Abstract

The rapid integration of algorithmic systems into high-stakes decision-making environments demands architectural guarantees that transcend raw statistical accuracy on sterile benchmarks. While policy institutions have established high-level normative requirements for AI governance, these guidelines are frequently presented as abstract regulatory objectives rather than operational engineering specifications. To bridge this gap, this paper proposes a comprehensive taxonomic framework that systematically decomposes the European Union’s mandates for Trustworthy AI into 21 distinct, measurable, and actionable design properties. These properties are structurally organized across three interdependent layers: the technical dimension (focusing on internal algorithmic integrity and mathematics), the socio-technical dimension (addressing human-computer interaction workflows), and the social dimension (aligning system deployment with legal compliance and ethical values).

A central claim of this framework is that computational efficiency and algorithmic observability are not peripheral performance conveniences, but indispensable structural pillars; without deterministic efficiency bounds, technical safety monitors collapse under production workloads, and without runtime observability telemetry, complex architectures remain unauditable black boxes. Crucial boundaries are formalized herein to resolve persistent conceptual overlaps, explicitly decoupling interpretability from user-facing explainability, and distinguishing the live engineering monitoring of observability from the forensic, long-term legal traceability of accountability. Finally, these dimensions are unified into a functional matrix that maps normative mandates directly to software architecture components. By transitioning from theoretical compliance to a strict engineering taxonomy, this framework provides a rigorous blueprint to ensure that algorithms deployed in contested environments remain structurally, interactionally, and socially worthy of trust under real-world duress.

1. Introduction

Algorithms are no longer confined to textbook exercises or isolated software components; they increasingly participate in decisions that affect access to loans, medical care, educational opportunities, employment, and the operational loops of cyber-physical systems. Search ranking, recommender systems, intrusion detection models, decision-support tools, and automated triage systems now influence both individual outcomes and institutional processes at scale. As a result, the relevant question is not simply whether algorithms are mathematically accurate on sterile benchmark datasets, but whether they can be trusted in contested environments where errors, opacity, or adversarial misuse carry substantial human and societal costs (Spiegelhalter, 2020).

This concern is driven by repeated evidence that many deployed systems remain vulnerable to adversarial manipulation, unfair treatment of underrepresented groups, privacy leakage, and catastrophic failures under shifting real-world distributions. These shortcomings do more than reduce technical performance; they undermine public confidence and expose a structural gap between what systems achieve in controlled laboratory evaluations and what they can justify in high-stakes deployment. In response, policy institutions, academic researchers, and industrial actors have increasingly converged on the language of trustworthy AI to organize requirements that transcend raw prediction quality (Li et al., 2023).

Within that broader landscape, the expression trustworthy algorithms (Grimmelikhuijsen, 2023) can be useful because it foregrounds a rigorous engineering question: what architectural properties and mathematical guarantees must a system possess before individuals or regulators are justified in relying on it? The answer is not exhausted by user perception. A system may be blindly trusted by users and still be unworthy of trust—operating as an epistemic black box—just as a system may be technically capable in narrow settings while remaining unsuitable for broader deployment because it is fragile, opaque, or impossible to audit dynamically.

This work develops a comprehensive taxonomic framework. It draws on the literature on trustworthy artificial intelligence but maintains a strict focus on algorithmic systems as engineered artifacts operating under resource constraints, strict computational environments, and legal-ethical expectations. In addition, efficiency and algorithmic observability are indispensable pillars: if an algorithm is prohibitively resource-intensive, it cannot meet latency or deployment constraints required for reliable operation; if it lacks the active runtime telemetry of observability, the system remains an unauditable black box. Without these structural properties, theoretical compliance dissolves into systemic vulnerability when systems transition from laboratory constraints to high-throughput production environments.

The remainder of this paper is organized as follows. Section 2 establishes the conceptual transition from psychological states of trust to engineered trustworthy algorithms, framing our work within the foundational requirements of the European Union. Section 3 introduces our three-dimensional mapping methodology, providing a holistic overview of the technical, socio-technical, and social layers. Sections 4, 5, and 6 deliver the core formalizations of the taxonomy, systematically defining the 10 structural properties of the technical dimension, the 6 operational vectors of the socio-technical layer, and the 5 institutional dynamics of the social dimension, respectively. Section 7 synthesizes these 21 building blocks into a functional architecture matrix, highlighting cross-cutting dynamics and alignment with normative mandates. Finally, Section 8 provides concluding remarks on the implications of this taxonomic framework for software engineering and algorithmic governance.

2. From trust in algorithms to trustworthy algorithms

Empirical research on algorithm aversion shows that people’s willingness to rely on algorithmic outputs can be lower than warranted by their actual performance, and that small observed errors can trigger a disproportionate loss of trust in otherwise more accurate systems (Dietvorst et al., 2015). At the same time, studies on trust in automation indicate that users may also over‑trust models because they project an aura of mathematical objectivity (Hoff and Bashir, 2015). Therefore, stronger formulations in the human–automation literature differentiate between systems that are trusted (a psychological state in the user) and systems that are trustworthy (an engineered property of the system’s reliability, predictability, and performance).

The European Commission’s ethics guidelines frame this through three broad pillars: AI should be lawful, ethical, and robust, both from a technical and a social perspective. In addition, from the ethical principles, the guidelines derives seven specific requirements that AI systems must take into account in order to put those principles into practice (European Commission, 2019):

• (1) Human agency and oversight, including fundamental rights, human agency and human oversight.
• (2) Technical robustness and safety, including resilience to attack and security, fall back plan and general safety, accuracy, reliability and reproducibility.
• (3) Privacy and data governance, including respect for privacy, quality and integrity of data, and access to data.
• (4) Transparency, including traceability, explainability and communication.
• (5) Diversity, non-discrimination and fairness, including the avoidance of unfair bias, accessibility and universal design, and stakeholder participation.
• (6) Societal and environmental wellbeing, including sustainability and environmental friendliness, social impact, society and democracy.
• (7) Accountability, including auditability, minimisation and reporting of negative impact, trade-offs and redress.

The term algorithm is used here in an intentionally broad sense. In contemporary practice, a deployed “algorithm” is rarely a single isolated procedure; it is often part of a pipeline that includes data collection, preprocessing, optimization, model inference, ranking, thresholding, logging, and human review. Accordingly, the trustworthiness of an algorithm should be understood as the trustworthiness of an algorithmic system, including its interfaces, assumptions, operating conditions, and governance arrangements.

A precise operational definition is therefore the following: trustworthy algorithms are algorithmic systems whose architectural design, socio-technical operation, and institutional governance justify calibrated reliance because they satisfy a comprehensive matrix of technical, interactional, and societal properties. As will be detailed in the subsequent section, this framework systematically operationalizes these properties across three interdependent layers: the technical, socio-technical, and social dimensions. Within this paradigm, attributes such as efficiency and governance are not treated as merely contextual conveniences or secondary performance metrics. Instead, they are established as foundational operational conditions required to ensure that all structural guarantees—including robustness, safety, and privacy—remain resilient, auditable, and scalable when the system transitions from isolated laboratory conditions to complex, high-throughput production environments.

3. Dimensional mapping

While the European Union’s requirements provide a robust normative foundation for Trustworthy AI (Smuha, 2019), applying these principles in complex computing environments requires a layered approach. Thus, evaluating algorithmic trustworthiness purely through statistical or architectural abstractions risks falling into what Selbst et al. (2019) define as sociotechnical traps. To mitigate this, our methodology expands the scope of evaluation into three distinct dimensions, ensuring that technical metrics (technical dimension) are explicitly mapped against human-computer interaction constraints (socio-technical dimension) and broader regulatory and social systems (social dimension):

• Technical dimension: Focuses on the system’s internal integrity, where requirements like robustness and security are codified into the architecture. This layer ensures that the system behaves as intended under specific conditions.
• Socio-technical dimension: Addresses the interaction between the system and its human operators. Here, we translate abstract requirements—such as transparency—into actionable properties like explainability and observability, enabling effective human intervention and oversight.
• Social dimension: Extends the scope beyond the machine, aligning the system’s design with societal values, legal compliance, and institutional governance. This layer ensures that the system is not only functional but also accountable, fair, and contestable within the broader societal framework.

By decoupling these layers, our methodology allows for a systematic assessment of trustworthiness, ensuring that engineering constraints are not just theoretical, but foundational pillars integrated from problem formulation through to final deployment.

Thus, while the European Ethics Guidelines for Trustworthy AI establish a comprehensive set of 7 high-level requirements for AI governance, these principles are often presented as regulatory objectives rather than operational specifications. To bridge this gap, our proposal systematically decomposes these normative requirements into actionable design properties. By mapping each European requirement to specific technical, socio-technical, and social properties, we transition from theoretical compliance to a concrete engineering methodology, allowing these properties to be embedded, monitored, and audited directly within the system’s architecture.

3.1 Technical dimension

Robustness, security, and resilience: The European Union’s requirement for Technical Robustness and Safety demands that computing systems behave reliably while minimizing and preventing unintentional harm. To operationalize this for granular architectural evaluation, we decompose this requirement into three distinct measurable properties. Security ensures the system incorporates proactive defenses against adversarial threats such as data poisoning and model evasion. Resilience measures the system’s fault tolerance and its capability to maintain essential functionalities under degraded or hostile conditions. Robustness guarantees continuous accuracy, reproducibility, and stable algorithmic behavior when confronted with novel, noisy, or unforeseen input data.

Safety and reversibility: Following the ALTAI (The Assessment List for Trustworthy Artificial Intelligence) mandate to establish structural risk metrics and safety criticality levels, safety is defined as the architectural guarantee against physical, mental, or operational harm. Complementarily, reversibility operationalizes the EU’s requirement for tested, failsafe fallback plans, ensuring the system can automatically roll back state alterations, restore baseline configurations, or safely halt operations to handle low-confidence states.

Containment: Although the European Union’s framework limits its operational scope to higher-level human override capabilities and safe abort procedures, containment is integrated into the technical dimension as a critical architectural property for fault isolation. While properties like resilience measure how the system maintains state under degradation, and reversibility dictates how it rolls back operations, containment addresses the spatial restriction of failures. It defines the system’s structural capacity to enforce sandboxing, virtualization, or microsegmentation boundaries (Weyns, 2021). This ensures that if a specific component or algorithmic path enters a low-confidence state or undergoes an adversarial compromise, the anomalous behavior is strictly localized and mathematically bounded. Consequently, containment is justified as a necessary architectural safeguard to prevent localized execution failures from propagating into catastrophic cascading collapses across the entire system topology.

Privacy, governance and integrity: To fully cover the Privacy and Data Governance requirement, we distinguish between data protection and data management. Privacy-Preserving Design mandates the integration of structural privacy principles—such as encryption, pseudonymization, and data minimization—throughout the system’s life cycle. Concurrently, governance addresses the EU’s strict demands regarding the quality, completeness, and provenance of training datasets, alongside enforcing strict access protocols and modification logging to ensure data lifecycle integrity.

Efficiency: Although the European Union’s framework does not explicitly mandate computational efficiency as a standalone core requirement—focusing its evaluation primarily on environmental sustainability and resource footprint—this property is integrated into the technical dimension as a critical structural constraint. In software and systems engineering, efficiency is not merely an optimization convenience; it is a foundational operational requirement. If an algorithmic system exhibits unbounded latency, excessive memory utilization, or poor complexity scaling, the theoretical guarantees regarding technical robustness, real-time safety, and interactive explainability inevitably collapse under the friction of high-throughput production environments. Therefore, aligned with the ‘Green AI’ paradigm (Schwartz et al., 2020), efficiency is justified as a necessary condition to ensure that internal risk management loops, safety audits, and real-time compliance operations can execute reliably within deterministic and viable computational bounds.

3.2 Socio-technical dimension

Interpretability and explainability: The European framework emphasizes Transparency as a core pillar of trust, which encompasses traceability and clear reasoning. In our methodology, interpretability represents the structural clarity and mathematical legibility of the model’s parameters. Explainability evaluates the system’s capability to translate its internal execution logic into contextual, human-readable explanations for both expert users and impacted subjects.

Observability: Although the European Union’s framework emphasizes traceability through the implementation of passive historical logging, observability is integrated into the socio-technical dimension as an active, runtime capability. Grounded in contemporary software engineering and site reliability paradigms, observability measures how effectively the internal, dynamic states of a complex system can be inferred solely by monitoring its external outputs and data streams (Majors et al., 2026). For advanced algorithmic architectures—such as retrieval-augmented generation pipelines or deep decision networks—historical logs are insufficient to diagnose real-time behavioral drifts. Introducing observability as a standalone property is justified because it provides the necessary infrastructure for live telemetry, runtime metric inspection, and execution path visualizations, thereby granting supervisors the functional visibility required for immediate algorithmic auditing and formal gap evaluation.

Supervision, reliance and communication: Directly answering the Human Agency and Oversight mandate, supervision evaluates the concrete implementation of governance mechanisms—specifically Human-in-the-Loop (HITL), Human-on-the-Loop (HOTL), and Human-in-Command (HIC) architectures—ensuring that operators possess explicit capabilities, such as a stop button, to override or abort operations. To prevent human autonomy from being diminished through over-reliance or manipulation, reliance measures the calibration of trust between the user and the machine. Lastly, communication fulfills the EU disclosure mandates by ensuring that end-users are pro-actively informed when interacting with an AI system, and that the system’s technical limitations, capabilities, and error rates are transparently conveyed.

3.3 Social dimension

Fairness and accessibility: Grounded in the fifth European requirement, Diversity, Non-discrimination and Fairness, fairness assesses the implementation of quantitative metrics and de-biasing pipelines to eliminate unfair historic or structural prejudices in algorithmic scoring. Expanding on this, accessibility operationalizes the egalitarian reachability of the socio-technical architecture. Rather than addressing superficial interface aesthetics, accessibility ensures that the system’s input channels, data pipelines, and transparency outputs are structurally designed to prevent the exclusion of vulnerable populations, transforming technical disclosure into a universally functional right.

Beneficence: Aligning with the sixth requirement, Societal and Environmental Well-being, beneficence extends evaluation boundaries beyond the immediate operator, assessing the algorithm’s long-term macro-effects on workforce dynamics, human de-skilling risks, social relationships, and democratic institutions.

Accountability and contestability: Directly answering the seventh and final requirement, Accountability, the EU mandates that trustworthy systems must be independently auditable and embedded within robust risk management workflows. In our framework, accountability mandates that execution traces, trade-off decisions, and impact assessments are securely preserved to facilitate external third-party audits. Concurrently, contestability translates the requirement for redress by design into an operational property, guaranteeing that subjects negatively impacted by an algorithmic outcome have clear, actionable, and architectural pathways to challenge, report vulnerabilities, and trigger manual rectification of decisions.

4. The technical dimension

The technical dimension includes properties that can be formalized, measured, tested, or engineered directly into the system.

• Containment: This property refers to the ability of an algorithm to limit the propagation and impact of faults, errors, or perturbations, ensuring that their effects remain confined to a bounded portion of the system rather than affecting the entire computation (Holzmann and Joshi, 2005). A trustworthy algorithm must possess architectural boundaries that restrict the blast radius of an unexpected execution failure, ensuring that a compromised component cannot induce a system-wide collapse. For instance, in autonomous or modular algorithmic pipelines, containment implies that if a specific sub-routine or external third-party API enters an infinite loop, returns corrupted outputs, or undergoes an adversarial exploit, the system isolates that execution path within a secure sandbox or independent container. Consequently, the core algorithmic architecture continues to function unimpeded, preventing the failure from cascading and destabilizing the entire operational environment.
• Efficiency: This property refers to the ability of an algorithm to achieve its intended objectives while minimizing the consumption of computational resources, such as execution time, memory usage, energy expenditure, or communication overhead (Knuth, 1997). A trustworthy algorithm must operate within viable, deterministic hardware bounds to guarantee scalability and operational viability in high-throughput production environments. For instance, in complex processing streams or real-time decision-making systems, efficiency implies that the algorithm utilizes advanced data structures or pruning techniques to minimize CPU and RAM consumption under peak workloads. Consequently, the architecture prevents performance degradation and resource exhaustion, ensuring that the system remains responsive and sustainable without exceeding infrastructure constraints.
• Governance: This property refers to the ability of an algorithmic or data-driven system to operate under established authority, control, and shared decision-making mechanisms, ensuring that data and algorithmic processes are managed, monitored, and enforced according to defined policies, responsibilities, and objectives throughout their lifecycle (DAMA, 2017). Crucially, governance focuses strictly on the administrative, legal, and procedural metadata of the data pipeline—the provenance and authorization channels—rather than the internal math of the files. A trustworthy algorithm must operate over data streams whose collection methods, authorization levels, and access protocols are transparently documented and strictly enforced throughout the system’s operational lifespan. For instance, in distributed automated architectures or retrieval-augmented pipelines, data governance implies that every operational document or training instance is automatically appended with immutable metadata detailing its origin, licensing constraints, and encryption standards. Consequently, the architecture guarantees that the algorithm processes information strictly within verified legal and operational boundaries, mitigating the risks of unauthorized data exposure or algorithmic non-compliance.
• Integrity: This property refers to the ability of an algorithm and its associated data to maintain accuracy, completeness, and consistency throughout their lifecycle, ensuring that information and computational processes remain unaltered except through authorized and intended operations (Wang and Strong, 1996). While data governance regulates the external administrative rules and lineage of the pipeline, Integrity demands that the actual, binary content of the data remains structurally uncorrupted and statistically valid across the entire execution loop. For instance, in high-throughput data processing or automated decision systems, integrity implies that the architecture continuously runs automated schema validations, checksum verifications, and statistical drift detection tests. Consequently, the system immediately flags and isolates any missing values, truncated fields, or corrupted inputs before they can propagate into the model and distort its operational outputs.
• Privacy: This property refers to the ability of an algorithm to protect personal and sensitive information from unauthorized access, disclosure, or inference, while preserving individuals’ control over when, how, and to what extent information about them is collected, used, and shared (Westin, A.F., 1968). Distinct from data governance, which manages administrative access rights and pipelines, privacy focuses on the algorithmic resistance against data leakage during both training and live inference phases. For instance, in distributed or cloud-based predictive architectures, privacy implies the strict implementation of techniques such as differential privacy or federated learning. Consequently, the algorithm guarantees that its mathematical outputs or updated parameters cannot be reverse-engineered by an adversary to reconstruct the identity or private profiles of the individual subjects contained in the underlying training distribution.
• Resilience: This property refers to the ability of an algorithm to sustain its required operation by adapting to, absorbing, and recovering from faults, disruptions, or unexpected environmental changes while continuing to achieve its intended objectives (Hollnagel et al., 2006). While Robustness demands that the algorithm maintains high calculation accuracy in the presence of noisy data, resilience focuses on survivability and infrastructure uptime during severe execution shocks. For instance, in decentralized decision networks or real-time control pipelines, resilience implies that if a primary compute node goes offline or network topologies experience sudden, critical dropouts, the architecture automatically activates fallback execution modes or degraded operational states. Consequently, the algorithm prevents sudden, hard system crashes and guarantees the continuity of vital processes under duress.
• Reversibility: This property refers to the ability of an algorithm or system to undo, roll back, or safely reverse the effects of previously executed actions, restoring a prior valid state without causing unacceptable loss, inconsistency, or harm (Vidal, 2023). Distinct from supervision, which centers on the supervisor’s active decision workflow, reversibility focuses on the underlying technical capability to safely unwind algorithmic actions without leaving corrupted dependencies behind. For instance, in automated database synchronization or transactional algorithmic workflows, reversibility implies that if an optimization path encounters an execution anomaly or is manually overridden, the system triggers an automated state-rollback sequence. Consequently, the architecture ensures that all modified data fields are reverted to their baseline states, preventing systemic corruption and leaving the environment ready for a safe manual intervention.
• Robustness: This property refers to the ability of an algorithm to maintain correct and good performance under imperfect inputs and varying operating conditions (Shahrokni and Feldt, 2013). While resilience ensures infrastructure uptime during hardware shocks, and reversibility dictates how to roll back states after a failure, robustness demands that the algorithm’s internal mathematical logic natively absorbs statistical anomalies without collapsing. For instance, in automated decision-making or pattern-matching systems, robustness implies that if the incoming input data streams experience significant structural drift or become partially corrupted during high-throughput execution, the algorithm does not experience a hard mathematical failure or output volatile, unpredictable anomalies. Instead, it safely and gracefully degrades its performance, adapting its internal thresholds to maintain a verified baseline of mathematical accuracy. Consequently, the architecture prevents algorithmic volatility, ensuring consistency even when production data diverges from training assumptions.
• Safety: This property refers to the ability of an algorithm to operate without causing unacceptable harm to individuals, organizations, or the environment, ensuring that its decisions, actions, and outputs remain within defined risk boundaries under both expected and unforeseen conditions (Leveson, 2016). While security focuses on defending the system against hostile, intentional adversaries, safety handles internal structural failures, unexpected edge-case miscalculations, and systemic crashes. For instance, in automated robotics or cyber-physical decision pipelines, safety implies that the algorithm is constrained by rigid, deterministic boundary-checking routines that override probabilistic models if an action risks physical collision. Consequently, the architecture intercepts and blocks unsafe execution commands before they manifest externally, ensuring that the system fails safely without causing collateral damage.
• Security: This property refers to the ability of an algorithm and its associated infrastructure to prevent, detect, and respond to unauthorized access, adversarial actions, and intentional threats that may compromise its operation, assets, or decision-making processes (Avizienis et al., 2004). While safety mitigates accidental execution risks, and Integrity guarantees the uncorrupted state of data structures, security explicitly addresses a hostile threat model where an attacker attempts to deliberately exploit software vulnerabilities. For instance, in high-stakes classification or authentication pipelines, security implies that the architecture incorporates proactive defenses against adversarial patches, data poisoning vectors, or model-evasion techniques. Consequently, the system maintains its operational defense and trust even when subjected to sophisticated, targeted attempts to deliberately deceive or compromise the algorithmic logic.

5. The socio-technical dimension

This dimension captures the fact that algorithms operate within human workflows, organizational routines, and decision environments rather than in a vacuum.

• Communication: This property refers to the ability of an algorithmic system to proactively communicate its purpose, capabilities, limitations, operational status, and potential risks to users and stakeholders, ensuring that interactions with the system are informed, transparent, and appropriately contextualized (Felzmann et al., 2020). Distinct from explainability, which details the rationale behind a single output, communication governs the systemic dissemination of what the model can and cannot achieve prior to or during operation. For instance, in deployed predictive or classification pipelines, communication implies that the interface dynamically displays clear notifications regarding the model’s error rates, the recency of its last update, and explicit warnings when a user input approaches out-of-distribution (OOD) thresholds. Thus, the architecture fosters a transparent operational environment, preventing misaligned expectations and mitigating user errors stemming from a lack of system awareness.
• Explainability: This property refers to the ability of an algorithm to generate human-understandable explanations that describe the reasons, causal factors, and decision-making processes underlying its outputs, predictions, or actions (Miller, 2019). While Interpretability focuses on the raw structural legibility of the model’s math, explainability demands post-hoc mechanisms that answer why a specific decision was reached in a user-centric language. For instance, in automated scoring or resource-allocation pipelines, explainability implies that when an application is rejected, the system automatically generates a natural language counterfactual explanation detailing the exact minimum adjustments needed to alter the outcome. That way, the architecture bridges the gap between machine execution and human comprehension, providing the foundation for user understanding and informed trust.
• Interpretability: This property refers to the degree of structural clarity and mathematical legibility of an algorithmic architecture, enabling its internal operations, weights, and parameter configurations to be directly inspected and comprehended by human experts without requiring post-hoc approximations (Lipton, 2018). While explainability addresses user-facing representations, interpretability focuses on the “glass-box” nature of the architecture itself, evaluating how easily a human can inspect the model’s inner mechanics. For instance, in risk-modeling or compliance pipelines, interpretability implies utilizing intrinsically constrained models—such as shallow decision trees or generalized additive models—where every internal weight and node split can be traced and verified. Consequently, the system allows engineers to mathematically audit the internal logic, facilitating the direct detection of structural flaws or hidden biases before deployment.
• Observability: This property refers to the ability of an algorithmic system to make its internal states and behaviors inferable from telemetry, logs, traces, metrics, and other external outputs, enabling real-time monitoring, diagnosis, and auditing of its operation (Majors et al., 2026). Distinct from explainability, which translates outputs post-hoc for users, observability provides real-time telemetry and structural visualization for engineers to audit execution while it occurs. For instance, in production-grade retrieval-augmented generation pipelines or graph-based decision networks, observability implies the implementation of live evaluation frameworks—such as RAGAS metrics or real-time token tracing dashboards—that display dynamic execution graphs and confidence scores. Thus, the architecture transforms opaque execution flows into clear diagnostic environments, allowing the instant detection of operational anomalies, semantic drifts, or silent failures.
• Reliance: This property refers to the degree to which users can appropriately depend on an algorithm’s outputs, recommendations, or decisions, aligning their reliance behavior with the system’s actual capabilities, performance, and limitations to avoid both over-reliance and under-reliance (Raees and Papangelis, 2026). While supervision establishes the technical buttons to override the system, reliance manages the psychological and cognitive interface cues that ensure those buttons are used correctly. For instance, in automated diagnostic or anomaly-detection pipelines, reliance implies that the interface intentionally introduces cognitive friction—such as forcing the supervisor to review key conflicting data points before accepting a high-confidence recommendation—to combat complacency and alert fatigue. Consequently, the architecture actively balances human-machine collaboration, ensuring the supervisor remains an alert, analytical, and meaningful participant within the Human-in-the-Loop (HITL) framework.
• Supervision: This property refers to the ability of human operators to monitor, supervise, intervene in, override, or halt the operation of an algorithm when necessary, ensuring that meaningful human control is maintained throughout the system’s lifecycle and decision-making processes (Amershi et al., 2019). Distinct from reversibility, which focuses on the automated technical rollback mechanism itself, supervision establishes the human-in-the-loop (HITL) or human-on-the-loop (HOTL) protocols required to govern execution. For instance, in high-stakes clinical decision or financial risk pipelines, supervision implies that the algorithm cannot finalize a critical action without an explicit digital signature from a verified human supervisor, who is presented with clear veto options. Thus, the system ensures that ultimate governance remains socio-technical, preventing unchecked algorithmic autonomy from dictating sensitive operational outcomes.

6. The social dimension

The social dimension addresses how algorithmic systems interact with legal norms, ethical values, social power, and public legitimacy.

• Accessibility: This property refers to the technical guarantee that an algorithmic system’s information pipelines, input mechanisms, and output disclosures are structurally designed to be perceived, operated, and understood by individuals with diverse cognitive, sensory, and physical capabilities (Shin, 2021). While explainability provides the rationale behind a decision, and contestability establishes the framework to challenge it, accessibility uniquely governs the multi-modal reachability of these assets. It demands that technical barriers do not prevent diverse populations from independently completing system inputs or consuming system outputs. For instance, in automated public service platforms, accessibility implies that the system natively operationalizes programmatic semantic tags, flexible text-to-speech rendering, and low-cognitive-load layouts across its entire execution workflow. Consequently, the architecture prevents systemic digital disenfranchisement, ensuring that the interface of the socio-technical system remains an open, non-exclusionary gateway for all citizens.
• Accountability: This property refers to the ability of an algorithmic system to provide sufficient evidence, documentation, and traceability to enable stakeholders to identify, evaluate, and assign responsibility for its decisions, actions, and impacts throughout its lifecycle (Diakopoulos, 2016). Distinct from observability, which monitors live runtime telemetry for engineering purposes, accountability focuses on creating immutable, forensic records of systemic design choices and operational histories for third-party evaluation. For instance, in high-stakes automated scoring or algorithmic auditing loops, accountability implies that every change in model parameters, data-filtering thresholds, or structural weights is securely logged in an immutable, tamper-evident ledger along with the human rationale behind that architectural modification. Thus, the architecture guarantees complete algorithmic and legal traceability, providing independent auditors with the necessary documentation to review, verify, and assign liability for algorithmic impacts.
• Beneficence: This property refers to the ability of an algorithmic system to produce beneficial outcomes for individuals and society, maximizing positive impacts and contributing to human welfare, social good, and sustainable progress (Floridi & Cowls, 2019). While efficiency measures internal computational limits like CPU consumption, beneficence extends to systemic externalities, such as the overall carbon footprint of training infrastructure, human de-skilling risks, or political polarization vectors. For instance, in massive automated moderation networks or large-scale optimization deployments, beneficence implies the implementation of continuous sociological and environmental auditing frameworks that track job displacement rates within affected sectors or total megawatt-hour consumption metrics. That way, the architecture ensures that the algorithm’s operational success does not come at the expense of public well-being, climate goals, or democratic stability.
• Contestability: This property refers to the ability of an algorithmic system to enable affected individuals and stakeholders to challenge, review, appeal, and seek correction of its decisions, outputs, or underlying processes, ensuring that potentially harmful, erroneous, or unfair outcomes can be meaningfully scrutinized and contested (Moreira et al., 2025). While supervision manages the live supervisor’s control during execution, contestability governs the rights and mechanisms available to the external end-user after a decision has been finalized. For instance, in automated credit scoring or automated recruitment pipelines, contestability implies that the system interface provides an immediate, legally binding “Dispute Decision” mechanism that automatically pauses the automated workflow and routes the entire case file—including its generated explanations—to an independent human review panel for manual override. Consequently, the architecture protects individual autonomy, transforming absolute algorithmic enforcement into an open, accountable process capable of manual correction.
• Fairness: This property refers to the ability of an algorithmic system to make decisions and produce outcomes in a manner that avoids unjustified discrimination, bias, or disparate treatment, ensuring that individuals and groups are treated equitably according to relevant and legitimate criteria (Barocas et al., 2023). Distinct from beneficence, which assesses broad macro-environmental consequences, fairness targets specific group and individual equity metrics inside the model’s predictive distribution. For instance, in algorithmic predictive profiling or resource-allocation systems, fairness implies that the ingestion pipeline actively applies mathematical de-biasing techniques—such as equalized odds or demographic parity optimization—to ensure the algorithm does not produce disparate impacts across protected attributes like gender, race, or socio-economic background. Thus, the architecture enforces mathematical equity, ensuring that automated classifications remain unbiased and legally compliant with non-discrimination standards.

7. Taxonomic synthesis and functional matrix

The operational success of this architectural framework lies not in the isolated verification of its 21 properties, but in their systematic, functional synthesis across the three proposed dimensions. Algorithmic trustworthiness is an emergent property of the entire socio-technical system. This chapter provides a unified matrix that bridges the gap between high-level European regulatory mandates and concrete software engineering specifications.

7.1 Inter-dimensional Dynamics and Cross-cutting Enablers

To prevent the “sociotechnical traps” described in Chapter 3, the architecture operationalizes cross-cutting dependency pathways where technical performance directly constrains social and human capabilities:

• The telemetry-governance loop: Observability (socio-technical) acts as the runtime enabler for both supervision (socio-technical) and accountability (social). Without active telemetry streams and RAGAS-grade metrics tracking live execution graphs, a human supervisor cannot exert meaningful control (supervision), and independent third-party auditors lack the dynamic provenance data required to assign forensic liability (accountability).
• The structural rights pathway: Fairness (social) and accessibility (social) form a defensive front for individual protection. While fairness mitigates mathematical and statistical demographic biases within the predictive distributions (technical dimension), accessibility ensures that the programmatic input/output pipelines are structurally reachable. A system cannot be deemed fair if its multi-modal transparency reports (explainability) are locked behind sensory or cognitive barriers.
• The survivability-control balance: Containment and resilience (technical dimension) provide the deterministic boundary-checking environments necessary for reversibility and supervision. Technical safety loops must isolate component failures (containment) within secure sandboxes to grant human operators the necessary time-window to execute manual overrides or trigger state-rollback sequences (reversibility) via explicit digital signatures.

7.2 The trustworthiness architecture matrix

The following functional matrix synthesizes the structural alignment of our taxonomy. It maps each of the 7 high-level European Union mandates to its corresponding operational properties, explicitly broken down by technical, socio-technical, and social evaluation layers:

EU Trustworthy AI Requirement	Technical Dimension	Socio-technical Dimension	Social Dimension
1. Human Agency & Oversight	Safety, Reversibility	Supervision, Reliance	Contestability
2. Technical Robustness & Safety	Robustness, Security, Resilience, Containment	Observability	Accountability
3. Privacy & Data Governance	Privacy, Integrity	Communication	Governance
4. Transparency	Integrity	Interpretability, Explainability, Observability	Accountability
5. Diversity, Non-discrimination & Fairness	Robustness	Communication	Fairness, Accessibility
6. Societal & Environmental Well-being	Efficiency	Communication	Beneficence
7. Accountability	Integrity, Reversibility, Containment	Observability, Supervision, Explainability	Accountability, Contestability

8. Conclusions and future work

The conceptualization of algorithmic trustworthiness can no longer rely on high-level ethical abstractions or passive, compliance-oriented checklists. As computing environments scale in complexity, transitioning from theoretical safety to verifiable execution requires a rigorous, multi-layered deconstruction of governance mandates into concrete software engineering constraints. This paper has systematically bridged this gap by decomposing the European Union’s foundational requirements into a unified taxonomy of 21 distinct, operational properties, structurally organized across three interdependent layers: the technical, socio-technical, and social dimensions.

The core contribution of this multi-dimensional architecture lies in its explicit enforcement of cross-cutting socio-technical dependencies. By separating the system’s internal mathematical integrity (technical dimension) from its operational human workflows (socio-technical dimension) and macro-environmental impacts (social dimension), the framework establishes that trustworthiness is an emergent property of the entire system topology rather than an isolated optimization metric. Crucial demarcations operationalized herein—such as decoupling the structural clarity of interpretability from user-centric explainability or distinguishing the real-time engineering telemetry of observability from the historical forensic auditability of accountability—provide the necessary lexicon to prevent the common sociotechnical traps that destabilize deployed architectures.

Furthermore, this synthesis elevates structural constraints like efficiency and containment into foundational pillars of algorithmic trust, proving that technical properties directly limit or enable social capabilities. A system cannot guarantee fairness or execute contestability if its interface assets are restricted by cognitive and sensory barriers under accessibility, nor can its ethical parameters survive if unbounded latency or resource exhaustion forces approximations that bypass safety monitors. Ultimately, by establishing a deterministic, auditable, and socio-technically grounded matrix, this taxonomy provides the engineering and regulatory communities with a precise architectural blueprint to ensure that algorithms designed to govern sensitive workflows remain structurally worthy of trust under real-world production duress.

While this taxonomic framework provides a comprehensive structural foundation, it also delineates clear avenues for future research. The natural progression of this work involves translating these 21 qualitative properties into quantitative, mathematically verifiable metrics that can be dynamically evaluated at runtime. Future research should focus on the empirical validation of this functional matrix within highly complex, resource-constrained environments—such as distributed IoT pipelines or large-scale digital twin architectures—where the tension between computational efficiency, real-time observability, and systemic safety is most acute. Ultimately, moving from trustworthy design to verifiable operation requires the continuous development of novel auditing frameworks capable of maintaining these architectural guarantees in an ever-evolving socio-technical landscape.

References

Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P.N., Inkpen, K. and Teevan, J., 2019, May. Guidelines for human-AI interaction. In Proceedings of the 2019 chi conference on human factors in computing systems (pp. 1-13).

Avizienis, A., Laprie, J.C., Randell, B. and Landwehr, C., 2004. Basic concepts and taxonomy of dependable and secure computing. IEEE transactions on dependable and secure computing, 1(1), pp.11-33.

Barocas, S., Hardt, M. and Narayanan, A., 2023. Fairness and machine learning: Limitations and opportunities. MIT press.

European Commission, High-Level Expert Group on AI (2019) Ethics Guidelines for Trustworthy AI. Luxembourg: Publications Office of the European Union.

Felzmann, H., Fosch-Villaronga, E., Lutz, C. and Tamò-Larrieux, A., 2020. Towards transparency by design for artificial intelligence. Science and engineering ethics, 26(6), p.3333.

Floridi, L. and Cowls, J., 2022. A unified framework of five principles for AI in society. Machine learning and the city: Applications in architecture and urban design, pp.535-545.

Data Administration Management Association ed., 2017. DAMA-DMBOK: Data Management Body of Knowledge. Technics Publications.

Diakopoulos, N., 2016. Accountability in algorithmic decision making. Communications of the ACM, 59(2), pp.56-62.

Dietvorst, B.J., Simmons, J.P. and Massey, C., 2015. Algorithm aversion: people erroneously avoid algorithms after seeing them err. Journal of experimental psychology: General, 144(1), p.114.

Grimmelikhuijsen, S., 2023. Explaining why the computer says no: Algorithmic transparency affects the perceived trustworthiness of automated decision‐making. Public Administration Review, 83(2), pp.241-262.

Hoff, K.A. and Bashir, M., 2015. Trust in automation: Integrating empirical evidence on factors that influence trust. Human factors, 57(3), pp.407-434.

Hollnagel, E., Woods, D.D. and Leveson, N. eds., 2006. Resilience engineering: Concepts and precepts. Ashgate Publishing, Ltd..

Holzmann, G.J. and Joshi, R., 2005, October. Reliable software systems design: Defect prevention, detection, and containment. In Working Conference on Verified Software: Theories, Tools, and Experiments (pp. 237-244). Berlin, Heidelberg: Springer Berlin Heidelberg.

Knuth, D.E., 1997. The Art of Computer Programming: Fundamental Algorithms, Volume 1. Addison-Wesley Professional.

Leveson, N.G., 2016. Engineering a safer world: Systems thinking applied to safety. MIT press.

Li, B., Qi, P., Liu, B., Di, S., Liu, J., Pei, J., Yi, J. and Zhou, B., 2023. Trustworthy AI: From principles to practices. ACM Computing Surveys, 55(9), pp.1-46.

Lipton, Z.C., 2018. The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery. Queue, 16(3), pp.31-57.

Majors, C., Fong-Jones, L. and Miranda, G., 2026. Observability engineering: achieving production excellence. ” O’Reilly Media, Inc.”.

Miller, T., 2019. ” But why?” Understanding explainable artificial intelligence. XRDS: Crossroads, The ACM Magazine for Students, 25(3), pp.20-25.

Moreira, C., Palatkina, A., Braca, D., Walsh, D.M., Leihn, P.J., Chen, F. and Hubig, N.C., 2025. Explainable AI Systems Must Be Contestable: Here’s How to Make It Happen. arXiv preprint arXiv:2506.01662.

Raees, M. and Papangelis, K., 2026. From Trust to Appropriate Reliance: Measurement Constructs in Human-AI Decision-Making. arXiv preprint arXiv:2604.23896.

Schwartz, R., Dodge, J., Smith, N.A. and Etzioni, O., 2020. Green ai. Communications of the ACM, 63(12), pp.54-63.

Selbst, A.D., Boyd, D., Friedler, S.A., Venkatasubramanian, S. and Vertesi, J., 2019. Fairness and abstraction in sociotechnical systems. Proceedings of the Conference on Fairness, Accountability, and Transparency, pp.59–68.

Shahrokni, A. and Feldt, R., 2013. A systematic review of software robustness. Information and Software Technology, 55(1), pp.1-17.

Shin, D., 2021. The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable AI. International journal of human-computer studies, 146, p.102551.

Smuha, N.A., 2019. The EU approach to ethics guidelines for trustworthy artificial intelligence. Computer Law Review International, 20(4), pp.97-106.

Spiegelhalter, D., 2020. Should we trust algorithms. Harvard Data Science Review, 2(1), p.1.

Vidal, G., 2023, October. From reversible computation to checkpoint-based rollback recovery for message-passing concurrent programs. In International Conference on Formal Aspects of Component Software (pp. 103-123). Cham: Springer Nature Switzerland.

Wang, R.Y. and Strong, D.M., 1996. Beyond accuracy: What data quality means to data consumers. Journal of management information systems, 12(4), pp.5-33.

Westin, A.F., 1968. Privacy and freedom. Washington and Lee Law Review, 25(1), p.166.

Weyns, D., 2020. An introduction to self-adaptive systems: A contemporary software engineering perspective. John Wiley & Sons.