or needs, but they require relevance measures that capture the dimensions users care about.A music recommender continuing a human-generated playlist is one implicit channel example.Source: §5.2 · p. 28
  • Conversational recommenders may offer a more natural interface for expressing preferences, but must interpret metaphorical, imprecise, or subjective language.Recommender personae can establish a conversational context, such as explorer, diplomat, or expert.Source: §5.2 · pp. 28–29

5.2.1 Meaningful explainability

  • Explanations can pursue distinct goals, including transparency, trust, scrutability, or satisfaction, and some goals may conflict.A transparent explanation may not increase trust if it is incomprehensible or reveals undesirable system behavior.Source: §5.2.1 · p. 29
  • Explanations may describe a model’s workings, come from a post-hoc model, or be built into the recommender; explainable-by-design methods avoid interpreting learned models.Deep-learning explanations remain an active research area.Source: §5.2.1 · p. 29

5.3 Making Tradeoffs

Recommender tradeoffs concern competing stakeholders and values, and social choice, fairness methods, and optimization offer ways to frame them. Their application still depends on difficult preference and utility judgments, and designers cannot avoid normative choices.

5.3 Making Tradeoffs

  • Recommender tradeoffs span harms and benefits across people, stakeholder interests, and values; the paper frames social choice, fairness, and ranking optimization as approaches to making them.It proposes that tradeoffs be informed by expressed opinions of users and other stakeholders.Source: §5.3 · pp. 29–30

5.3.1 Tradeoffs in theory: Social Choice

  • Social choice represents each stakeholder’s preferences as utility over outcomes, combines them through a social welfare function, and selects policies to maximize expected welfare.The framework can express short- or long-term and individual or group outcomes, linking normative preferences to algorithmic design.Source: §5.3.1 · p. 30
  • Preference elicitation is demanding and may capture uninformed, coerced, addicted, unavailable, or strategically misrepresented preferences; defining outcomes and whose preferences count also requires normative choices.Behavioral feedback can signal what people value but cannot be naively treated as their true preferences.Source: §5.3.1 · pp. 30–31

5.3.2 Fairness and Tradeoffs among Stakeholders

  • Fairness can involve users, providers, or both, and research still faces challenges in defining it, measuring it, and designing efficient recommendation algorithms.The text distinguishes consumer fairness, provider exposure fairness, and joint consumer-provider fairness.Source: §5.3.2 · p. 31
  • Average performance can conceal subgroup disparities; methods have equalized prediction error rates on a commercial platform, but reducing utility for well-served users may be inappropriate when gains need not impose losses.Differences in group size or activity can affect training data and contribute to disparities.Source: §5.3.2 · p. 31
  • Items and producers compete for attention, and position bias can leave less-relevant items disproportionately overlooked; exposure-fairness methods seek to align attention with item value.Other normative goals include providing exposure to less-popular artists or showing qualified candidates from different groups at equal rates.Source: §5.3.2 · pp. 31–32
  • Multi-objective optimization can balance stakeholder utility and fairness, including Pareto-efficient rankings and approaches that trade small gains for well-served users for larger gains elsewhere.The cited methods also include contextual bandits and game-theoretic approaches, but these methods had not seen practical deployment in the passage’s account.Source: §5.3.2 · p. 32

5.3.3 Optimizing Tradeoffs

  • Ranking-weight optimization requires designers to define overall utility across metrics; Bayesian optimization can search weights and run data-gathering experiments, while interactive utility elicitation remains uncommon in practice.The tradeoff may involve choosing between more items viewed and greater time spent or reported satisfaction.Source: §5.3.3 · pp. 32–33

6. Policy Approaches

Policy approaches can shape recommender design, but translating broad societal goals into technical requirements is difficult. The section reviews risk management, accountability, and policy–technology translation challenges.

6. Policy Approaches

  • External governance spans government, regulators, and other authorized bodies; because recommenders serve varied products, the paper surveys policy categories rather than prescribing specific policies.It highlights gaps between policy terminology and recommender technology, including cases where requested capabilities may not yet exist.Source: §6 · p. 33

6.1 Risk Management vs. Value Sensitive Design

  • Risk-based frameworks focus on potential harms and generally mitigate them after systems are built, while value-sensitive approaches can require considering values during design.The proposed DSA and DSOSA illustrate risk assessment and mitigation; privacy-by-design provisions and content-diversity obligations illustrate design-phase approaches.Source: §6.1 · pp. 33–34
  • Regulators must define harms precisely yet broadly, but no single metric set suits every recommender, and deciding what values are acceptable remains difficult.Even useful measures leave open the normative question of what level counts as acceptable.Source: §6.1 · p. 34

6.2 Accountability

  • Transparency and audits can support accountability, but implementation must address disclosure goals, privacy, security, intellectual property, and limits on auditor capacity.Production code may be hard for outsiders to interpret, while shared user data can pose privacy risks and external auditors may lack resources or technical skills.Source: §6.2 · pp. 34–36
  • A review of 62 audits found eight of recommender systems, seven focused on distortion; some found no echo-chamber effects but found few news sources dominated results.The reported findings concerned Google News, Facebook News Feed, and Apple News.Source: §6.2 · p. 35
  • Trajectory audits using proxies may not model real behavior, while observational studies with real users cannot robustly infer recommender effects; on-platform experiments may be needed.The passage notes that experiments would require extensive collaboration between operators and external researchers.Source: §6.2 · pp. 35–36

6.3 Translating Between Policy and Technology

  • Policy terms such as “main parameters” and “not based on profiling” are difficult to map to useful controls, since systems have many interacting settings and profiling can include stated interests.An option based only on explicit user controls is presented as an alternative to one excluding profiling.Source: §6.3 · p. 36
  • “Amplification” lacks a well-defined measure for most recommenders because plausible baselines can mislead; proposed bridges include policy-maker education, sandboxes, and evidence-based policy.A raw-content-prevalence baseline can make spam appear to reduce amplification, while chronological feeds are not a natural baseline for many services.Source: §6.3 · p. 37

7. Open Problems

The paper identifies open problems in defining values, measuring them, and incorporating them into recommender systems, alongside challenges in establishing effects and coordinating research and policy.

Process for defining values and resolving trade-offs

  • There is no widely agreed process for prioritizing and measuring values or adjudicating trade-offs among values and stakeholders.Multi-stakeholder metric construction and user juries are possible approaches, but remain underdeveloped.Source: §7 · pp. 37–38

Better measurements

  • Values are difficult to operationalize, and social-science or policy measures may need refinement for specific recommender contexts.AI metric collections such as IEEE 7010 offer useful compendiums but may not transfer directly to contexts such as news recommendation or targeted advertising.Source: §7 · p. 38

Controls that people want to use

  • Most user controls are used by only a few percent of users, so useful designs may elicit feedback during ordinary use while distinguishing felt agency from effective control.Immediate feedback on setting changes, voting systems, or a “respect” button are suggested possibilities.Source: §7 · p. 38

Non-behavioral feedback

  • Behavioral signals can misrepresent preferences, while outcomes such as well-being may not be identifiable from on-platform behavior alone.The paper proposes surveying users and using the resulting data to build and validate models that predict survey responses as algorithmic objectives.Source: §7 · p. 38

Long-term outcomes

  • Most recommenders optimize immediate responses, while values-relevant long-term outcomes could be managed by systems with personalized optimization.Emerging reinforcement-learning methods may help, but may require cheap, accurate individual-level proxy metrics.Source: §7 · pp. 38–39

Causal inference of human outcomes

  • Sampling, confounding, and missing counterfactuals make recommender effects difficult to establish; on-platform experiments may measure group-level effects, while long-term effects remain especially hard to study.Individual case studies can be instructive but are difficult to generalize to platform scale.Source: §7 · p. 39

Industry-academic research collaborations

  • Important questions may require private platform data or on-platform experiments, but collaboration is hindered by access, privacy, security, integrity, intellectual-property, competition, and cost concerns.The paper points to privacy- and security-enhancing methods and legal or policy approaches to set collaboration terms.Source: §7 · p. 39

Interdisciplinary Policy-making

  • Policy and technical communities have substantial knowledge and terminology gaps, alongside value and political differences touching rights such as privacy and freedom of expression.The paper concludes that effective policy-making must be collaborative and interdisciplinary, though how to achieve this remains unclear.Source: §7 · p. 39

8. Conclusion

Recommender systems affect individual and social life and can serve or harm multiple interests. The paper synthesizes interdisciplinary perspectives on values, design, measurement, and regulation to identify directions for further research.

8. Conclusion

  • Recommenders can serve consumers, producers, and platforms, yet may also cause harmful effects or miss opportunities to do good.Building human values into them raises philosophical, regulatory, political, sociological, and technical challenges.Source: §8 · p. 39
  • The paper identifies relevant values, reviews industry practice and value measurement, and surveys design techniques and developing approaches to recommender regulation.It also describes modern recommender design and an illustrative process for shifting a value in a production system.Source: §8 · pp. 39–40
  • The authors consider a general theory of value-sensitive recommender design premature because the field is still emerging.Instead, they offer a menu of design techniques applied or likely applicable to production systems.Source: §8 · p. 40
  • The synthesis aims to provide shared language, starting points, and research directions for building human values into recommender systems.It also identifies a substantial knowledge and orientation gap between technology and policy communities.Source: §8 · p. 40

Appendix A: Table of Values

This appendix presents a multi-stakeholder-informed, non-comprehensive selection of broadly important human values relevant to recommender systems, pairing them with interpretations or definitions, candidate indicators or measures, and example design or implementation ideas. The values span civic engagement and accessibility to sustainability, progress, and meaningful work.

Appendix A: Table of Values

  • The values were selected from varied sources and refined through multi-stakeholder consultation, but are neither comprehensive nor ranked.The table is presented as one list of broadly important themes.Source: Appendix A · p. 41
  • Usefulness means providing a useful service, with retention and satisfaction surveys as possible indicators and highly rated items as an example design.The table also frames usefulness as helping users find relevant items.Source: Appendix A · p. 41
  • Liberty includes protecting users’ ability to explore information and express themselves, with transparency about suppression and appeals for removed content as example designs.The table also describes protecting human rights and freedoms and avoiding undue suppression of user posts.Source: Appendix A · p. 41
  • Control and agency mean enabling users to revise preferences, customize recommendations, and pursue their goals without platform manipulation.The table suggests recommender controls and frames agency as intentional action and control.Source: Appendix A · p. 42
  • Privacy, safety, and self-expression encompass user choice over personal-data use, avoiding harm, and empowering users to express and present their identity.Example designs include controls to limit data use and content-creation tools; the table also calls for reliable systems that do not harm living beings or their environment.Source: Appendix A · p. 42
  • Well-being and health values call for supporting life satisfaction and mental and physical health, including through screen-time management and healthy-habit support.The table treats well-being as multidimensional and allows objective and subjective measures; it also calls for accurate health information.Source: Appendix A · p. 43
  • Inspiration, personal growth, and recognition mean showing motivating content, supporting users’ development, and enabling others to acknowledge their contributions.The table offers inspiration measures and educational recommenders as examples relevant to these values.Source: Appendix A · pp. 43–44
  • Knowledge and connection values call for informing users about relevant topics and supporting social ties; civic engagement is also included as a value.The table gives news knowledge quizzes and recommendations for people or groups with similar interests as examples.Source: Appendix A · p. 44

Civic engagement and community

  • Civic engagement and community are linked to measures of participation, belonging, volunteering, and charitable giving, alongside ideas such as community-oriented recommendations and donation affordances.Examples include voter turnout, peaceful demonstrations, time in voluntary organizations, neighborhood belonging, and the General Sense of Belongingness Scale.Source: Appendix A · pp. 44–45

Accessibility, Inclusiveness

  • Accessibility and inclusiveness call for assessing diverse users’ experiences and applying accessibility guidelines and formats such as screen readers, transcripts, and captions.The passages name visually impaired and hard-of-hearing users as examples for assessment.Source: Appendix A · p. 45

Tolerance, Constructive Discourse; Duty; Care, Compassion, Empathy

  • Tolerance, constructive discourse, duty, and care are represented through polarization measures, civil-conversation goals, personal responsibility, and empathy measures.The passages also describe care in terms of context, interdependence, relationships, and responsibilities to others.Source: Appendix A · p. 45

Fairness, Equality, Equity

  • Fairness, equality, and equity are framed as just distributions of benefits, harms, and risks, with recommender fairness metrics and training or serving objectives as candidate approaches.The table notes that stakeholders can have different objectives and needs from a system.Source: Appendix A · pp. 45–46

Diversity

  • Diversity includes varied sources and cultural expression, with slate and feed metrics, methods for increasing topical and source diversity, and perceived-diversity and satisfaction measures.The listed methods also include novelty and coverage.Source: Appendix A · p. 46

Accountability; Transparency & Explainability

  • Accountability and transparency are associated with corporate accountability indices, explanations, audits, user problem-resolution processes, system cards, and disclosures about ranking or moderation changes.The table characterizes meaningful transparency as useful, accessible, comprehensive, and truthful, and explainability as tailored to users’ needs.Source: Appendix A · p. 46

Accuracy; Tradition, History; Environmental Sustainability

  • Accuracy and credibility can be assessed through misinformation prevalence and engagement, while tradition and sustainability are linked to local content, carbon footprints, and reporting model energy use.Recommenders are also noted as potentially directing users toward climate-friendly options.Source: Appendix A · pp. 46–47

Progress; Labor

  • Progress is linked to human development, while labor is framed through meaningful work and well-being measures and recommendations for jobs users would want.The passage notes that job insecurity, low employability, and unemployment can harm well-being beyond income loss.Source: Appendix A · p. 47
Loading 2207.10192v1…