Source-linked AI summary
Causal Learning for Socially Responsible AI
Lu Cheng, Ahmadreza Mosallanezhad, Paras Sheth, Huan Liu
TL;DR
AI’s potential risks have increased interest in socially responsible AI, and this survey examines causal learning as one approach. It reviews causal tools and their applications to SRAI, concluding that causal learning offers an effective means of addressing these concerns while identifying open problems.
Problem
AI’s potential for oppression and calamity, alongside concerns about ethical challenges and undesirable outcomes, motivates approaches to socially responsible AI.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
The survey identifies causal discovery’s connection to socially responsible AI as an important future research area because causal graphs underpin many approaches in this field.
Source: §7 · p. 6Causal discovery’s connection to socially responsible AI is not well understood.
Source: §7 · p. 6
Abstract
from arXiv · showhide
There have been increasing concerns about Artificial Intelligence (AI) due to its unfathomable potential power. To make AI address ethical challenges and shun undesirable outcomes, researchers proposed to develop socially responsible AI (SRAI). One of these approaches is causal learning (CL). We survey state-of-the-art methods of CL for SRAI. We begin by examining the seven CL tools to enhance the social responsibility of AI, then review how existing works have succeeded using these tools to tackle issues in developing SRAI such as fairness. The goal of this survey is to bring forefront the potentials and promises of CL for SRAI.
1 Introduction
AI promises efficient human-like performance but also carries risks of oppression and calamity; the survey examines causal learning as an approach to socially responsible AI.
1 Introduction
- AI can improve daily life through efficient, precise performance, yet its workings are not fully understood and regulation remains limited.These concerns have brought socially responsible AI into the spotlight.Source: §1 · p. 1
- Without knowledge of the data-generating process, AI may inherit societal patterns, embed unexamined framing choices, and lack a basis for identifying causal relations.The passage identifies causal relations and cause-effect connections as ingredients for socially responsible AI.Source: §1 · p. 1
- The survey examines seven causal-learning tools, reviews four tools applied to bias mitigation, fairness, transparency, and generalizability or invariance, and identifies open problems.It frames causal learning as learning causality from causal knowledge and real-world data.Source: §1 · p. 1
2 Causal Tools for Socially Responsible AI
The section presents causal learning as a hierarchy of association, intervention, and counterfactual reasoning, then connects its tools to socially responsible AI.
2 Causal Tools for Socially Responsible AI
- The hierarchy moves from association, which captures statistical relations, to interventions on the data-generating process and counterfactual questions about alternate outcomes.Interventional distributions use do-calculus; in a causal graph, intervention removes incoming arrows to the changed variable.Source: §2 · p. 1
- Causal assumptions make AI systems more transparent and testable by letting researchers examine whether assumptions fit available data.Do-calculus can eliminate confounding through the back-door criterion, while counterfactual analysis asks what would have happened under a different condition.Source: §2 · pp. 1–2
- Mediation analysis separates an intervention’s direct and indirect effects, helping explain how a cause-effect relationship arises.The section links this explanatory role to generating explanations.Source: §2 · p. 2
- Adaptability concerns generalizing across environments, while causal graphs can help recover relationships when data are incomplete.Causal discovery seeks directionality from observational data and can provide a path toward true causal graphs.Source: §2 · p. 2
- The survey proposes a taxonomy and focuses on causal assumptions, do-calculus, counterfactual analysis, and adaptability in relation to bias mitigation, fairness, transparency, and generalizability or invariance.Source: §2 · p. 2
3 Bias Mitigation
The section reviews propensity scoring, counterfactual data augmentation, and causal graphs as approaches to bias mitigation, alongside practical constraints on their use.
3 Bias Mitigation
- The section situates bias mitigation as important to fair decisions and reviews propensity scoring and counterfactual data augmentation among several debiasing approaches.It identifies causal inference as one option alongside adversarial training and reinforcement learning.Source: §3 · p. 2
3.1 Propensity Score
- Inverse propensity scoring uses observation propensities to address selection and position bias in recommender systems.Randomized rankings can degrade user experience [Joachims et al., 2017]; an offline approach instead uses multiple feedback types to learn an unbiased ranking model [Guo et al., 2020b].Source: §3.1 · p. 2
3.2 Counterfactual Data Augmentation (CDA)
- Counterfactual data augmentation adds intervention-revised examples to training data to reduce spurious correlations and undesired bias.For gender-bias mitigation, one approach replaces gendered word pairs in sentence templates; its assumption is that matched pairs should receive the same outcome across genders [Lu et al., 2020b].Source: §3.2 · pp. 2–3
3.3 Causal Graphs
- A causal graph can separate user interest from item popularity to create a recommender based on users’ real interest rather than popularity-driven pseudo-interest.The cited approach creates separate embeddings for real interest and pseudo-interest caused by popularity.Source: §3.3 · p. 3, Figure 2 · p. 3
3.4 Discussions
- Propensity-based methods require correctly specified real-world scores, while randomized experiments may be impractical for ethical or financial reasons and may reduce utility.Designing interventions for CDA is also challenging, and causal graphs can rely on assumptions that may not reflect real-world biases.Source: §3.4 · p. 3
4 Fairness
The survey reviews causal approaches to fairness, framing fairness in terms of causal effects of sensitive attributes and examining counterfactual and intervention-based methods. It also highlights limits involving counterfactual validity and identifiability.
4 Fairness
- Fairness research using causal learning estimates the effects of sensitive attributes, such as gender, on AI outcomes and reviews causal fairness beyond de-biasing.The survey points to [Makhlouf et al., 2020] for a comprehensive review of causal fairness.Source: §4 · p. 3, §4.1 · p. 3
4.1 Fairness via Causal Modeling
- Counterfactual fairness requires an individual’s decision to remain the same in the actual and counterfactual worlds where their sensitive attribute differs.The definition compares outcome probabilities across sensitive-attribute values; its use can be constrained when counterfactual quantities are unidentifiable.Source: §4.1 · p. 3
- Intervention-based methods quantify discrimination through causal paths from sensitive attributes to outcomes, using average causal effects under post-intervention distributions.The survey describes methods by and that distinguish direct, indirect, or disallowed paths.Source: §4.1 · pp. 3–4
4.2 Discussions
- Counterfactual fairness is more general than intervention-based fairness but depends on manipulable sensitive attributes and counterfactuals whose validity and similarity are difficult to assess.The survey also notes that CF-based models may fail when the identifiable assumption is violated.Source: §4.2 · p. 4
5 Transparency
The survey presents causal interpretability as a way to produce human-friendly explanations for model decisions. It reviews model-based causal effects and counterfactual explanations, while noting evaluation and social-context challenges.
5 Transparency
- Causal interpretability aims to answer why a model made a decision, addressing concerns that opaque algorithms can erode user trust.The survey distinguishes interventional and counterfactual interpretability approaches and refers readers to.Source: §5 · p. 4
5.1 Model-based Interpretations
- Model-based interpretation maps a neural network to a structural causal model and uses do-calculus to estimate each component’s effect on the output.The model can be reduced by marginalizing hidden neurons to estimate the average causal effect of an input feature on an output [Chattopadhyay et al., 2019]; related approaches use CNNs, dependence plots, or latent-factor relations [Narendra et al., 2018; Zhao and Hastie, 2021; Mart´ınez and Marca, 2019].Source: §5.1 · p. 4
5.2 Counterfactual Explanations
- Counterfactual explanations describe how minimally changing input features could change a model’s predicted outcome to a predefined alternative.A generative framework uses attribute editing and balances closeness to the desired output against distance from the observed features; λ controls that balance.Source: §5.2 · p. 4
- Alternative counterfactual-generation approaches use adversarial examples with constrained feature changes or class prototypes to guide interpretable perturbations.Prototypes are mean encodings of instances in a class [Van Looveren and Klaise, 2019].Source: §5.2 · pp. 4–5
5.3 Discussions
- Evaluation is difficult without ground-truth causal relations, and counterfactual scenarios depend on contextual choices, including how social categories are defined.The survey stresses that causal assumptions should be stated clearly, even when ground truth and well-defined categories are available.Source: §5.3 · p. 5
6 Invariance/Generalizability
The survey reviews causal approaches to invariance and generalizability, motivated by spurious correlations and distribution shifts across environments. It covers causal-parent discovery, invariant representations, and counterfactual data augmentation, alongside their assumptions and limitations.
6 Invariance/Generalizability
- Data heterogeneity and shifts between training and test environments can undermine generalization, motivating invariant prediction based on stable causal relations.The survey notes that interventions change environments while direct causal relations remain invariant.Source: §6 · p. 5
6.1 Invariant Causal Prediction (ICP)
- ICP searches across environments for feature subsets with invariant coefficients or residuals, whose intersection is a subset of the target’s direct causes.It assumes unconfoundedness; a nonlinear extension was proposed for settings beyond linear models and discrete environment separation [Heinze-Deml et al., 2018].Source: §6.1 · p. 5
6.2 Invariant Risk Minimization (IRM)
- IRM learns a representation that predicts accurately while eliciting an invariant predictor across environments, without requiring recovery of the target’s direct causes.Prediction performance may decrease when spurious correlations are excluded, and IRM cannot guarantee their complete removal because it depends on the environments supplied.Source: §6.2 · p. 5
6.3 CDA for Invariance/Generalizability
- Causal data augmentation adds counterfactual examples to expose models to out-of-distribution scenarios and reduce learning of spurious training patterns.Applications include counterfactual images generated from shape, texture, and background mechanisms, visual question answering, and counterfactual experiences for reinforcement learning [Sauer and Geiger, 2021; Pitis et al., 2020].Source: §6.3 · pp. 5–6
6.4 Discussions
- Unconfoundedness is a stringent, often impractical assumption for invariant prediction; nonlinear IRM settings also lack formal latent-variable results and have underexplored risks.The survey notes ICP is more interpretable for discovering causal features, while CDA can relieve the training-data conditional-independence assumption; minimally different examples are an alternative when counterfactual generation is infeasible.Source: §6.4 · p. 6
7 Summary and Open Problems
The survey presents causal learning (CL) as an effective means of uncovering data-generating processes (DGPs) for socially responsible AI (SRAI), and outlines open directions involving privacy, causal discovery, and long-term impacts.
7 Summary and Open Problems
- CL is presented as an effective means of pursuing SRAI because it seeks to uncover DGPs, and as a component of AI systems addressing societal issues.The paper frames SRAI as protecting, informing, and preventing or mitigating AI harms.Source: §7 · p. 6
Privacy-preserving
- Using CL to enhance privacy remains barely studied, although causal methods could remove sensitive information and create privacy-preserving representations.The passage identifies this as a potential direction rather than an established result.Source: §7 · p. 6
Making explicit causal assumptions
- Explicit causal assumptions can make models more testable and transparent, but disputed assumptions call for sensitivity analysis and care against unethical choices.Researchers are urged to make assumptions with humility.Source: §7 · p. 6
Causal discovery
- Causal discovery’s connection to SRAI remains poorly understood, despite its role in checking assumptions and whether interventions are correctly applied.The passage identifies causal discovery as an important future research area because causal graphs underpin many SRAI approaches.Source: §7 · p. 6
Mediation analysis
- Causal mediation analysis can improve transparency by examining indirect pathways, such as whether job type mediates race’s effect on loan granting.The passage also proposes mediation analysis for tracing how neurons influence algorithmic decisions directly or indirectly.Source: §7 · p. 6
Missing data
- The paper proposes extending CL to broader missing-data settings, including graphical-model performance guarantees when data are Missing Not At Random.The cited example is Mohan and Pearl (2021).Source: §7 · p. 6
Long-term impact
- Most SRAI work overlooks long-term commitments, and static loan-fairness criteria may reduce minority groups’ credibility scores over time [Liu et al., 2018].The passage says this neglect hinders the efficiency and efficacy of existing work toward SRAI.Source: §7 · p. 6
Causal tools and libraries for SRAI
- Existing CL libraries include Causal ML, DoWhy, and Causal Discovery Toolbox, and CL models for SRAI could be integrated into them.The passage presents integration as a possibility.Source: §7 · p. 6