CARE Principles: AI Recourse & Data Ethics

• CARE Principles are a set of guidelines uniting model interpretability and ethical data governance, emphasizing coherence, actionability, recourse, and soundness.
• They support multi-objective optimization in counterfactual explanations by balancing classifier flip, feature sparsity, and data manifold proximity.
• They also drive participatory data curation by granting community control and ethical oversight in low-resource language technologies.

The CARE Principles are a family of foundational guidelines that shape the development and evaluation of machine learning explanations, actionable recourse systems, and the governance of data—particularly for historically marginalized or under-resourced communities. They appear prominently in two distinct but complementary domains: (1) the technical framing and optimization of counterfactual explanations in model interpretability, and (2) the ethical and sociotechnical governance of language data curation, especially in low-resource linguistic contexts.

1. Definitions and Motivations of CARE Principles

The CARE acronym encapsulates four core desiderata: Coherence, Actionability, Recourse (Validity), and Soundness in the context of counterfactual generation and actionable recourse (Rasouli et al., 2021); and Collective Benefit, Authority to Control, Responsibility, and Ethics for data governance, especially low-resource language technologies (Ubois, 11 Dec 2025). While their instantiations differ across application domains, each dimension defines criteria for trustworthy, community-aligned, and practical interventions in both technical and societal systems.

In counterfactual explanations:

• Coherence requires counterfactuals $x'$ to preserve the statistical or causal relationships among features, ensuring mutual consistency of altered and unaltered features.
• Actionability mandates $x'$ respect immutable features and domain/user constraints; only feasibly alterable attributes may change.
• Recourse (Validity) compels $x'$ to flip the classifier output to the target $y'$ with minimal, sparse feature changes.
• Soundness insists $x'$ be a plausible, in-distribution point strongly connected to the data manifold to avoid unrealistic, artifact-exploiting recommendations.

In data governance:

• Collective Benefit obliges corpus work to serve the speech community, not just external or commercial stakeholders.
• Authority to Control asserts community veto power over data sourcing, annotation, and licensing.
• Responsibility assigns all pipeline actors accountability for representativeness and social impacts.
• Ethics demands concrete, ongoing ethical reflexivity at each operational stage.

The motivation for embedding these principles—either in model interpretability pipelines or corpus curation—is to preclude unimplementable, unethical, or unrepresentative outcomes by integrating contextual constraints and community priorities from the outset.

2. Formalization in Counterfactual Explanation: Multi-Objective Optimization

In model-agnostic counterfactual generation, the CARE framework formalizes these principles as a multi-objective optimization problem over candidate recourse points $x'$:

• Minimize: $L_\text{outcome}(x')$ for classifier flip (hinge loss), $d(x, x')$ (Gower mix-type distance), $\|\Delta(x, x')\|_0$ (feature sparsity), $L_\text{coherency}(x, x')$ (coherency cost via data-driven regression scores), and $L_\text{action}(x, x')$ (user/domain constraint violations).
• Maximize: $\operatorname{prox}(x')$ (negative Local Outlier Factor, i.e., data manifold proximity) and $\operatorname{conn}(x')$ (HDBSCAN high-density connectivity indicator).

The optimization is vectorized as: $$\min_{x'} \left[L_\text{outcome}, d, \|\Delta\|_0, L_\text{coherency}, L_\text{action}\right],\quad \max_{x'} \left[\operatorname{prox}, \operatorname{conn}\right]$$ subject to $x' \in \mathcal{X}$ (with numerical/categorical bounds).

The mapping is as follows:

Principle	Optimization Term(s)	Operational Objective
Recourse (Validity)	$L_\text{outcome} + d + \|\Delta\|_0$	Ensure classification flip with proximity and sparsity
Soundness	$\max \operatorname{prox},\operatorname{conn}$	Guarantee data manifold inlier status
Coherence	$\min L_\text{coherency}$	Preserve inter-feature dependencies
Actionability	$\min L_\text{action}$	Enforce user/domain constraints

This formalism enables compatibility with black-box models across tabular classification and regression tasks (Rasouli et al., 2021).

3. Modular Architectures and Solution Strategies

The actionable recourse pipeline grounded in CARE is modular, organizing computation into four pluggable modules:

1. Validity (always active): Enforces classification flip, data closeness, and sparsity.
2. Soundness (optional): Imposes manifold constraints through class-specific LOF/HDBSCAN clustering.
3. Coherency (optional): Extracts pairwise statistical dependencies from training data to restrict incoherent feature edits.
4. Actionability (optional): Codifies domain/user constraints over feature changeability and ranges.

Switching modules on/off recovers specialist prior methods and enables systematic benchmarking.

Optimization is carried out via the NSGA-III evolutionary algorithm:

• Handles mixed-scalar objectives (distance, indicator functions).
• Produces a diverse Pareto front of recourse options.
• Employs cross-over (≈60%) and polynomial mutation (≈30%), typically over ≈10 generations, with adaptively chosen population size.

At explanation time, the architecture queries only on $f(x')$ and performs efficient, lightweight scoring—never requiring model internals.

4. Application in Low-Resource Language Data Governance

In the context of language technology for low-resource languages (LRLs), the CARE principles (here: Collective Benefit, Authority to Control, Responsibility, and Ethics) have been instantiated within the "Data Care" framework (Ubois, 11 Dec 2025):

• Collective Benefit: Corpus work (e.g., ComText SR) is guided by community steering groups to align with educational, cultural, and minority-language stakeholder needs.
• Authority to Control: Data-licensing protocols require explicit opt-in by source-rightsholders, as exemplified by the TESLA team's workflow, contrasting with non-consensual scraping.
• Responsibility: Annotation is iteratively improved in joint workshops; every correction and bias audit is publicly tracked, versioned, and documented.
• Ethics: Community review boards vet new projects for consent, privacy, and dual-use risks, producing enforceable "Ethics Charters" governing subsequent model use.

The operational pipeline encompasses:

1. Stakeholder Mapping: Identify affected parties and domain experts.
2. Collective Benefit Workshops: Elicit, document, and prioritize community-desired applications.
3. Authority to Control Protocols: Conduct rights audits (ownership, licensing, permissions) and log them in machine-readable tables.
4. Responsibility-Driven Annotation: Pilot annotation, coverage metrics, and structured mid-cycle bias audits.
5. Ethics Review Gate: Mandatory risk assessment, ethics approval prior to model training.
6. Documentation and Release: Publish comprehensive datasheets, rights tables, and decision logs.

This methodology is operationalized as a repeatable workflow (e.g., within Snakemake or Nextflow), ensuring each CARE component is executed and documented.

5. Empirical Performance, Impact, and Limitations

In counterfactual frameworks (Rasouli et al., 2021), experimental results on standard tabular datasets (Adult Income, COMPAS, Default of Credit Card Clients, Boston House Prices) and models (neural networks, gradient boosting) showed:

• Near 100% proximity and connectedness to the data manifold (vs. 60–70% for baselines).
• Zero coherency violations (vs. up to 14% for baselines).
• Competitive distance and sparsity metrics (minimal loss in simplicity or feasibility).
• High diversity of explanations ($d_F \approx 0.8$, $d_v \approx 0.6$).
• Runtime per instance (0.1–0.3 s) matching or exceeding prominent baselines such as DiCE and CFPrototype.

For language data curation (Ubois, 11 Dec 2025), a plausible implication is that up-front CARE alignment produces more accurate, culturally authentic, and community-trusted resources—particularly critical in post-colonial or digitally marginalized linguistic contexts. However, CARE-based methodologies are resource-intensive, often burdened by legal, institutional, and scalability barriers, especially in jurisdictions lacking copyright exceptions for AI.

6. Broader Implications and Recommendations

The transposition of CARE principles across technical and sociotechnical boundaries demonstrates their utility as integrating frameworks for both model interpretability and responsible data stewardship. In each case, the embedding of CARE up-front, rather than as posterior correction, reframes bias and illegitimacy as problems of design and governance, not mere artifacts of algorithmic deployment.

Recommendations for practitioners include:

• Initiate all recourse and data governance projects with explicit CARE-aligned planning.
• Institutionalize rights management, bias auditing, and ethics review as continuous, documented stages.
• Leverage lightweight, automatable tooling for compliance (e.g., Rights Tables, datasheets, workflow checklists).
• Engage affected communities in both technical and policy oversight.

The Serbian case in particular underlines that CARE is essential—not an optional luxury—for inclusive, sustainable, and trustful language technology development (Ubois, 11 Dec 2025).

7. Comparative Table: CARE Principles Across Domains

Domain	CARE Principle(s)	Operationalization
Counterfactual AI	Coherence, Actionability, Recourse (Validity), Soundness	Multi-objective loss, modular pipeline, model-agnostic
Data Governance	Collective Benefit, Authority to Control, Responsibility, Ethics	Participatory curation, rights management, ethical review

This cross-domain instantiation highlights the adaptability and foundational status of CARE as a guiding paradigm for both rigorous technical systems and ethical data governance.