ERS: Ethical Risk Scoring System

• Ethical Risk Scoring (ERS) System is a methodology that quantifies and aggregates ethical risks in automated systems using explicit principles, scenario attributes, and domain-specific metrics.
• ERS frameworks employ mathematical models such as composite sums, fuzzy logic rules, and MIP optimizations to derive interpretable, repeatable risk scores across diverse domains.
• Empirical evaluations in criminal justice, AI alignment, and governance demonstrate ERS’s potential to guide risk mitigation and ethical decision-making in high-stakes environments.

An Ethical Risk Scoring (ERS) System is a formal methodology for quantifying, aggregating, and comparing the ethical risks associated with automated systems—AI, ML models, data collection workflows, or decision-support infrastructures. ERS systems operationalize the otherwise qualitative and theoretically diverse landscape of ethical concerns, structuring them into interpretable, repeatable, and auditable numerical scores based on explicit principles, scenario attributes, and domain-specific metrics. Modern ERS frameworks span criminal justice interventions, AI and LLM alignment, organizational and governance risk management, and adversarial harm quantification, employing causal inference, fuzzy logic, social welfare functions, multi-theory consensus, and expansive harm taxonomies (Barabas et al., 2017, Dyoub et al., 2 Jul 2025, Zeng et al., 2024, Khan et al., 24 Jan 2026, Khan et al., 23 Jan 2026).

1. Theoretical Foundations and Purpose

ERS systems are motivated by the need to render ethical and societal risks quantitatively tractable. Central constructs include:

• Probability × Impact framing: Risk is often modeled as the likelihood of an ethically negative event multiplied by its potential severity (impact) (Felländer et al., 2021).
• Multi-theory integration: Recent ERS frameworks synthesize ethical imperatives from utilitarianism, deontology, virtue ethics, care ethics, rights-based and contract theories, Rawlsian justice, environmental and pragmatic lenses (Khan et al., 24 Jan 2026, Khan et al., 23 Jan 2026).
• Modality and domains: Systems are tailored to specific application areas (criminal justice, LLMs, data governance, adversarial AI), but share a common methodology of principled decomposition into risk dimensions or harm types.

A defining principle is that ethical risks are not reducible to fairness or bias detection alone, nor to accuracy metrics, but arise from the intersection of system design, data provenance, societal structures, and harm externalities (Barabas et al., 2017, Felländer et al., 2021).

2. Taxonomies, Risk Dimensions, and Harm Typologies

ERS frameworks operationalize ethical risk by selecting, structuring, and weighting key risk dimensions:

Multi-level Harm Taxonomies

• HARM66+ taxonomy: Defines 66+ harm types, structured into Exo-Human (environmental, technological, infrastructural, corporate, and sociopolitical) and Endo-Human (physical, psychological, identity, social, legal, financial) categories, each with subtypes, and mapped to dominant ethical theories for cross-paradigm validity (Khan et al., 23 Jan 2026).
• Victim entity taxonomy: Classifies impacted entities as individuals, groups, institutions, non-human life, environmental systems, technological artifacts, or normative constructs.

Dimension Examples (Selected Frameworks)

Framework	Core Dimensions	Reference
DRESS-eAI	Legal, Societal, Governance	(Felländer et al., 2021)
Multi-ethical consensus	Harm Mitigation, Data Ownership, Subject Rights	(Khan et al., 24 Jan 2026)
AES/LLM ERS	Harm Probability, Severity (toxicity), Essay Quality	(Kim et al., 9 Jan 2026)
Fuzzy Framework (ff4ERA)	Physical harm, Autonomy violation, Trust loss, etc.	(Dyoub et al., 28 Jul 2025)
HARM66+	11 major categories, 66+ sub-types	(Khan et al., 23 Jan 2026)

Harm attributes are further formalized with normatively critical variables: irreversibility ($I_h$), duration ($D_h$), base severity ($S_0$), with composite harm weights defined as $w_h = S_0 \cdot [1 + \alpha I_h + \beta D_h ]$ (Khan et al., 23 Jan 2026).

3. Formal Scoring Models and Mathematical Aggregation

ERS systems mathematically encode ethical risk via explicit scoring formulas. Representative approaches include:

Composite and Weighted Sums

• In DRESS-eAI, normalized answers to scenario-specific questions ($x_i$) are mapped to subscores ($r_j$), then combined into composite legal, societal, and governance risk scores ($R_L$, $R_S$, $R_G$), and finally aggregated:

$$\mathrm{ERS} = \alpha R_L + \beta R_S + \gamma R_G \quad ; \quad \alpha + \beta + \gamma = 1$$

Rule-Based and Fuzzy Logic Systems

• Fuzzy ERS employs membership functions and fuzzy “if–then” rule sets, with risk aggregations:

$$R^* = \frac{\int_{0}^{100} r \cdot \mu_R(r)\, dr}{\int_{0}^{100} \mu_R(r)\, dr}$$

Validation and verification utilize fuzzy Petri nets and dynamic standard coverage (Dyoub et al., 2 Jul 2025).

• ff4ERA generalizes this by propagating certainty factors and analytic hierarchy weights:

$$\mathrm{ERS}_i = \mathrm{ERM}_i \times \mathrm{CF}_i \times w_i$$

where $\mathrm{ERM}_i$ is a defuzzified risk magnitude, $\mathrm{CF}_i$ the propagated certainty, and $w_i$ the FAHP-derived weight (Dyoub et al., 28 Jul 2025).

Social Welfare and Fairness-Aware Optimization

• Decision scorecards can be optimized with mixed-integer programming (MIP) frameworks, balancing predictive accuracy, group fairness constraints (e.g., statistical parity, equal opportunity), and interpretability:

$$\min_{w,\delta} \frac{1}{n} \sum_i b_i\,\psi_i(w) + \bar{\rho}\delta + \lambda_0 \Vert w \Vert_0 + \epsilon \Vert w \Vert_1 \quad \text{s.t.}\ G(w) \leq \delta$$

Knobs for cost weights $b_i$, fairness penalties $\rho_i$, and interpretability regularizers $\lambda_0$ enable tailoring to policy trade-offs (Yang et al., 2021).

Harm Instance Aggregation

• The HARM66+ scheme aggregates per-harm-instance scores:

$$\mathrm{ERS} = \frac{ \sum_{h \in H} w_d(d(h))\,w_c(c(h))\,w_h }{ \sum_{h \in H} w_d(d(h))\,w_c(c(h)) }$$

where $w_d$ (category), $w_c$ (domain), and $w_h$ (composite) weights embed normative consensus and stakeholder preferences (Khan et al., 23 Jan 2026).

4. Implementation, Validation, and Workflow

Implementation varies by context but follows a characteristic sequence:

1. Scoping and factor elicitation: Identify relevant ethical principles, dimensions, or harm types. Map phases/stakeholders/data flows to these dimensions (Khan et al., 24 Jan 2026, Felländer et al., 2021).
2. Questionnaire/survey administration: Structured surveys instantiate scenario-specific variables. For LLM data, for example, forty weighted questions span ethical sourcing, transparency, harm, and rights (dimensions S, T, H, R) (Khan et al., 24 Jan 2026).
3. Fuzzification/rule application: For fuzzy systems, measurement variables are mapped to linguistic terms with defined membership functions and aggregated via formal rule bases (Dyoub et al., 2 Jul 2025, Dyoub et al., 28 Jul 2025).
4. Scoring, aggregation, and normalization: Scores are computed as described above. Thresholds for actionable risk bands are either set by expert panel consensus, empirical calibration, or fixed analytically (Felländer et al., 2021, Khan et al., 24 Jan 2026).
5. Verification/validation: Techniques include dynamic validation (rule/test-case matching), structural verification via fuzzy Petri nets, sensitivity analysis (local and global), and recurring audit cycles (Dyoub et al., 2 Jul 2025, Dyoub et al., 28 Jul 2025).
6. Reporting and remediation: ERS results inform governance processes, access controls, system deployment gating, and targeted mitigations (Felländer et al., 2021).

5. Fairness, Proxy Bias, and Calibration

Ensuring the ethical integrity of ERS outputs demands multi-layered fairness constraints and proxy detection:

• Direct attribute exclusion and monitoring: Protected attributes (e.g., race, gender) must be excluded as direct covariates yet monitored for residual impact through fairness metrics such as demographic parity, equalized odds, and calibration-within-groups (Barabas et al., 2017).
• Proxy variable auditing: Mutual information between candidate covariates and protected attributes is computed; high-MI variables are orthogonalized or dropped (Barabas et al., 2017).
• Fairness-aware optimization: Scorecard frameworks integrate fairness directly into the optimization objective, enabling explicit tuning of fairness-utility trade-offs with clear parameter interpretation (Yang et al., 2021).
• Calibration and bias adjustment: For LLM risk propensity assessment, baseline subtraction and z-scoring facilitate comparability across personas and models, supporting systematic detection of role-based or group-based bias (Zeng et al., 2024).

6. Extensibility, Modularization, and Domain Adaptation

ERS frameworks are designed for modular adaptation:

• Taxonomic modularity: The HARM66+ harm hierarchy maintains stability in its top levels but enables incremental extension for emerging harms with rigorous assignment rules and versioning (Khan et al., 23 Jan 2026).
• Domain weighting and stakeholder customization: Category and domain weights can be tuned a priori (e.g., to prioritize environmental over social harm) or adjusted post hoc in light of organizational priorities or pilot data (Khan et al., 23 Jan 2026, Khan et al., 24 Jan 2026).
• Rule and survey expansion: New audit items or fuzzy rules can be appended as standards evolve, retaining backwards compatibility in scoring structures (Dyoub et al., 2 Jul 2025, Dyoub et al., 28 Jul 2025).
• Integration with enterprise/governance systems: ERS pipelines are deployed within risk management infrastructures for continuous reassessment and audit trail generation (Felländer et al., 2021).

7. Empirical Evaluations and Use-case Examples

Applied ERS systems have been piloted and evaluated in a range of settings:

• Criminal justice: Data-driven, intervention-oriented ERS cuts iatrogenic feedback loops, shifting focus from punitive prediction to causal, needs-based intervention allocation (Barabas et al., 2017).
• Organizational AI deployments: DRESS-eAI pilots in hiring and tax fraud detection identified actionable deficits in governance and data bias, enabling pre-deployment remediation and fostering cross-functional ethical oversight (Felländer et al., 2021).
• LLM risk profiling: DOSPERT/EDRAS batteries with role-play capture granular, cross-persona risk attitudinal differences in LLMs; ERS flagging supports both stability and bias detection (Zeng et al., 2024).
• Adversarial and security risks: The extensible HARM66+ taxonomy reconciles philosophical rigor with operational tractability, supporting high-resolution harm scoring in adversarial ML and cyber-physical domains (Khan et al., 23 Jan 2026).
• Essay scoring pipelines: ERS layers in harmful-content detectors and severity metrics to robustify quality scoring and flag LLM failure modes (Kim et al., 9 Jan 2026).
• Data pipeline audit: ERS is embedded in LLM data harnessing protocols to enforce source rights, transparency, harm reduction, and subject-rights compliance (Khan et al., 24 Jan 2026).

These empirical applications demonstrate that ERS methodologies, grounded in diverse but formally explicit ethical theories, can be tailored, verified, and deployed across high-stakes sociotechnical systems for principled risk anticipation, mitigation, and governance.