PaperlayerView exact revision on arXiv

Source-linked AI summary

Socially Responsible AI Algorithms: Issues, Purposes, and Challenges

Lu Cheng, Kush R. Varshney, Huan Liu

arXiv:2101.02032v5cs.CYcs.AI

TL;DR

Responsible-AI research has focused on fairness and bias in scoring and classification, leaving broader AI-human connections to examine. This survey proposes a systematic framework for socially responsible AI and concludes that current understanding remains insufficient.

  • Problem

    Research on responsible AI has focused on fairness and unwanted bias in scoring and classification, leaving a need to examine broader AI-human connections relevant to safety, privacy, and society.

    Source: §1 · p. 2

  • Method

    The survey proposes a systematic framework covering socially responsible AI’s subjects, causes, means, objectives, and roles in protecting and informing users and preventing negative impacts.

    Source: §1 · p. 2

  • Results

    The survey concludes that current understanding of socially responsible AI is insufficient and that future efforts are needed.

    Source: §5 · p. 27

  • Takeaways & Limitations

    The paper frames socially responsible AI as serving users through protection, information, and prevention of negative AI impacts.

    Source: §1 · p. 2

    Existing socially responsible AI interventions may be ineffective, inaccurate, or dangerously misguided when introduced into a new social context.

    Source: §5 · p. 28

Abstract

from arXiv · show

In the current era, people and society have grown increasingly reliant on artificial intelligence (AI) technologies. AI has the potential to drive us towards a future in which all of humanity flourishes. It also comes with substantial risks for oppression and calamity. Discussions about whether we should (re)trust AI have repeatedly emerged in recent years and in many quarters, including industry, academia, healthcare, services, and so on. Technologists and AI researchers have a responsibility to develop trustworthy AI systems. They have responded with great effort to design more responsible AI algorithms. However, existing technical solutions are narrow in scope and have been primarily directed towards algorithms for scoring or classification tasks, with an emphasis on fairness and unwanted bias. To build long-lasting trust between AI and human beings, we argue that the key is to think beyond algorithmic fairness and connect major aspects of AI that potentially cause AI's indifferent behavior. In this survey, we provide a systematic framework of Socially Responsible AI Algorithms that aims to examine the subjects of AI indifference and the need for socially responsible AI algorithms, define the objectives, and introduce the means by which we may achieve these objectives. We further discuss how to leverage this framework to improve societal well-being through protection, information, and prevention/mitigation.

1. Introduction

The survey argues that socially responsible AI requires a holistic view of how AI systems affect people and society, beyond algorithmic fairness. It defines AI responsibility and presents a framework for protecting and informing users and preventing or mitigating AI’s negative effects.

Introduction

  • AI’s broad sociotechnical reach brings both promise and peril; a report found 59% of tech workers had worked on products they felt harmed society, and over 25% of affected AI workers quit (Miller C, 2019).The introduction also notes public and worker activism over technology’s negative social impacts.Source: §1 · p. 1
  • The survey moves beyond fairness-focused work on scoring and classification to examine connections among AI aspects and their effects on safety, privacy, and society.It aims to complement surveys treating fairness, explainability, and privacy as separate topics with a holistic account of AI systems and humans.Source: §1 · p. 2
  • The authors define AI social responsibility through principles, means, and objectives, and organize its four responsibilities as functional, legal, ethical, and philanthropic.The pyramid is adapted from Carroll et al.’s (1991) Corporate Social Responsibility pyramid and encompasses efforts across disciplines.Source: §1 · p. 2
  • The SRAs framework covers subjects, causes, means, and objectives, and describes roles in protecting and informing users and preventing negative AI impacts.It subsumes existing topics such as fairness and interpretability.Source: §1 · p. 2
  • The survey identifies open problems and challenges, including understanding why AI systems are needed and defining new AI ethics principles and policies.The authors hope these discussions will prompt future SRA research.Source: §1 · p. 3

2. Social Responsibility of AI

This section frames Social Responsibility of AI as encompassing technical and societal concerns, defines its value-driven purpose, and organizes AI duties in a four-part pyramid.

2.1 What is Social Responsibility of AI?

  • Social Responsibility of AI is a human value-driven process that uses responsible algorithms to generate shared value by improving AI capability and societal benefit.Its guiding principles include fairness, transparency, accountability, reliability and safety, privacy and security, and inclusiveness.Source: §2.1 · p. 3

2.2 The Pyramid of Social Responsibility of AI

  • The proposed pyramid adapts Carroll’s CSR model into functional, legal, ethical, and philanthropic AI responsibilities.Functional competence undergirds the rest, and legal responsibilities require meeting at least minimal legal requirements. Ethical duties include doing what is right, just, and fair, and cannot be sacrificed for functional goals; philanthropy is desirable but not ethically required. The responsibilities may remain in dynamic tension.Source: §2.2 · pp. 3–5, Figure 1 · p. 4

2.3 Comparisons of Similar Concepts

  • Socially Responsible AI provides a systematic view that subsumes similar concepts and includes functional, legal, ethical, and philanthropic responsibilities.The survey focuses on ethical and philanthropic responsibilities, while noting that functional and legal responsibilities are usual focuses in AI research and development.Source: §2.3 · p. 5, Table 1 · p. 5

3. Socially Responsible AI Algorithms (SRAs)

The section defines Socially Responsible AI Algorithms (SRAs) as systems that prioritize all stakeholders—especially minoritized and disadvantaged users—to make just, trustworthy decisions. It presents a framework for their responsibilities, roles, and ongoing user feedback.

3. Socially Responsible AI Algorithms (SRAs)

  • AI researchers must develop accurate, reliable, trustworthy algorithms, but human biases and uncertainties in supervised-learning pipelines can lead to bias and discrimination.Examples include imbalanced data, ill-defined annotation criteria, oversimplified loss functions, and unexplainable results.Source: §3 · p. 6
  • SRAs prioritize all stakeholders, especially minoritized and disadvantaged users, to make just and trustworthy decisions while protecting and informing users, mitigating negative impacts, and maximizing long-term benefit.They continually receive user feedback to help accomplish expected social values.Source: §3 · p. 6
  • SRAs integrate functional objectives, such as profit, and societal objectives, such as transparency, while aiming to meet or exceed business objectives.Source: §3 · p. 6
  • The framework links SRAs’ internal mechanisms and ethical responsibilities with philanthropic and ethical roles, using end-user feedback to help achieve expected social values.The framework is illustrated in Figure 3.Source: §3 · p. 6, Figure 3 · p. 7

3.1 Subjects of Socially Indifferent AI Algorithms

Socially indifferent AI algorithms can harm anyone, but minorities and disadvantaged groups often face greater harms that vary by context. Individuals are also exposed when personal data are collected and used without meaningful consent or control.

3.1 Subjects of Socially Indifferent AI Algorithms

  • Minorities and disadvantaged groups often suffer the most frequent and severe harms, including racial bias in COMPAS risk labels (Angwin, Larson, Mattu, & Kirchner, 2016).Google mislabeled an image of two Black people as “gorillas”, and job-ad algorithms favored males.Source: §3.1 · p. 7
  • The subjects of harm depend on context: Princeton Review tutoring prices burdened higher-income areas and Asians, while poor people without internet may be missing from training data.Residents of higher-income areas were charged twice as much as the general public and people in high-density zip codes; Asians were 1.8 times more likely to pay higher prices regardless of income.Source: §3.1 · p. 7
  • Purpose-driven data collection and use can make every individual a subject of harm when personal data are used without consent or control, causing privacy leakage and distrust.Collected data can include contact information, browsing and session records, social media consumption, and location; few users control where, how, or by whom their data are used.Source: §3.1 · p. 8

3.2 Causes of Socially Indifferent AI Algorithms

The paper identifies formalization, measurement errors, bias, data misuse, and confusing correlation with causation as causes of socially indifferent AI algorithms (Mehrabi et al., 2019).

3.2.1 Formalization

  • Formalization can erase social context, rely on proxy labels, encode oversimplified objectives, and use benchmarks that steer models away from task goals.Loss functions focused only on profit or loss may select strategies stakeholders consider unethical; data annotation criteria and annotator consistency can also be problematic (Getoor, 2019; Beale, Battey, Davison, and MacKay, 2019).Source: §3.2.1 · p. 8

3.2.2 Measuring Errors

  • Non-representative training and test samples can make reported performance misleading, as later testing may reveal deterioration despite zero training error.The learned model may fail to represent the true model when the samples do not reflect the target population or have sufficiently similar distributions.Source: §3.2.2 · p. 8, Figure 4 · p. 9

3.2.3 Bias

  • The review identifies data bias, including heterogeneity and imbalanced subgroups, automation bias from favoring automated decisions over contradictory information, and algorithmic bias that can encode societal prejudices in data.High-dimensional text risks overfitting, while data heterogeneity challenges the i.i.d. assumption. Skynet extracted 80 features per user from cellular traffic in Pakistan with only 7 known terrorists and identified an Al Jazeera reporter covering Al Qaeda as a potential terrorist.Source: §3.2.3 · p. 9

3.2.4 Data Misuse

  • Data misuse without users’ consent, awareness, knowledge, or choice can devalue personal data and undermine privacy and trust.Examples include the Facebook-Cambridge Analytica scandal involving millions of users’ data and advertisers exploiting sensitive information for tailored campaigns (Caba˜nas, Cuevas, Arrate, & Cuevas, 2020).Source: §3.2.4 · p. 10

3.2.5 Correlation vs Causation

  • Correlation can aid prediction but not establish causation for decisions; confounders create spurious associations, and subgroup trends can reverse when aggregated (Simpson’s Paradox).At UC Berkeley, male applicants had a higher overall admission rate, while department-level rates for women were equal or higher (Bickel, Hammel, & O’Connell, 1975).Source: §3.2.5 · pp. 10–11

3.3 Objectives of Socially Responsible AI Algorithms

Socially Responsible AI Algorithms aim to rebuild trust in AI through three concrete objectives: fairness, transparency, and safety. These objectives address decision-making, understandable explanations, and safe, accurate, reliable system behavior.

3.3.1 Fairness

  • Fairness is the absence of prejudice or favoritism based on individual or group characteristics (Mehrabi et al., 2019), but remains difficult to define or guarantee in practice.A system fair in a specific context may still contribute to biased decisions because policymakers and the environment also shape decision-making. The six discrimination types are direct, indirect, systemic, statistical, and explainable or unexplainable discrimination, depending on whether group differences can be justified.Source: §3.3.1 · pp. 11–12

3.3.2 Transparency

  • Transparency means understandable explanations, with distinct goals for developers, deployers, and users and global explanations of systems versus local explanations of contextual decisions.Explanations may be hacked or manipulated, undermining trust; Slack et al. (2020) studied attacks on LIME and SHAP. Privacy and security are necessary conditions for pursuing transparency.Source: §3.3.2 · p. 12

3.3.3 Safety

  • Safety requires AI systems to be accurate, reliable, secure, and robust amid uncertainty, volatility, and flux (Varshney & Alemzadeh, 2017; Leslie, 2019).Accuracy requires an appropriate performance measure: precision, recall, and F1 may be more suitable than accuracy for extremely imbalanced classification data. Robustness means functioning reliably and accurately under adversarial attacks, perturbations, and implementation errors.Source: §3.3.3 · pp. 12–13

3.4 Means Towards Socially Responsible AI Algorithms

The section surveys four means for socially responsible AI algorithms: interpretability and explainability, adversarial machine learning, causal learning, and uncertainty quantification. It outlines their roles in transparency, robustness, and responsible decision-making.

3.4.1 Interpretability and Explainability

  • Interpretability methods make AI decisions more transparent, especially in high-stakes applications, and can be applied before, during, or after model training.Pre-model methods examine data; in-model methods use intrinsically interpretable models or constraints; post-model methods explain trained models. Explanations should balance audience comprehension with correctness and have expressive power, translucency, portability, and manageable complexity.Source: §3.4.1 · pp. 13–14

3.4.2 Adversarial Machine Learning

  • Adversarial machine learning addresses model robustness against imperceptible crafted examples and attacks that exploit an AI system’s threat surface.Attacks include evasion at testing, poisoning during training, and exploratory probing; adversary capabilities range from data injection and modification to logic corruption, and from white-box to black-box access. Examples include adversarial-example generation and GAN-based attacks.Source: §3.4.2 · pp. 14–15

3.4.3 Causal Learning

  • Causal learning underpins socially responsible AI and supports fairness, discrimination reduction, and bias alleviation across applications.Structural causal models and potential-outcome frameworks ground causal effect estimation and discovery; propensity-score weighting is a commonly used approach. Applications include debiasing recommendation systems and word embeddings.Source: §3.4.3 · pp. 15–16
  • Causal methods can explain model-component effects and provide human-friendly counterfactual explanations, while causal relationships are expected to remain invariant and robust across environments.Counterfactual explanations may have multiple valid versions, the Rashomon effect; causal approaches also complement transfer learning, domain adaptation, and lifelong learning.Source: §3.4.3 · pp. 16–17

3.4.4 Uncertainty Quantification

  • Uncertainty quantification supports transparency and decision-making because deployed AI encounters samples unlike its training data, with epistemic uncertainty harder to quantify than aleatory uncertainty.Prediction intervals include multi-step methods such as Bayesian and bootstrap methods, and direct methods such as lower-upper-bound estimation, which can handle arbitrary target distributions with reduced computation; direct interval forecasting further narrows intervals and shortens computation time.Source: §3.4.4 · pp. 17–18

4. Roles of SRAs

SRAs should fulfill ethical responsibilities before their philanthropic aims, since benefits may be insignificant when ethical duties fail. Their roles span protecting, informing, and preventing or mitigating harms to improve human life.

Roles of SRAs

  • SRAs must ensure ethical responsibilities first, because philanthropic benefits may be insignificant when AI fails to meet its ethical duties.The paper argues that the roles of SRAs encompass both ethical and philanthropic responsibilities.Source: §4 · p. 19
  • SRAs can improve human life by protecting users’ personal information, informing through fake-news early detection, and preventing or mitigating harms such as cyberbullying.The survey focuses on privacy-preserving methods and data dignity for protection, and on disinformation, abusive language, and unwanted bias for informing and preventing or mitigating.Source: §4 · p. 19
  • The survey uses cyberbullying as a representative form of abusive language because related work across its many forms is vast, diverse, and multidisciplinary.Examples of other forms include hate speech and profanity.Source: §4 · p. 19

4.1 Protecting

The protecting dimension aims to shield people—especially those most vulnerable or at risk—from harm caused by AI. It focuses on privacy-preserving methods and data dignity, which extends protection to users’ control over their data.

4.1 Protecting

  • The protecting dimension seeks to shield humans, especially vulnerable or at-risk people, from harm, injury, and negative impacts of AI systems.It includes protecting personal data and users’ interactions with AI systems.Source: §4.1 · p. 19

4.1.1 Privacy-Preserving

  • Privacy threats include direct breaches and inferred exposure through model access or outputs; defenses are grouped into private data aggregation, private training, and private inference.Sensitive-data use raises concerns about illegitimate use and disclosure (Boulemtafes, Derhab, & Challal, 2020; Beigi & Liu, 2020).Source: §4.1.1 · pp. 19–20
  • Aggregation methods may be context-free, such as differential privacy, or context-aware, such as information-theoretic privacy (Schaefer, Boche, Khisti, & Poor, 2017).Removing identifiers has proved unsatisfying (Sweeney, 2002; Narayanan & Shmatikov, 2008; Homer, Szelinger, Redman, Duggan, Tembe, Muehling, Pearson, Stephan, Nelson, & Craig, 2008); k-anonymity seeks to prevent re-identification.Source: §4.1.1 · p. 20
  • Private training commonly uses differential privacy, semantic security, or encryption, while private inference adapts similar methods for deployed inference-as-a-service systems.Encryption methods include homomorphic encryption and secure multi-party computation; inference methods also include information-theoretic privacy, differential privacy, and these encryption approaches.Source: §4.1.1 · p. 20

4.1.2 Data Dignity

  • Data dignity goes beyond privacy by giving users control over data use and compensation, including rights to assess risks and benefits, set purposes, negotiate terms, and be found, analyzed, or forgotten.These aspects give users autonomy over their data and its use.Source: §4.1.2 · pp. 20–21

4.2 Informing

The informing dimension aims to deliver information to users promptly, especially potential negative outcomes, using disinformation, cyberbullying, and bias detection as examples.

4.2.1 Disinformation Detection

  • Disinformation detection is difficult because it spans domains, changes rapidly, and requires interdisciplinary work; approaches use content, social context, comments, fact-checking, explanations, and cross-domain methods.Comments can provide weak supervision for early detection, while knowledge bases can support automatic fact-checking; research also addresses causal understanding of fake-news dissemination.Source: §4.2.1 · pp. 21–22

4.2.2 Cyberbullying Detection

  • Cyberbullying detection must model repeated aggression, power imbalance, and context, not just abusive content; methods draw on content, sentiment, user, network, and temporal features.Textual features are reported as the most informative predictors; challenges include defining bullying characteristics, annotation, and severe class imbalance.Source: §4.2.2 · pp. 22–23

4.2.3 Bias Detection

  • Bias detection is difficult when decision data are inaccessible and algorithms opaque; statistical disparity can flag possible discrimination but does not establish it.Methods include exploratory data analysis, the 80% selection-rate rule, regression, outcome tests, benchmarking, and automated tools; regression findings depend on correct specification, and outcome tests address unobserved decision factors.Source: §4.2.3 · p. 23

4.3 Preventing/Mitigating

When earlier protections fail, socially responsible AI must prevent or mitigate harms from disinformation, cyberbullying, and bias. The section outlines interventions ranging from information controls and social support to technical fairness methods and broader accountability measures.

4.3.1 Disinformation Prevention/Mitigation

  • Disinformation mitigation can govern news visibility through recommendation and ranking systems and use early detection to limit its spread and harms.Because engagement-driven systems can increase the visibility of popular posts and trending content, governing that visibility is one proposed approach.Source: §4.3.1 · p. 24
  • Disinformation interventions include counter-cascades, content flagging, and behavior-based identification of susceptible users.Counter-cascades aim to inoculate network nodes against false news; user flagging can draw on crowd judgment, while behavior patterns can help identify vulnerable groups.Source: §4.3.1 · p. 24

4.3.2 Cyberbullying Prevention/Mitigation

  • Cyberbullying prevention combines technological controls, victim support, education, and administrative policy, requiring interdisciplinary collaboration.Examples include parental controls, counseling, victim-support apps, and efforts to educate both victims and bullies.Source: §4.3.2 · p. 25
  • Cyberbullying strategies often fall short because adolescents’ online bullying behavior is complex and nuanced.The passage notes that prevention efforts remain comparatively limited despite extensive work on cyberbullying detection.Source: §4.3.2 · p. 25

4.3.3 Bias Mitigation

  • Bias mitigation uses pre-processing, in-processing, or post-processing, depending on whether training data or the model can be modified.Pre-processing transforms training data; in-processing modifies training to incorporate fairness objectives or constraints; post-processing reassigns predictions using a defined function and holdout set.Source: §4.3.3 · p. 25
  • Fairness methods address protected groups through individual, group, or subgroup measures, with classification and regression adapting fairness definitions and metrics to their targets.Counterfactual fairness asks whether an individual would receive the same decision under a different sensitive attribute; generalized fairness quantification remains challenging.Source: §4.3.3 · p. 26
  • Beyond fair machine learning, bias mitigation calls for transparency, bias-impact statements, diverse design, cultural sensitivity, and consideration of civil-rights protections.Sharing enough information to support replication can help users understand algorithms; impact statements can probe potential bias, while proposed legal clarification includes thresholds for disparate treatment.Source: §4.3.3 · pp. 26–27

5. Open Problems and Challenges

The survey finds that socially responsible AI remains insufficiently understood and identifies challenges spanning causal and contextual design, governance, defenses, ethics, human involvement, long-term responsibility, industry practice, and social good.

Causal Learning

  • Causal learning can address robustness, explainability, and cause-effect obstacles while helping AI understand “why” and transfer knowledge across similar domains.Early SRA efforts applied interventions, counterfactuals, do-calculus, and propensity scoring to fairness and interpretability.Source: §5 · p. 27

Context Matters

  • Social and algorithmic contexts shape whether SRA interventions work, making context-aware design essential to avoid ineffective or misguided systems.Fair-ranking effectiveness can depend on job contexts, while subtle algorithmic choices can amplify feedback loops in predictive policing.Source: §5 · p. 28

Responsible Model Release and Governance

  • Responsible model release and governance should document intended uses, system characteristics, evaluations, ethical considerations, caveats, and recommendations across the AI life cycle.The survey calls for effective release documentation and governance tools to increase transparency, manage risk, and build trust.Source: §5 · p. 28

AI Defenses

  • AI defenses should make systems provably robust by probing for weaknesses and automatically intercepting potentially problematic inputs.The survey notes that AI defenses remain early-stage because systems are vulnerable to attack and can fail in surprising ways.Source: §5 · pp. 28–29

AI Ethics Principles and Policies

  • Ethical AI principles need clearer operational guidance and broader input from minority and disadvantaged groups, rather than being shaped mainly by powerful actors and researchers.One proposal grounds principles in autonomy, beneficence, and justice to help organizations navigate value trade-offs.Source: §5 · p. 29

Understanding Why; Humans in the Loop

  • Clarifying AI’s intended goals and involving humans to contextualize high-stakes objectives can reduce socially indifferent failures, including when subgroup fairness thresholds differ.Human judgment can help calibrate different cutoffs for two subgroups.Source: §5 · pp. 29–30

Long-Term Effect

  • Long-term responsibility is difficult because social values and fairness notions evolve, and feedback loops can undermine even static fairness criteria.The survey identifies measurement and temporal modeling as important to evaluating fairness over time.Source: §5 · p. 29

Responsible AI Gap in Industry; Interdisciplinary Research; SRAs for Social Good

  • Bridging industry’s Responsible AI Gap requires sustained organizational action and interdisciplinary, decentralized research beyond a narrow set of fields.Social-good applications can support protection, information, prevention, fundraising, and greenlighting grants or other resources.Source: §5 · pp. 30–31

6. Conclusion

The survey broadens Social Responsibility of AI beyond scoring and classification, proposing a framework that connects AI’s functional and societal aspects to responsibilities and human well-being. It argues that doing good alone is insufficient and calls for further work on socially responsible AI.

6. Conclusion

  • The survey broadens the scope of Social Responsibility of AI and defines socially responsible algorithms as integrating functional and societal aspects.It frames AI responsibility across functional, legal, ethical, and philanthropic levels, with improving humanity as an objective.Source: §6 · p. 31
  • The framework investigates ethical responsibility through its subjects, causes, objectives, and means, while assigning AI roles in protecting and informing users and preventing or mitigating harm.The survey highlights philanthropic and ethical responsibilities as necessary for these intended uses.Source: §6 · p. 31
  • Drawing on moral license, the survey argues that simply asking AI to do good is insufficient and calls for further research on open problems and challenges.It urges AI technologists and researchers to develop socially responsible systems and steer AI toward a beneficial future.Source: §6 · p. 31

Related papers

Loading 2101.02032v5…