The Seven Requirements of Trustworthy AI: Technical, Data, and Governance Methods

From principles to actionable requirements

This article is the English translation of Chapter 2 of the "Ethical Guidelines for Trustworthy AI" 2019 edition, corresponding to printed pages 14–23. Previous chapter Proposes respect for autonomy, harm prevention, fairness, and interpretability; this chapter translates these into seven requirements and explains the technical and organizational methods. Historical policies, standards, and research advances reflect the context of the 2019 original edition.

The previous chapter also includes a recurring reminder throughout the following text: while AI can bring significant benefits, it may also generate unforeseeable, detectable, and measurable risks, such as impacts on democracy, the rule of law, distributive justice, and human cognition. Appropriate risk-mitigation measures should be implemented in proportion to the level of risk.

Who is responsible for implementing the seven requirements?

These requirements apply to different stakeholders throughout the lifecycle. Developers refer to individuals who research, design, or develop AI; deployers refer to public or private organizations that use AI in business processes or provide products and services to others via AI; end users refer to individuals who interact directly or indirectly with the system; and the broader society includes all others who are directly or indirectly affected.

Developers should implement requirements into design and development processes; deployers should ensure that the systems and products and services they provide meet the requirements; end users and the broader society should be aware of these requirements and be able to demand that they are upheld.

The following requirements cover system, individual, and societal levels, but are not exhaustive. They are equally important and interdependent, and should be continuously implemented and evaluated throughout the entire lifecycle; the ordering does not imply hierarchy. Specific implementation should also take into account domain, use case, conflicting considerations, and the extent of impact. While most requirements apply to all AI systems, particular attention should be given to systems that directly or indirectly affect people; in certain industrial applications, the relevance of some requirements may be limited.

View textual explanation of the diagram

The seven requirements are equally important, mutually supportive, and may involve tensions that require management. Implementation and evaluation should occur throughout development, deployment, usage, and subsequent improvement, with specific actions adjusted according to domain, purpose, and potential risks. The numbering corresponds to the original text and does not imply a sequence where one requirement should be completed before moving to the next.

Some requirements are reflected in existing laws. Practitioners remain responsible for complying with applicable general laws and sector-specific regulations; this chapter does not override these obligations.

I. Human Initiative and Supervision

Based on the principle of respecting human autonomy, AI should support human autonomy and decision-making, promote user agency, fundamental rights, and human oversight, thereby helping to foster a democratic, prosperous, and equitable society.

Basic rights

AI can both promote and undermine fundamental rights. For example, it can help individuals track personal data and increase access to education, supporting the right to education; yet its scope and capabilities may also harm rights.

When such risks exist, a basic rights impact assessment should be conducted prior to system development to determine whether the risks can be mitigated; if rights are restricted, the assessment should evaluate whether the restriction can be justified as necessary for protecting the rights and freedoms of others in a democratic society. At the same time, a mechanism for receiving external feedback should be established to enable potential rights infringements to be identified and reported.

Human initiative

Users should be able to make informed, autonomous decisions about AI. They need knowledge and tools to fully understand and interact with the systems, and to reasonably evaluate or question the systems when possible. The system should assist people in making better, more informed choices aligned with their own goals.

Some systems exploit subtle mechanisms that influence behavior, even leveraging unconscious processes to engage in unfair manipulation, deception, displacement, or behavioral shaping, threatening individual autonomy. Therefore, user autonomy should be at the center of system functionality.

A key point is the right not to be subject to decisions based solely on automated processing which produces legal effects or similar significant impacts. The original text refers to GDPR Article 22; specific conditions and exceptions should be checked in the original regulation, and this statement should not be understood as a blanket ban on all forms of automation.

Human supervision

Oversight can help prevent AI from undermining autonomy or producing other adverse effects, and governance mechanisms can be implemented in three ways:

It is also essential to ensure that public enforcement agencies can carry out oversight in accordance with their responsibilities. The intensity of oversight should be based on the application and potential risks, and should be coordinated with other security and control measures. The more limited the human capacity for oversight under equal conditions, the greater the need for comprehensive testing and stringent governance.

II. Technical Reliability and Security

Technical robustness is closely related to the prevention of harm. System development should adopt risk-preventive approaches, operating reliably as intended, minimizing unintended and accidental harm, and preventing unacceptable harm. This also applies to scenarios involving changes in the operating environment, as well as situations where humans or human agents interact with the system in adversarial ways. The physical and mental integrity of individuals should be protected.

Defending against attacks and security protection

Like other software, AI should avoid vulnerabilities that can be exploited by attackers. Attacks may target data, such as data poisoning; model components, such as model leakage; or the software and hardware infrastructure. Adversarial attacks can alter data, system behavior, and decisions, potentially causing systems to halt completely.

Malicious actions or unintended scenarios can damage systems and data; inadequate security processes can also lead to erroneous decisions, even physical harm. To consider a system secure, one should account for unintended uses, dual-use applications (civilian and military), and malicious abuse, and implement preventive and mitigative measures.

Research should foster a virtuous cycle of understanding attacks, developing defenses, and improving evaluation methods, promoting collaboration between AI and security fields. Relevant parties should also jointly establish cross-border security norms and a culture of mutual trust to advance international cooperation.

Backup plan and general safety

The system should have contingency plans in place when issues arise. For example, switching from a statistical method to a rule-based method, or requesting human intervention before proceeding. It is also necessary to consider scenarios where human intervention is not immediately available.

The system should accomplish its intended tasks without harming life or the environment, and should minimize unintended consequences and errors. Different application domains require established processes for identifying, clarifying, and assessing risks. The level of protection depends on the risk, which in turn is related to the system's capabilities; if the development process or the system itself inherently involves particularly high risks, safety measures should be developed and tested in advance.

Accuracy

Accuracy refers to a system's ability to make correct judgments, such as correctly categorizing information, or making correct predictions, recommendations, or decisions based on data and models. Clear and robust development and evaluation processes help identify, mitigate, and correct unexpected risks arising from inaccurate predictions.

If occasional prediction errors are inevitable, the system should be able to indicate the likelihood of such errors. High accuracy is especially critical when directly affecting human life.

Reliability and reproducibility

A reliable system works appropriately across different inputs and contexts. This is crucial for review systems designed to prevent unintended harm. Reproducibility refers to a system's ability to exhibit the same behavior when the same experiment is repeated under identical conditions, helping researchers and policymakers accurately describe the system's behavior.

Replication files can document and reproduce each development step, from research and initial data collection to final results, facilitating testing and replication.

Three. Privacy and Data Governance

Privacy is a fundamental right particularly vulnerable to AI impacts and closely linked to harm prevention. Protecting privacy requires appropriate data governance, encompassing data quality, integrity, relevance to deployment domains, access protocols, and the capability to process data in a privacy-preserving manner.

Privacy and Data Protection

Privacy and data protection should be ensured throughout the entire lifecycle. This includes not only the information initially provided by users, but also the information formed during interactions about users, such as outputs generated for a particular individual or their reactions to recommendations.

Digital behavior records may not only reflect preferences but also enable systems to infer orientation, age, gender, religion, or political views. To earn human trust in the data collection process, it must be ensured that the collected data is not used for illegal or unfair discrimination.

Data Quality and Integrity

Dataset quality is crucial to system performance. Data collected may contain societal biases, inaccuracies, and errors, which should be addressed prior to training. Data integrity should also be ensured, as malicious data can alter system behavior, particularly in systems that continuously learn.

The workflows and datasets used across planning, training, testing, and deployment stages should be tested and documented. Third-party systems should also be subject to the same scrutiny and documentation, not limited to only internally developed systems.

Data access

Organizations processing personal data should establish access agreements, regardless of whether the data subject is a system user. The agreements should clearly specify who can access data and under what circumstances. Access permissions should be granted only to individuals who are qualified, capable, and genuinely required to do so.

IV. Transparency

Transparency and explainability are closely related, involving data, systems, and business models.

Traceability

Record data and processes that led to decisions at the highest possible standard, including data collection, labeling, algorithms, and decisions made by the system itself. This enables identification of root causes and helps prevent recurrence; traceability also supports auditing and explanation.

Explainability

Explainability includes explaining the technical process, as well as human decisions, such as why a particular application domain was chosen. Technically, humans should be able to understand and trace AI decisions.

Sometimes a trade-off between interpretability and accuracy is necessary: increasing interpretability may reduce accuracy, while improving accuracy might compromise interpretability. When systems have significant impacts on people's lives, they should be able to provide appropriate explanations for their decision-making processes. These explanations should be timely and tailored to the professional level of the relevant individuals, such as ordinary people, regulators, or researchers.

Should also explain to what extent AI influences organizational decision-making, system design choices, and deployment rationale, thereby achieving business model transparency.

Communication

AI should not pretend to be human; people have the right to know they are interacting with AI, and systems should therefore be identifiable. When basic rights require it, people should be allowed to choose to interact with a human.

In addition, capabilities, accuracy, and limitations should be appropriately communicated to practitioners and end users to help form realistic expectations.

Five. Diversity, Non-discrimination, and Fairness

Inclusion and diversity should be promoted throughout the entire lifecycle. This requires not only considering and involving affected parties but also ensuring equal access and treatment through inclusive design.

Avoid unfair bias

Training and runtime data may contain unintended historical biases, incomplete content, or governance issues. Persisting these biases could lead to direct or indirect bias and discrimination against certain groups, exacerbating marginalization.

Harm may also arise from deliberate exploitation of consumer biases or unfair competition, such as collusion to set identical prices or market opacity. Discriminatory biases that can be identified should be removed as early as possible during the data collection phase.

Development approaches, including the algorithmic programming itself, may also carry unfair biases. Oversight processes should be established to clearly and transparently analyze the system's objectives, constraints, requirements, and decisions. Recruiting individuals from diverse backgrounds, cultures, and disciplines to promote varied perspectives should be encouraged.

Accessibility and Universal Design

Especially in consumer-facing applications, systems should be user-centered, ensuring that people of all ages, genders, abilities, and characteristics can access products and services. People with disabilities are present in all social groups, and accessibility in technology is particularly important for them.

The system should not adopt a single approach that fits all users' thinking, but rather follow general design principles and applicable accessibility standards to serve as broad a user base as possible, enabling everyone to participate fairly in existing and emerging computer-assisted activities and to use assistive technologies. The original text references EN 301 549 and connects it to public procurement requirements for "designing for all" and the Convention on the Rights of Persons with Disabilities.

Stakeholder engagement

Consult directly and indirectly affected parties throughout the entire lifecycle. Feedback should be regularly collected after deployment, and long-term mechanisms for participation should be established. For example, throughout the entire process of organizing the implementation of AI, ensure that workers are informed, consulted, and involved.

VI. Social and Environmental Well-being

Based on principles of fairness and harm prevention, broader societal groups, other sentient beings, and the environment should also be considered stakeholders throughout the lifecycle. Sustainability and ecological responsibility should be encouraged, promoting the use of AI to address global challenges such as the Sustainable Development Goals. Ideally, AI should benefit all of humanity, including future generations.

Sustainable, environmentally friendly AI

AI has the potential to help address urgent societal challenges, but the methods used to achieve this should also be as environmentally friendly as possible. The impacts of development, deployment, usage, and the entire supply chain should be assessed (such as examining resource and energy consumption during training) and prioritizing approaches with lower environmental harm, while encouraging measures that ensure environmental sustainability across the entire supply chain.

Social impact

Widespread exposure to social AI in education, work, caregiving, and entertainment may alter our understanding of social agents, social relationships, and attachment. Here, social AI refers to systems that simulate social interaction and communicate with humans through physical robots or virtual avatars, potentially reshaping social cultural practices and the structure of social life.

AI may enhance social abilities or potentially degrade them, thereby affecting mental and physical well-being. Therefore, its impacts must be carefully monitored and considered. The original text cites research on robot-assisted autism therapy in human-led treatment settings, demonstrating potential benefits in improving social and communication skills in children, without claiming that all social AI has such effects.

Society and Democracy

In addition to the individual level, an assessment of the impacts on institutions, democracy, and society as a whole during development, deployment, and use should also be conducted. Special caution is required when involving democratic processes, encompassing not only policy decisions but also electoral contexts.

Seven、Accountability

Accountability complements the above requirements and is closely tied to fairness. Mechanisms must be established to ensure that systems and their outcomes are held accountable at every stage (during development, deployment, and throughout use) so that responsibility is assigned and oversight is maintained.

Auditable

Allow evaluation of algorithms, data, and design processes, but this does not mean that business models and intellectual property information must always be completely disclosed. Accessibility of internal and external audits, as well as relevant evaluation reports, contributes to credibility. For applications involving fundamental rights, especially those with security-critical functions, independent audits should be possible.

Minimize and report negative impacts

Should be able to report which actions or decisions led to a particular outcome, as well as address the consequences of that outcome. It is especially important to identify, assess, document, and mitigate potential negative impacts for both directly and indirectly affected parties.

Reporters, non-governmental organizations, trade unions, and others should receive appropriate protection when raising legitimate concerns. During the development, deployment, and use of AI systems (both before and during implementation) methods such as red-teaming and algorithmic impact assessments can be employed to mitigate negative effects, with evaluations proportionate to the associated risks.

Weighing

Conflicts among requirements may arise, and when trade-offs are necessary, they should be addressed using rational, systematic methods within the current technological capabilities. First, identify the involved interests and values, then clearly acknowledge the conflict, and assess the risks to ethical principles and fundamental rights associated with the trade-offs.

If no ethically acceptable trade-offs exist, development, deployment, or use should not proceed in its current form. Every trade-off decision should have a rationale and be appropriately documented; those making the trade-off decisions are responsible for the manner in which they are made and should continuously review whether the decisions remain appropriate, modifying the system as necessary.

Internal or external ethics and field experts, ethics committees, and meaningful consultation with those who may be adversely affected can help identify conflicts and find solutions. The original text also emphasizes the role of higher education and research institutions in cultivating professional ethics talent.

relief

When adverse effects of unfairness occur, there should be readily accessible mechanisms sufficient to provide appropriate redress. The knowledge that recourse is available when problems arise is a fundamental condition of trust; special attention should be given to vulnerable individuals and groups.

The implementation method requires continuous evaluation

Technical and non-technical approaches can both be used to meet requirements across the entire lifecycle. Methods should be continuously evaluated, and changes to implementation processes should be reported with justification, such as why a particular design, development, or deployment decision was made.

The system continuously evolves and operates in dynamic environments, making trustworthy AI a continuous process. As shown in the original Figure 3, rights, principles, and values are first translated into requirements, then implemented through both technical and non-technical methods; subsequently, it enters a cycle of "development, deployment, analysis, re-design, and development," centered on continuous evaluation and justification. It cannot be considered complete after a single pre-release review.

The following methods can complement each other or be used interchangeably under specific circumstances. Different requirements and sensitivity levels may necessitate different implementation approaches. This list is neither exhaustive nor a mandatory checklist; the maturity of these methods varies, with some already available as early as 2019, while others are still under research.

Technical methods

Integrate trust requirements into the system architecture

Convert requirements into processes or constraints on processes and embed them in the architecture. For example, specify allowed behaviors or states that must always be followed, prohibit behaviors or states that must never be breached, or combine both, even employing complex behaviors that can be proven to ensure compliance. Independent processes can monitor the system during operation to verify adherence to constraints.

Systems that learn and dynamically adjust their behavior can be viewed as non-deterministic systems that may exhibit unexpected behavior. Using the perception–planning–action cycle as a framework, requirements should be included at all three stages: during perception, identify the environmental elements needed to comply with the requirements; during planning, consider only plans that satisfy them; during action, restrict behavior to the scope that fulfills them.

This architecture provides only a general and not entirely accurate description, but it offers anchor points for constraints and policies to be mapped into specific modules, helping the entire system become trustworthy and perceived as trustworthy.

Integrate ethics and legal governance from the design phase

Methods for integrating value into design establish a clear connection between abstract principles and specific implementations. The key idea is to translate compliance with regulations into system design. Organizations should begin by identifying early the impacts and the regulations needed to avoid negative consequences.

Privacy-preserving design, security design, and similar approaches have become common practices. To earn trust, the system's processes, data, and outputs should be secure and capable of withstanding adversarial data and attacks; it should also include a security shutdown mechanism and be able to resume operations after being forcibly halted due to an attack.

Explainable method

To build trust in a system, it's necessary to understand why it acts the way it does and why it arrives at a particular interpretation. Explainable AI (XAI) research aims to understand underlying mechanisms and find solutions.

The original text from 2019 stated that this remains an open challenge for neural networks: it is difficult to establish intuitive connections between trained numerical parameters and their outcomes, and small changes in data can lead to dramatic shifts in interpretation, such as mistaking a school bus for an ostrich; attackers can also exploit such vulnerabilities. Interpretability research not only helps explain model behavior to users, but is also crucial for reliable deployment.

Testing and Verification

The uncertainty and context-dependence of AI make traditional testing insufficient. Concepts or representations used by the system may only reveal issues when exposed to sufficiently realistic data. Monitoring of the underlying models during both training and deployment is essential to verify stability, robustness, and operation within well-understood, predictable boundaries; planning outcomes should align with inputs, and decision mechanisms should allow validation of the underlying processes.

Testing should begin early and span the entire lifecycle, particularly not being overlooked after deployment; it should cover data, pre-trained models, environments, and overall behavior. Testing should be designed and executed by as diverse a group of people as possible, using a variety of metrics to cover different categories from multiple perspectives.

Consider having a trustworthy and diverse red team intentionally attempt to disrupt the system to identify vulnerabilities, or implement vulnerability rewards to encourage external parties to discover and responsibly report errors and weaknesses. Finally, verify that outputs or actions are consistent with previous processing results and compare them against established policies to ensure no violations of constraints occur.

Service quality indicators

Appropriate metrics can be established to help foster an understanding within the system regarding whether security has been adequately considered during development and testing. These metrics can evaluate algorithm training and testing, or alternatively use traditional software metrics such as functionality, performance, usability, reliability, security, and maintainability.

Non-technical approaches

regulation

Regulatory measures for trustworthiness were already in place at the time, such as product safety laws and liability frameworks. Which regulations need to be amended, adjusted, or introduced to balance protective and promotive effects remains a question discussed in the accompanying policy and investment recommendations.

Code of Conduct

Organizations can commit to adopting the guidelines and incorporate trustworthy AI into their corporate responsibility charters, key performance indicators, code of conduct, or internal policies. Organizations developing or using AI can document their intentions and support these intentions with standards based on fundamental rights, transparency, and harm prevention.

Standardization

Design, manufacturing, and business practices can serve as quality management tools that enable buyers, consumers, organizations, research institutions, and governments to recognize and incentivize ethical behavior through procurement.

In addition to traditional standards, certification and accreditation systems, professional ethics guidelines, and design standards aligned with fundamental rights can also serve as co-governance mechanisms. The original text cites ISO standards and the IEEE P7000 series, and discusses the possibility of future trustworthy AI labels, using specific technical standards to verify requirements such as safety, technical robustness, and transparency. This represents a 2019 vision and cannot be used to claim that any product has received certification today.

Authentication

It is not feasible to expect every individual to fully understand how AI operates and its impacts, so organizations could demonstrate system transparency, accountability, and fairness to the public. Certification should be aligned with domain-specific and technology adoption standards, appropriately integrating relevant industrial and social norms within each context.

But authentication cannot replace accountability and should be accompanied by an accountability framework, including responsibility statements, review processes, and redress mechanisms.

Hold accountability through a governance framework

Organizations should establish internal and external governance frameworks to ensure ethical responsibilities for decisions regarding development, deployment, and use of AI. For example, appointing an AI ethics officer, or forming internal and external ethics committees to provide oversight and recommendations, can be effective. Certification standards and institutional policies can also play a role.

Should maintain communication with industry or public oversight bodies, sharing best practices, discussing challenges, and reporting new issues. Such mechanisms can complement but not replace statutory oversight, such as data protection officers or similar measures required by data protection laws.

Through education and awareness-building to cultivate ethical thinking

Trustworthy AI encourages all stakeholders to be informed and actively involved. Communication, education, and training disseminate knowledge about potential impacts and empower individuals to understand that they can participate in shaping societal development.

The stakeholders include designers and developers, enterprise and individual users, as well as people who have not purchased or used the system but have still been subject to decisions made by it, and society as a whole. Basic AI literacy should be enhanced across society; one prerequisite for public education in this field is ensuring that ethical professionals in this domain receive appropriate training.

Stakeholder engagement and social dialogue

The benefits of AI should be accessible to everyone, requiring open dialogue and participation from social partners, stakeholders, and the public. Many organizations have conducted stakeholder group discussions on AI and data analytics, including legal, technical, and ethical professionals, consumer representatives, and labor representatives.

Actively seeking participation and discussion of purposes and impacts helps evaluate outcomes and methods, especially in complex situations.

Diverse and inclusive design team

The more independent the system is in carrying out tasks, the more diverse the teams responsible for design, development, testing, maintenance, deployment, and procurement should be in reflecting user and societal diversity. This helps objectively consider different perspectives, needs, and goals.

Ideal teams are not only diverse in gender, culture, and age, but also in professional backgrounds and skill sets.

Next article: Comprehensive assessment of the issue checklist , convert the requirements into actionable, itemized inquiry questions. Also align with those addressed to advanced system developers" full text of the Hiroshima AI Behavior Guidelines Comparison.

Reference and Extended Reading

"Original text: Ethics Guidelines for Trustworthy AI, April 8, 2019, European Union, printed pages 14–23." Original entry for EU Digital Strategy . This page is an unofficial English translation, edited for sectioning, terminology, and formatting; Figure 2 is a Chinese redrawn version, and the lifecycle relationship in Figure 3 is presented in text form. Substantive additions from the original text's footnotes have been integrated into the relevant paragraphs.

"yi" EU Reuse Guidelines , the original document uses " CC BY 4.0 The translated text and annotated diagrams on this page are provided under the same license, without implying endorsement by the original institution.

"Public query entry mentioned in the original text:" GDPR Original Text Verify automated decisions and the applicability conditions of data protection. Text in Chinese of the United Nations Convention on the Rights of Persons with Disabilities Explain accessibility and equal participation; EN 301 549 Standards Database Information on ICT accessibility standards and related versions is available through the above link. Access to this link does not imply that any system has been verified as compliant.