How do application examples help in understanding the framework?
This article is the English translation of the final section of the "Ethical Guidelines for Trustworthy AI" 2019 edition, corresponding to printed pages 32–38, and includes opportunities, key concerns, conclusions, and a glossary. The sections preceding this are in order. Overall framework 、 Four principles 、 Seven requirements and evaluation checklist .
Example text illustrates which developments and uses are worth encouraging, and which may contradict their values or raise specific concerns. A balance should be struck between what should be done and what is feasible, while carefully considering what should not be done. Cases, predictions, and figures are retained in the context of 2019 and are not presented as current performance metrics.
Opportunities brought by trustworthy AI
Trustworthy AI has the potential to help address urgent challenges such as an ageing population, widening social inequalities, and environmental pollution. This is also reflected in the United Nations Sustainable Development Goals. The following discusses how to encourage European AI strategies that respond to these issues.
Climate Action and Sustainable Infrastructure
Addressing climate change should be a global policy priority. Digitalization and trustworthy AI have the potential to reduce human impact on the environment, improve energy and natural resource efficiency. For example, by combining big data, energy demand can be identified more accurately, enabling energy infrastructure to better align with consumption patterns.
The original text uses projects such as smart grid and energy storage as examples, and mentions the BRIDGE initiative helping various projects develop a shared understanding of cross-domain issues; the Encompass project is also cited as an example. These represent directions and project descriptions, not equivalent to this site verifying the current performance of each project.
In public transportation, intelligent transportation systems can reduce congestion, optimize routes, assist visually impaired individuals in navigating more independently, and improve energy-efficient engine performance, thereby supporting decarbonization and reducing environmental impact. The original text uses examples from Fabulos' urban mobility solutions, PRO4VIP's visual impairment support, and UP-Drive's vehicle progressive automation and collaboration to illustrate a safer, more inclusive, and affordable transportation future.
Original citation: At the time, the World Health Organization stated that approximately one person dies in traffic accidents every 23 seconds, and it was believed that AI could reduce fatalities through faster reaction times and better adherence to rules. Note: This figure is a historical reference from the original text; it does not represent current statistics, and cannot be used to conclude that autonomous driving has achieved corresponding safety improvements. Current traffic injury data can be queried through the WHO link at the end of the article.
Health and Well-being
Trustworthy AI can make treatments smarter and more targeted, helping prevent life-threatening diseases, with some applications already in use. Medical professionals may be able to analyze complex health data more accurately and thoroughly, even providing personalized prevention before disease onset.
The original text lists REVOLVER's research on cancer evolution and individualized treatment, MURAB's exploration of more accurate biopsies and earlier diagnosis; Live INCITE calls for innovative electronic health solutions to help individuals adjust their lifestyle according to their own circumstances during the perioperative period, improving healthcare outcomes.
In response to aging, AI and robots can assist caregivers, support elderly care, and enable real-time monitoring of individuals' conditions. The original text cites the CARESSES study, which adapts robot behavior and speech according to elderly individuals' cultural and habitual preferences, as well as the ALFRED virtual assistant that helps promote physical activity; the EMPATTICS project explores how information and communication technologies, including AI, can be used collectively to plan interventions and monitor physical and mental changes.
The MyHealthAvatar project provides a digital representation of long-term health status, serving as a lifelong health partner through applications and online platforms, and exploring predictive capabilities for stroke, diabetes, cardiovascular disease, and hypertension. ENRICHME targets individuals experiencing gradual cognitive decline, integrating environmental assistive living platforms and mobile service robots for long-term monitoring and interaction, supporting independent and active living.
On a larger scale, AI can analyze medical and treatment trends, aiding in earlier disease detection, improving drug development efficiency, and enhancing targeted therapies, ultimately holding promise for saving lives. The original text also mentions Sophia Genetics using statistical inference, pattern recognition, and machine learning to process genomic and imaging data. All of the above represent potential applications or projects as originally presented and do not constitute medical advice nor independently validate their effectiveness.
Quality education and digital transformation
Technological, economic, and environmental changes require society to be more proactively prepared. Governments, industries, educational institutions, and unions have a responsibility to help the public transition into the digital age and acquire the skills needed for future work.
Trustworthy AI can more accurately predict which jobs and careers are at risk, which new roles will emerge, and the skills required, helping governments, labor unions, and industries plan skill development and job transition training, and offering new pathways for individuals concerned about job loss.
AI can also be used to reduce educational inequity by providing personalized, adaptive learning programs that enable individuals with different learning abilities to acquire new qualifications, skills, and competencies, and potentially improve learning speed and quality from basic to higher education. The original text cites examples such as the MaTHiSiS affective sensing learning environment, Watson Classroom, and Century Tech.
Key Concerns Raised by AI
When any component of trustworthy AI is violated, significant concerns may arise. Many of the following issues fall within the scope of legally binding requirements, but even if compliance is demonstrated, they may still not address all ethical concerns. Awareness of the adequacy of rules and principles may evolve, and thus this non-exhaustive list may be reduced, expanded, or updated.
Identification, tracking, and continuous monitoring of individuals
AI enables public and private entities to more efficiently identify individuals. Examples of technologies that can be widely deployed include facial recognition, and techniques that identify people based on biometric data under circumstances where individuals are not voluntarily providing such information, such as lie detection, inferring personality from micro-expressions, and automatic voice recognition. Note: The listing of certain technologies does not imply that the original text or this site validates the reliability of such inferences regarding personality, emotions, or other psychological attributes.
Automated identification may serve ethical purposes, such as detecting fraud, money laundering, or terrorist financing. However, automated identification also raises significant legal and ethical concerns, potentially having unforeseen psychological and societal impacts. To preserve autonomy, the use of such technologies must be proportionate.
It must be clearly established whether, when, and how automated identification is permitted, and it must distinguish between "identifying an individual" and "continuously tracking an individual," as well as between targeted surveillance and mass surveillance. The application of technology must be grounded in clear legal basis, with the original text pointing to GDPR Article 6.
If consent is the legal basis, practical and effective measures should enable people to give meaningful, verifiable consent to automated identification by AI. This also applies to so-called anonymous data that can be linked back to individuals. The original text notes that many online consent mechanisms lead consumers to agree without careful consideration, making it difficult to regard that consent as effective.
A system hiding its AI identity
People should know whether they are interacting directly with a person or a machine. Practitioners have a responsibility to reliably implement this, for example by providing clear and transparent indications, or by allowing individuals to request and verify whether they are interacting with AI.
There are also boundary cases, such as human speech filtered through AI. Human-machine confusion may lead to attachment, influence behavior, or reduce recognition of the unique value of humans. Therefore, humanoid robots and AI-driven avatars should undergo careful ethical evaluation.
Score of citizens violating fundamental rights
Society should protect the freedom and autonomy of all individuals. Any form of civic rating may undermine autonomy and threaten the principle of non-discrimination, and therefore should only be used when there is a clear justification, proportionate measures, and fairness.
Large-scale evaluations of an individual's "moral character" or "ethical integrity" by public or private entities across various aspects of life pose a threat to these values, especially when such evaluations violate fundamental rights, are disproportionate, or lack a clear and communicated legitimate purpose.
Descriptive scoring in narrow domains such as education, e-learning, or driver's license evaluation also requires fully transparent processes that clearly explain the rationale, purposes, and methodologies. The original wording contains a potential ambiguity; in context, it means that transparency alone cannot prevent discrimination, ensure fairness, or serve as a universal solution to scoring issues.
Whenever possible, it is preferable to allow individuals to exit without adverse consequences; otherwise, mechanisms for questioning and correcting scoring must be in place. This is especially important when there is a power imbalance. To uphold fundamental rights and where necessary in a democratic society, exit options must be ensured in technical design.
Lethal autonomous weapon systems
The 2019 original text stated that a number of unknown countries and industries are researching and developing lethal autonomous weapons, ranging from missiles capable of selecting targets to learning systems that could determine who, when, and where to attack without human intervention.
This raises fundamental ethical concerns: a potentially unprecedented and uncontrollable arms race could emerge, leading to almost complete loss of human control in military contexts, while fault risks remain unaddressed.
Supports the European Parliament's resolution of 12 September 2018 and related efforts, calling for a swift development of a common, legally binding stance on human control, oversight, accountability, the implementation of international human rights law and international humanitarian law, and military strategy, and links to the peace objectives under Article 3 of the EU treaties. This record reflects the position at the time and does not replace current legal and diplomatic developments.
Potential long-term issues
The original argues that at the time, AI development was still centered on specific domains, requiring trained researchers and engineers to clearly define objectives. Looking further out in time, one might envision certain major long-term concerns.
When adopting a risk-based approach, consideration should be given to "unknown unknowns" and black swan events. A black swan event refers to a rare occurrence with significant impact, which may have never been observed, thus making its probability estimation typically highly uncertain. Given the potentially severe consequences and uncertain development trajectories, such events should be regularly assessed.
The original footnote lists then-future concepts such as general artificial intelligence, artificial consciousness, artificial moral agents, superintelligence, and transformative AI, while noting that many people consider these to be unrealistic. Here, the original version's explanation of the disagreements is retained, and these concepts or any specific timeline are not presented as facts.
Conclusion
AI will continue to have positive impacts in business and society, but related risks and other negative effects need to be appropriately and reasonably managed.
AI has both transformative and disruptive potential. Recent advances have been driven by large volumes of digital data, progress in computing and storage capabilities, and scientific and engineering innovations in methods and tools. It will continue to impact society and the public in ways that remain unimaginable.
Precisely for this reason, trustworthy systems are needed. Only when the technology and the processes and people behind it are trustworthy can people have confidence and fully enjoy the benefits. Thus, trustworthy AI is the foundational goal of this guide.
Trustworthy AI consists of three components: legality, adhering to all applicable laws and regulations; ethical considerations, following principles and values; and robustness, avoiding unintended harm even in good intentions, both technically and socially. Each component is necessary but insufficient on its own. The ideal state is coordination and mutual overlap, with efforts to resolve conflicts when they arise.
Chapter one presents fundamental rights and related ethical principles; Chapter two lists seven requirements along with technical and non-technical implementation methods; Chapter three provides an assessment checklist to facilitate implementation. The final opportunities and concerns are intended to stimulate further discussion.
The original views the public-centered, treaty-based position of the European Union as a unique foundation of Europe, and aims to support innovation and technological development through trustworthy AI, promoting well-being at both individual and collective levels, so that everyone can benefit from AI while respecting fundamental rights, democracy, and the rule of law.
Glossary
The following terms are defined as used in this guide to aid in understanding the original text; these definitions do not represent universally accepted definitions across all fields.
Artificial Intelligence or AI systems
AI systems are software systems designed by humans, and may also include hardware. Given a complex objective, it perceives the environment by acquiring data, interprets structured or unstructured data, reasons or processes information based on the acquired knowledge, determines the best course of action to achieve the objective, and takes action within physical or digital environments.
It can use symbolic rules or learn numerical models; it can also analyze how previous actions affect the environment and adjust behavior. Humans can either directly design or use AI to optimize the design.
As a discipline, AI includes machine learning, such as deep learning and reinforcement learning; machine reasoning, such as planning, scheduling, knowledge representation and reasoning, search, and optimization; and robotics, including control, perception, sensors, actuators, and the integration of other technologies into information-physical systems. A separate document, "Definition of AI: Core Capabilities and Scientific Disciplines," provides further elaboration.
"AI Practitioner"
Development, including research, design, and data provision; deployment, including implementation, or use of AI by individuals and organizations; this guide specifically excludes individuals and organizations using systems solely in their capacity as end users or consumers.
AI System Lifecycle
Including the development phase, covering research, design, data provision, and limited testing; the deployment phase, covering implementation; and the usage phase.
Auditable
The system's algorithms, data, and design processes are capable of being evaluated. This does not mean that business models and intellectual property information must always be made public. Establishing traceability and logging mechanisms early in the design phase aids in audits.
Deviation
A bias or inclination towards people, objects, or positions. Sources of bias in AI are numerous: data collection and training, designers' perspectives on specific contexts, online learning and interactive adaptation, and personalized recommendations and information streams.
Bias is not always due to human bias or manipulated data collection; it can also arise because systems are used only in limited contexts and cannot be generalized to other situations. Bias can be beneficial or harmful, intentional or unintentional; when it leads to discrimination or unfair outcomes, the guideline refers to it as "unfair bias."
Ethics
Ethics belongs to philosophy, studying questions such as what kinds of actions are good, what value human life has, what justice is, and what a good life entails. It is typically divided into four main areas:
- Metaethics: the study of the meaning and reference of normative statements, and whether they have truth values, and how to determine them.
- Normative ethics: explores standards of right and wrong, evaluates specific actions, and determines the pathways for moral behavior.
- Ethics: empirical research on human moral behavior and beliefs.
- Applied Ethics: In facing specific, often historically novel contexts or domains of action, research has an obligation or permission to address questions about what ought to or is allowed to be done. Reality is frequently constrained by time pressure and bounded rationality; AI ethics typically falls within this domain, focusing on normative issues related to design, development, implementation, and use.
In ethical discussions, "moral" refers to specific behaviors, customs, and practices that actually exist within a particular culture, group, or individual at a given point in time; "ethical," on the other hand, denotes a systematic academic evaluation of such actions and behaviors. English translations may overlap; context should be used to distinguish between the two.
Ethical AI
Development, deployment, and use of ethical AI, including treating fundamental rights as special moral rights, as well as relevant ethical principles and core values. It is the second of the three components of trustworthy AI.
AI centered on people
Placing human value at the center of development, deployment, use, and monitoring, respecting fundamental rights, including those enshrined in EU treaties and the Charter. The common foundation is human dignity and the unique, inalienable moral status.
This also requires considering the natural environment and other forms of life within human ecosystems, and adopting a sustainable approach that supports the full development of future generations.
Red team testing
Conducted by the Red Team or an independent group from an adversarial perspective or stance, challenging the organization and enhancing effectiveness, particularly for identifying and addressing potential security vulnerabilities.
Reproducibility
Does AI exhibit the same behavior when repeated under identical conditions?
Stable AI
It includes technical robustness appropriately achieved under specific conditions such as application domain or lifecycle stage, as well as social robustness that fully considers the operational context and environment. This helps prevent unintended harm, even when intended for good, and constitutes the third component of trustworthy AI.
Stakeholders
All individuals and entities involved in research, development, design, deployment, or use of AI, or who are directly or indirectly affected by AI, including enterprises, organizations, researchers, public services and other institutions, civil society, governments, regulators, social partners, individuals, residents, workers, and consumers.
Traceability
Ability to continuously track system data, development, and deployment processes is typically achieved through documented, identifiable records.
Trust
The original study draws on research that defines trust as a belief in benevolence, capability, integrity, and predictability, and the willingness to rely on another party in a risky situation, or a combination of both.
Although "trust" is typically not considered a direct attribute of machines, this guide emphasizes that trust should not only be placed in systems that are legitimate, ethical, and robust, but also in the people and processes involved throughout the system's lifecycle.
Trustworthy AI
Also includes the three components of legitimacy, ethical compliance, and technical and social stability. Trustworthiness extends not only to the system itself but to all processes and actors throughout its lifecycle.
Vulnerable individuals and groups
Due to the diversity of circumstances, there is no universally accepted or widely agreed-upon uniform legal definition. What constitutes vulnerability often depends on the context: temporary life stages such as childhood or illness, market factors like information asymmetry or market power, economic factors such as poverty, identity factors such as gender, religion, or culture, and other factors may all play a role.
Article 21 of the Charter on grounds of non-discrimination may be referenced, including gender, race, skin color, ethnic or social origin, genetic characteristics, language, religion or belief, political or other opinions, minority status, property, place of birth, disability, age, and sexual orientation. Other laws provide specific protections for certain groups. No list is exhaustive, and the list may change over time. Vulnerable groups refer to a group of people who share one or more vulnerable characteristics.
Here completes the substantive content of the original guide. You can return to Trustworthy AI and Ethical Frameworks or use " evaluation checklist Check the specific system.
Reference and Extended Reading
"Original text: Ethics Guidelines for Trustworthy AI, April 8, 2019, European Union, printed pages 32–38." Original entry for EU Digital Strategy . This page is an unofficial English translation, with adjustments to sectioning, terminology, and layout; factual knowledge supplements from examples and footnotes are retained, and historical figures and predictions are clearly annotated. The personal identity list and administrative details on the final page are outside the scope of this knowledge translation.
"yi" EU Reuse Guidelines , the original document uses " CC BY 4.0 This page's translation is provided under the same license and does not imply endorsement by the original organization.
Query the public issues involved in the original text: WHO Road Traffic Injury For querying update statistics; UN Sustainable Development Goals Chinese Entry Point Convenient for verifying the target; European Parliament 2018 autonomous weapon resolution Verify historical positions.
Historical projects can be queried independently: BRIDGE Energy Project Collaboration 、 Fabulos Public Transportation Project 、Sophia Genetics . The entry is for understanding the original case example and does not indicate that these services have received recommendations from this site or that their effectiveness has been verified by this site.