PaperlayerView exact revision on arXiv

Source-linked AI summary

Connecting the Dots in Trustworthy Artificial Intelligence: From AI Principles, Ethics, and Key Requirements to Responsible AI Systems and Regulation

Natalia Díaz-Rodríguez, Javier Del Ser, Mark Coeckelbergh, Marcos López de Prado, Enrique Herrera-Viedma, Francisco Herrera

arXiv:2305.02231v2cs.CYcs.AIcs.LG

TL;DR

AI’s expanding reach across professional and social domains heightens the need for a holistic account of trustworthy systems. This paper synthesizes ethical principles, ethics, regulation, and technical requirements, concluding that auditability, accountability, and related requirements are central in medical AI.

  • Problem

    AI’s expanding applications and potential impacts create a need for a holistic vision of trustworthy AI across system life cycles.

  • Method

    The paper synthesizes ethical principles, AI ethics, legislation, technical requirements, and auditing practices across four axes.

  • Results

    Auditability and accountability are central to trustworthy medical AI, alongside ethics, data governance, and transparency throughout the AI life cycle.

  • Takeaways & Limitations

    Responsible AI systems may require different degrees of trustworthiness compliance depending on their application domain.

  • Takeaways & Limitations

    Regulatory sandboxes face insufficient methodological assessment of AI’s societal impacts and unresolved needs for cross-border standardization.

Abstract

from arXiv · show

Trustworthy Artificial Intelligence (AI) is based on seven technical requirements sustained over three main pillars that should be met throughout the system's entire life cycle: it should be (1) lawful, (2) ethical, and (3) robust, both from a technical and a social perspective. However, attaining truly trustworthy AI concerns a wider vision that comprises the trustworthiness of all processes and actors that are part of the system's life cycle, and considers previous aspects from different lenses. A more holistic vision contemplates four essential axes: the global principles for ethical use and development of AI-based systems, a philosophical take on AI ethics, a risk-based approach to AI regulation, and the mentioned pillars and requirements. The seven requirements (human agency and oversight; robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental wellbeing; and accountability) are analyzed from a triple perspective: What each requirement for trustworthy AI is, Why it is needed, and How each requirement can be implemented in practice. On the other hand, a practical approach to implement trustworthy AI systems allows defining the concept of responsibility of AI-based systems facing the law, through a given auditing process. Therefore, a responsible AI system is the resulting notion we introduce in this work, and a concept of utmost necessity that can be realized through auditing processes, subject to the challenges posed by the use of regulatory sandboxes. Our multidisciplinary vision of trustworthy AI culminates in a debate on the diverging views published lately about the future of AI. Our reflections in this matter conclude that regulation is a key for reaching a consensus among these views, and that trustworthy and responsible AI systems will be crucial for the present and future of our society.

1. Introduction

The paper presents a holistic vision of trustworthy AI spanning ethical principles, AI ethics, regulation, and technical requirements. It connects these dimensions to responsible AI systems, auditing, and regulatory sandboxes for human-centered deployment.

  • Holistic framework: Trustworthy AI is framed as a holistic, systemic approach built on legal, ethical, and technical robustness and seven key requirements.The requirements include human agency and oversight, robustness and safety, privacy and data governance, transparency, fairness, societal and environmental wellbeing, and accountability.
  • Four-axis approach: The paper organizes trustworthy AI around four axes: ethical-use principles, philosophical AI ethics, risk-based regulation, and technical requirements.Figure 1 presents these axes as the route from trustworthy AI toward responsible AI systems.
  • Requirements: Each trustworthy-AI requirement is examined through what it means, why it is needed, and how it can be addressed technologically.The paper offers a technical overview rather than an exhaustive review.
  • Responsible AI: Responsible AI systems are defined as the result of connecting trustworthy AI’s ethical, philosophical, regulatory, and technical dimensions.Their design is proposed to be guided by regulatory sandboxes, while regulation is analyzed for practical applicability.
  • Scope and motivation: The paper bridges theory and practice to support human-centered AI integration and examines debates on AI’s status, moratoria, and international regulation.Its stated aim is to clarify how these elements can enable natural and sustainable integration into everyday human life.

2. Principles for ethical use and development of Artificial Intelligence

This section surveys ethical AI principles through international recommendations, industry practice, and European trustworthy-AI foundations. It emphasizes human rights, safety, fairness, privacy, transparency, accountability, sustainability, and stakeholder participation across the AI life cycle.

  • Section 2 overview: The section analyzes three principle declarations: UNESCO’s ethics recommendation, Telefónica’s Responsible AI by Design in Practice, and an additional declaration identified in Section 2.3.The analysis moves from a general international recommendation to an industry perspective and then to a further principle declaration.
  • UNESCO Recommendation: UNESCO’s recommendation grounds AI ethics in human rights and fundamental freedoms, applying its values and principles to all actors across the AI system life cycle.The principles may evolve through legislation and business guidelines while complying with international law and the United Nations Charter.
  • UNESCO Recommendation: UNESCO’s principles address proportionality and non-harm, safety and security, fairness and non-discrimination, sustainability, privacy and data protection, human oversight, transparency, and accountability.They also include awareness and literacy, plus multi-stakeholder, adaptive governance and collaboration.
  • Industry principles: Telefónica’s five Responsible AI principles require fair, transparent and explainable, human-centered, privacy- and security-by-design AI, extending these principles to third parties.Human-centered AI is aligned with the UN Sustainable Development Goals, while privacy and security standards apply throughout all life cycles.
  • European foundations: The European Commission’s trustworthy-AI foundations are based on four ethical principles, including respect for human autonomy, prevention of harm, and fairness.These principles are grounded in fundamental rights and guide trustworthy development, deployment, and use of AI systems.

3. A philosophical approach to Artificial Intelligence ethics

A philosophical approach to AI ethics frames ethics as a philosophical inquiry and proposes five actions aligned with European Commission guidelines to support ethical AI. These actions examine assumptions, AI’s applications and impacts, future policies, and the prioritization of AI among global problems.

  • Philosophical foundation: Ethics, as a subfield of philosophy, examines questions about good actions, human life, justice, and the good life.This frames the philosophical basis for examining AI ethics.
  • Five ethical actions: The ethical vision of AI comprises five main actions aligned with European Commission ethics guidelines.The five actions are intended to help attain ethical AI.
  • Five ethical actions: Use philosophy and science to critically examine assumptions about AI and human roles, including claims about artificial general or human-level AI.Large language models can appear human-like while functioning differently from human brains and making different mistakes.
  • Five ethical actions: Assess AI’s diverse and pervasive applications attentively, paying attention to the specific details of each application.Examples include search, text generation, and commercial-product recommendations.
  • Five ethical actions: Address pressing ethical and social problems in current AI applications, including privacy, safety, responsibility, and explainability.The passage illustrates these concerns through chatbot risks and questions about responsibility when multiple people contribute to a technology.
  • Five ethical actions: Critically investigate AI policies and ask whether public attention to AI should be prioritized alongside problems such as climate change and poverty.Existing principles and ethical lists may not suffice because gaps remain between principles, technology implementation, standards, and legal regulation.

4. Artificial Intelligence regulation: A risk-based approach

The European Commission’s AI Act proposal presents a horizontal, human-centric and risk-based framework that regulates AI uses through obligations proportionate to their risks. It distinguishes four risk levels, prohibiting unacceptable-risk practices and imposing stringent requirements on high-risk systems, while implementation details remained under debate.

  • Risk-based regulatory framework: The AI Act proposal is the first attempt at horizontal AI regulation, using a technology-neutral definition and tailoring requirements to the risks of specific AI uses.The framework regulates usages rather than models or technologies themselves.
  • Risk-based regulatory framework: Its four risk levels are minimal or no risk, limited risk, high-risk, and unacceptable risk.Minimal-risk systems may be used freely, limited-risk systems face information and transparency duties, high-risk systems face stringent obligations, and unacceptable-risk systems are prohibited.
  • High-risk AI systems: High-risk AI systems can significantly affect people’s life chances or threaten safety and fundamental rights, requiring conformity assessment before market placement.Examples include law-enforcement systems and systems governing access to education; the proposal also includes post-market enforcement and governance at European and national levels.
  • Risk categories: Unacceptable-risk systems that clearly threaten people’s safety, livelihoods, or rights would be prohibited in the EU market.Examples include social scoring, manipulative systems, dark patterns, and real-time remote biometric identification for law enforcement in public spaces.
  • Regulatory approach: The EU’s human-centric approach contrasts with Chinese government-centric and US industry-owned-data approaches by focusing regulation on AI applications and their societal risks.This is the position defended in the paper.
  • Open regulatory issues: Amendments under consideration addressed fundamental-rights impact assessments, high-risk scenario coverage, prohibited practices, copyrighted training content, and general-purpose AI systems.The status of these issues indicates that important implementation details were still being finalized.

5. Trustworthy Artificial Intelligence: Pillars and Requirements

Trustworthy AI requires confidence that systems act as intended, supported by explanations, robustness, privacy, and freedom from bias. It is a multifaceted condition involving systems, actors, and processes across the AI life cycle, organized around lawfulness, ethics, robustness, and seven requirements.

  • Concept of Trustworthiness: Trustworthiness is confidence that an AI system or model will act as intended when facing a given problem.This technical sense of trustworthiness supports user trust in the model.
  • Concept of Trustworthiness: Trust can be strengthened through decision explanations, robustness under different circumstances, privacy protection, and resistance to data biases.Understanding how a model works and produces decisions can increase user confidence.
  • Holistic Perspective: Trustworthiness is necessary for realizing AI’s potential social and economic benefits and extends beyond systems to actors and processes across the AI life cycle.The paper therefore calls for a holistic and systemic analysis of the contributing pillars and requirements.
  • Pillars and Requirements: The framework is based on three pillars—lawfulness, ethics, and robustness—and analyzes seven HLEG requirements throughout the following subsections.The requirements begin with human agency and oversight, technical robustness and safety, privacy and data governance, transparency, and diversity, non-discrimination, and fairness.

5.1. The three pillars of trustworthy Artificial Intelligence

Trustworthy AI rests on three necessary but individually insufficient pillars: lawfulness, ethics, and robustness. These pillars should work synergistically, although tensions can arise among legal, ethical, and operational demands.

  • Pillar foundations: Each pillar is necessary but not sufficient alone; together, they provide the foundation from which trustworthy AI requirements are formulated.Requirements may contribute to one or several pillars.
  • The three pillars: Lawful AI complies with applicable horizontal and domain-specific laws and regulations.Examples include the European General Data Protection Regulation and rules for high-risk medical or financial applications.
  • The three pillars: Ethical AI adheres to ethical principles and values beyond legal compliance, addressing questions that regulation may not yet resolve.Large-language-model democratization and deepfake misinformation illustrate the continuing relevance of ethics.
  • The three pillars: Robust AI avoids unintentional harm and operates safely and reliably across technical performance and confidence, and social usage and context.Robustness therefore spans both technical and social perspectives.
  • Pillar interactions and requirements: The three pillars should act synergistically, but tensions may arise because legal outcomes are not always ethical and ethical concerns may require conflicting legal amendments.The HLEG guidelines connect these pillars to seven trustworthy-AI requirements and provide technical, nontechnical, and ALTAI self-assessment methods.

5.2. Requirement 1: Human agency and oversight

Human agency and oversight require AI systems to support human autonomy, informed decision-making, and fundamental rights while enabling appropriate human supervision. In practice, oversight can range from intervention in every decision cycle to supervision of design, monitoring, and broader impacts, with mechanisms tailored to application and risk.

  • Requirement 1: Human agency and oversight: AI systems should empower people to make informed decisions, support human autonomy, and protect fundamental rights through appropriate oversight mechanisms.Oversight may use human-in-the-loop, human-on-the-loop, or human-in-command approaches.
  • Requirement 1: Human agency and oversight: Human agency and oversight preserve autonomy and control by enabling users to supervise, evaluate, and freely adopt or override AI decisions.These safeguards address manipulation, deception, herding, conditioning, and decisions made without human involvement.
  • Implementation: Oversight mechanisms depend on application area and potential risk, with largely domain-specific methods shaped by supervisor capabilities, background, interfaces, and system design.More structured toolkits are needed for smooth domain-specific implementation.
  • Human oversight: Human-in-the-loop allows intervention in every decision cycle, whereas human-on-the-loop involves intervention during system design and monitoring.These represent different degrees of human involvement in supervising AI-based systems.
  • Human oversight: Human-in-command enables supervisors to oversee an AI system’s overall activity and broader impacts while ensuring that its decisions can be overridden.The impacts include economic, societal, legal, and ethical dimensions.

5.3. Requirement 2: Technical robustness and safety

Technical robustness and safety require AI systems to prevent or minimize harm, remain secure, reliable, accurate, reproducible, and resilient to errors, attacks, and changing conditions. These aims are addressed through technical robustness, safety, and reproducibility methods across the system life cycle.

  • Requirement 2: Technical robustness and safety: Robustness and safety encompass resilience to attacks, fallback plans, general safety, and accurate, reliable, reproducible operation while minimizing unintentional and intentional harm.The requirement concerns systems being secure, reliable, and robust to errors and interruptions.
  • Requirement 2: Technical robustness and safety: Operational changes, malicious interactions, and deviations from expected behavior make mitigation, safety monitoring, and fallback functionalities necessary in real-world and risk-critical applications.Reliability and reproducibility support verification of expected operation and performance across deployment contexts.
  • 5.3.1. Technical robustness: Technical robustness can be organized into levels covering standard training, distribution shifts, single risks, multiple risks, universal robustness, and human-aligned augmented robustness.Distribution-shift methods include concept-drift adaptation, out-of-distribution detection, and class-incremental learning; single-risk robustness includes model inspection and adversarial defenses.
  • 5.3.1. Technical robustness: Robustness should be central to risk management and accountability, supported by maintenance frameworks, life-cycle tracking, and passive or active monitoring of quantitative metrics.Monitoring can measure robustness over the data, model, or both.
  • 5.3.2. Safety: AI safety requires human-value alignment and concrete protocols addressing hazards, adversarial threats, system inspection, anomaly detection, emergent capabilities, safety objectives, value learning, and proxy gaming.Backdoored models are an identified safety risk because poisoned training data can induce incorrect behavior in selected scenarios.

5.4. Requirement 3: Privacy and data governance

Privacy and data governance require protection, quality, integrity, relevance, and legitimate access throughout the AI system life cycle. Trustworthy implementation combines privacy-aware technologies with traceable governance, authorized access, and lawful data-use protocols.

  • What does it mean?: Privacy and data protection must be maintained across design, training, testing, deployment, and operation, alongside data quality, integrity, relevance, and legitimate access protocols.These elements define the requirement throughout the AI system life cycle.
  • Why is it important for trustworthiness?: AI systems must prevent sensitive personal information from being revealed during data processing, storage, and retrieval, while enabling data-use traceability and privacy verification.Unmet guarantees undermine user trust and compliance with legislation such as the European GDPR.
  • How can this requirement be met in practice?: Federated learning trains models locally on decentralized devices and sends only numerical model updates to a central server for aggregation.This supports global-model learning without moving all underlying data to a central location.
  • How can this requirement be met in practice?: Homomorphic computing processes encrypted data without deciphering it, while differential privacy adds calibrated noise to reduce individual-identification risks.Differential privacy balances protection against performance degradation, and authorized parties alone can decrypt homomorphic-computing results.
  • How can this requirement be met in practice?: Data governance should restrict access to qualified staff with explicit need and competence, govern integrity and access broadly, and support fair, lawful data sharing and use.European initiatives include the Data Governance Act and Data Act, which address data sharing, re-use, access, and conditions of use.

5.5. Requirement 4: Transparency

Transparency in trustworthy AI ensures that relevant stakeholders receive appropriate information about systems, data, models, and business models. It is implemented through traceability, explainability, and audience-adapted communication across the AI system life cycle.

  • What does it mean?: Transparency ensures appropriate information reaches relevant stakeholders through simulatability, decomposability, and algorithmic transparency.These levels concern human understanding of the model, its parts and behavior, and the process producing outputs.
  • Why is it important for trustworthiness?: Trustworthy AI requires transparent data, systems, and business models, with timely explanations adapted to stakeholders and system traceability ensured.People should understand system capabilities and limitations and know when they are interacting with AI.
  • Traceability: Traceability tracks data and development and deployment processes through documented records, supporting auditing and audience-appropriate transparency.Traceability and logging should begin during early design stages of AI-based systems.
  • Traceability: Provenance tools support data and decision lineage, while blockchain mechanisms may help preserve the integrity and provenance of training data and associated quality, bias, and fairness information.These mechanisms contribute to transparency by making data origins and model decisions easier to track.
  • Explainability and communication: Explainability supports algorithmic auditing by helping validate and understand black-box models, while communication must disclose AI interaction, performance, capabilities, limitations, and explanations in audience-adapted formats.Explanations should reflect the audience’s background and the motivation for generating them.

5.6. Requirement 5: Diversity, non-discrimination and fairness

Requirement 5 encompasses diversity, non-discrimination, accessibility, stakeholder participation, and fairness to prevent unjustified restrictions on human choice and reduce harms from biased automated decisions. It is implemented across the AI life cycle through inclusive design, bias mitigation, fairness techniques, and auditing against fairwashing.

  • What does it mean?: Diversity, non-discrimination and fairness require avoiding unfair bias, fostering diversity, ensuring accessibility, and involving stakeholders throughout the AI system life cycle.These dimensions share the purpose of preventing AI systems from deceiving humans or unjustifiably limiting their freedom of choice.
  • Why is it important for trustworthiness?: The requirement broadens AI’s impact across social groups while minimizing harms when learned data contains hidden biases that may marginalize vulnerable groups.The passage identifies biased data as a source of negative implications for automated decisions.
  • How can this requirement be met in practice?: Implementation methods address diversity, non-discrimination, accessibility, universal design, stakeholder participation, fairness, and unfair-bias avoidance through distinct approaches.The paper organizes these methods across Subsections 5.6.1–5.6.3.
  • Diversity, non-discrimination and accessibility: Diversity should be built into data, models, stakeholders, and design processes, including minority representation, participatory design, universal design, and post-deployment feedback.Active learning can integrate users’ feedback during model learning and support interactivity.
  • Fairness and bias mitigation: Fairness techniques operate before, during, or after model training, while fairwashing can falsely suggest ethical compliance and requires scrutiny when auditing opaque models.Pre-processing transforms data, in-processing modifies learning, and post-processing changes outputs; fairness may trade off with accuracy or privacy.

5.7. Requirement 6: Societal and environmental wellbeing

Societal and environmental wellbeing requires AI systems to benefit present and future generations while remaining sustainable, environmentally friendly, and socially responsible. In practice, this requirement combines environmental sustainability with societal wellbeing, supported by technical strategies and attention to fairness, privacy, transparency, and human oversight.

  • What does it mean?: AI systems should benefit all humankind across present and future generations without progressively depleting natural resources or disrupting ecological balance.This requirement includes sustainability, environmental friendliness, and careful assessment of social wellbeing.
  • Why is it important for trustworthiness?: AI can support positive social change and climate-change mitigation, but computationally intensive training can increase greenhouse-gas emissions and related social and ethical challenges.The passage notes that training a single AI model can emit CO2 emissions comparable to five cars in the truncated example.
  • How can this requirement be met in practice?: The requirement is addressed through two complementary angles: sustainability and environmental wellbeing, and societal wellbeing.These are treated in Subsections 5.7.1 and 5.7.2, respectively.
  • Sustainability and environmental wellbeing: Sustainable AI spans models, data, algorithms, and hardware across design, training, and deployment, using co-design, shared practices, metrics, and standards to reduce carbon footprints.Proposed measures include environmental-impact assessment, smart-data selection, model compression, and efficiency as an evaluation metric and price tag.
  • Societal wellbeing: AI can improve productivity, public administration, policy making, urban planning, and quality of life, but socially consequential decisions require fairness, privacy, transparency, human oversight, and compliance with human rights and law.The examples include autonomous routine tasks, administrative process improvements, climate-change visualization, flood prediction, and urban-heat identification.

5.8. Requirement 7: Accountability

Accountability requires mechanisms that assign responsibility for AI systems and their outcomes across development, deployment, maintenance, and use, with auditability enabling assessment and attribution. In practice, trustworthy accountability combines auditing, risk-aware oversight, impact reporting, and algorithmic redress, especially in high-risk and generative-AI contexts.

  • What does it mean?: Accountability requires mechanisms assigning responsibility for AI development, deployment, maintenance, use, and outcomes, with auditability supporting assessment and attribution.Auditability assesses algorithms, data, and design processes to attribute results to actions based on AI-system outputs.
  • Why is it important for trustworthiness?: Practical auditability tools should verify neural-network properties and track traceability, data quality, and integrity beyond explainability.Examples of desirable properties include stability, sensitivity, relevance, and reachability.
  • Why is it important for trustworthiness?: Accountability supports recourse for proven wrong or unfair AI decisions and requires compliance, answerability, reporting, oversight, attribution, and consequence enforcement.Redress procedures can correct or reverse outcomes considered wrong, preserving user trust after adverse or unfair impacts.
  • How can this requirement be met in practice?: In high-risk scenarios, accountability tools assign responsibility across design, development, and deployment while using policy toolkits, post-hoc analysis, and causal inference.Accountability is linked to fairness, and emerging trade-offs should be assessed for risks to ethical requirements and fundamental rights.
  • How can this requirement be met in practice?: Stable, verifiable auditing mechanisms are especially urgent for generative AI because unidentified machine-generated content can enable confusion, deception, public-opinion manipulation, and fake news.Verifiable claims offer falsifiable claims supported by evidence and arguments to demonstrate AI-system properties credibly.

6. Trustworthy Artificial Intelligence from theory to practice and regulation: responsible Artificial Intelligence systems

The section develops responsible AI systems as a practical extension of trustworthy AI, requiring auditable, accountable, legally compliant, and ethical operation throughout the system life cycle. It emphasizes regulatory sandboxes, standardized assessment, and dynamic regulation to address implementation challenges and emerging AI applications.

  • Practical trustworthy AI: Practical trustworthy AI requires mapping principles and requirements into operative protocols that can be automated, verified, and audited through standardized blueprints and models.The section identifies regulatory sandboxes as practical regulatory scenarios for implementing trustworthy AI.
  • Responsible AI systems: A Responsible AI system ensures auditability and accountability during design, development, and use according to applicable domain regulations and specifications.Its decisions must also be legally compliant and ethical.
  • Responsible AI systems: Auditability validates conformity with sectoral regulations, AI-wide regulations, and application-specific constraints, while accountability establishes liability for audited system decisions.Auditability may require transparency, including explainability methods, and accountability can involve different levels of trustworthy-AI compliance.
  • Regulatory sandboxes: Regulatory sandboxes support pre-market assessment and adaptable domain-specific requirements, motivated by the need for harmonized testing protocols, monitoring plans, and conformity procedures.Under the European approach and AI Act, sandboxes help adapt trustworthy-AI requirements and regulation to the system’s practice domain.
  • Regulatory sandboxes: Sandboxes remain immature because methodological impact assessments, cross-border standardization, rapid algorithmic auditing, and validated high-risk systems are unresolved challenges.Governments must balance EU coordination with national procedures to avoid regulatory implementation conflicts.
  • Future regulation: Dynamic regulation and independent periodic assessment are needed to support responsible AI systems as novel applications emerge, including neurotechnology risks to fundamental rights such as mental privacy.Sandbox results and certification activities are expected to improve systems and enable their progressive deployment in practical scenarios.

7. From the Artificial Intelligence moratorium letter to regulation as the key for consensus

The debate over AI’s rapid development, existential risks, safety, and public benefits underscores the need for urgent ethical and regulatory frameworks. The paper argues that regulation can create consensus among diverging perspectives, supported by responsible-AI technologies, auditability, and accountability.

  • Debate and regulatory urgency: The moratorium letter highlights the gap between rapidly advancing, high-powered AI systems and regulation.It calls for refocusing AI research on accuracy, safety, interpretability, transparency, robustness, alignment, trustworthiness, and loyalty.
  • Debate and regulatory urgency: Experts emphasize urgent control, alignment with human benefit, and rigorous safety checks before powerful AI systems enter public use.The discussion links AI’s non-evolved goals to the alignment problem and compares pre-release checks with pharmaceutical safety testing.
  • Debate and regulatory urgency: The controversy over AI’s potential existential threat has intensified the need for ethical and regulatory frameworks governing whether and how AI systems are trusted and used.Other proposals call for global licensing, international cooperation, intergovernmental oversight, standard-setting, and treating AI extinction risk as a global priority.
  • Policy proposals: Policy proposals emphasize improved legal liability, lifecycle-wide regulatory scrutiny, transparency about training data, and human supervision of automated systems.These measures aim to address harms, corporate design and data choices, bias, and false or misleading information.
  • Regulation and consensus: The paper positions regulation as a key for consensus, requiring technologies, methodologies, and tools that support responsible-AI development, auditability, and accountability.Despite risks in conflating trustworthiness with risk acceptability, regulation is presented as a driving force for consolidating diverging views about AI’s benefits and restrictions.

8. Concluding remarks

The manuscript presents a holistic account of trustworthy AI and emphasizes auditability and accountability as practical foundations of responsible AI systems. It concludes that responsible AI should guide AI design throughout the life cycle, while regulation should focus on products and their use rather than scientific progress.

  • 8. Concluding remarks: The paper synthesizes the principles, pillars, and requirements that trustworthy AI systems must meet.It approaches trustworthy AI from a holistic perspective for researchers, practitioners, and newcomers.
  • 8. Concluding remarks: Auditability and accountability are identified as essential practical conditions for responsible AI systems.Auditability concerns conformity to regulations, while accountability concerns explaining how decisions are issued.
  • 8. Concluding remarks: Medical AI recommendations place auditability and accountability at their core, alongside ethics, data governance, and transparency.The domain is presented as risk-critical and especially dependent on trust when adopting technological advances.
  • 8. Concluding remarks: Trustworthy AI assessment should examine all essential elements in regulatory sandboxes and establish clear accountability protocols for each domain of practice.Responsible AI systems should be the final output and goal of current AI designs and developments.
  • 8. Concluding remarks: The conclusion argues that regulation should target AI products and their usage rather than scientific progress.The paper frames life-cycle consideration of trust and requirements as key to designing and developing responsible AI systems.
Loading 2305.02231v2…