Prof. Yucheng Zhang
AI has quietly penetrated every aspect of life (The invisible penetration and social impact of artificial intelligence) → This widespread penetration has raised ethical issues concerning human-machine collaboration, privacy, fairness, dignity, governance, and sustainability (Key ethical issues and question framework) → The root cause of these issues lies in AI’s unprecedented data and computing power, enabling it to both benefit society and potentially pose risks (Sources and dual effects of AI power) → Therefore, we need to rethink “what constitutes good technology” from an ethical perspective and clarify the importance of ethics in technological development (Necessity of ethical perspective and definition of ethics) → Finally, by introducing systematic ethical analysis methods, we can translate ethical principles into specific design and governance practices (Method and output)
| Category | Covered Topics |
|---|---|
| Human-Machine Collaboration and Decision Ethics | Healthcare, justice, education, designer responsibility |
| Privacy, Autonomy, and Human Agency | Public nature of privacy, challenges to autonomy |
| Fairness, Justice, and Trust | Fairness, justice, trustworthy AI |
| Human Dignity and Bioethics | Autonomous weapons, autonomous driving |
| Governance and Social Order | Algorithmic governance, social control |
| Sustainability and Future Ethics | AI and environment |
Sources of power: unprecedented data collection capabilities + immense computational power
Characteristics: ubiquitous, widely accepted, opaque in operation, and often ‘invisible’ in application
Dual effects: accelerating research, driving innovation, and improving life vs. bringing new concerns and challenges
Judgment: AI can solve old problems but will also generate new challenges
What is ethics: a systematic reflection on moral issues (a branch of philosophy)
More than a technical issue: to make AI beneficial to humanity, ethical analysis must be central
Meaning of morality: the totality of individuals/groups expressing ‘what is good/just’ through opinions, decisions, and actions
AI ethics takes “responsibility” as its core starting point (The close connection between AI ethics and the concept of “responsibility”) → First, clarify the basic connotations and categories of responsibility (role/moral/professional, etc.) (Basic connotations and categories of responsibility) → Then distinguish the two types of responsibility paths in technology ethics (Two types of responsibility in technology ethics) → Passive responsibility: emphasize interpretability and attributability to define fault and undertake liability (Passive responsibility (looking backward / after the fact)) → Active responsibility: embed values and protections (e.g., VSD) in the front-end of design to prevent risks and promote positive outcomes (Active responsibility (looking forward / before the fact)) → Ultimately form forward-looking guidance for technological development, while facing the dual challenges of multiple stakeholders and high uncertainty (Guiding technological development: necessity and challenges)
Core issue: Once powerful technologies emerge, who is responsible?
Context shift: As AI gains autonomy and decision-making power, the concept of responsibility needs to be re-examined
General definition: Being accountable for one’s own actions and their consequences
Role responsibility: Responsibility tied to specific roles (e.g., parent, contractual role)
Example: Moral responsibility vs. professional responsibility
Moral responsibility: Obligations, norms, and duties based on moral reasoning
Professional responsibility: Grounded in the professional role and must not exceed moral boundaries
Passive responsibility
Active responsibility
Applicable scenario: Accountability after an undesirable event
Example: In an autonomous driving collision — who is responsible: the driver, the designer, or the manufacturer?
Two key elements:
Explainability: The ability to explain decisions and actions
Blameworthiness: Being a legitimate target of blame, based on common criteria such as:
Misconduct: e.g., engineers failed to test the recognition system adequately
Causal contribution: acts or omissions that causally contributed to the event (e.g., premature system rollout)
Foreseeability: the harm was foreseeable (e.g., failure to test sufficiently caused preventable collisions)
Freedom of action: acting without coercion; even under corporate pressure, autonomy implies moral agency
Core idea: Prevent negative consequences and promote positive outcomes through proactive action
Significance: Especially crucial in the design stage of impactful technologies like AI
Methodological transformation: Integrating values and ethics into early design and embedding them into technology
Value-Sensitive Design (VSD): A systematic approach that treats moral values as design requirements
Example: Designing autonomous vehicles with privacy protection as a guiding principle
Necessity: Providing directional guidance for technological development
Two main challenges:
Current situation: Multiple response strategies exist, but continuous improvement is required
Taking “autonomous driving ethics” as the entry point, we first clarify the scope of discussion and core concept - Operational Autonomy, which refers to the system’s ability to perform tasks without human intervention (Topic positioning and definition of “Operational Autonomy”) → Then evaluate its moral pros and cons from a macro perspective: the improvement of safety and efficiency vs. the risks to privacy, freedom and responsibility (Moral advantages and disadvantages overview) → Focus on the key scenario - the “inevitable crash” issue, and reveal that AI faces real moral choices (Key scenario: inevitable crash) → Guide students to engage in ethical thinking through two typical dilemmas (e.g., “crash into the helmeted person or the unhelmeted person”) (Group discussion — two typical moral dilemmas) → Further analyze using different ethical frameworks: Consequentialism pursues minimal harm, while Deontology emphasizes moral principles and the boundary of rights (Analysis under different ethical frameworks) → Then discuss the complexity of responsibility attribution: the breakage in the responsibility chain among manufacturers, programmers and car owners (Challenges in assigning responsibility) → Furthermore, demonstrate how ethics influences policy-making, taking the principles of the German Ethics Commission as an example to transform theory into governance practice (From ethics to policy: The German case (2016-2017)) → Finally, summarize: Ethical analysis does not provide a single answer, but helps society strike a balance between innovation and responsibility (Summary).
Discussion value: Use autonomous driving ethics as a lens to explore AI-related issues through multiple frameworks
Operational autonomy: The ability of a system to perform tasks without continuous human supervision, adapt to dynamic environments, and respond to unexpected situations; corresponds to levels of autonomous driving (L1–L5):
L1: Automation of basic functions such as acceleration and cruise control
L2–L4: Gradual enhancement of perception, decision-making, and control across scenarios
L5: Full autonomy under all conditions
Pros: Improve traffic safety (more law-abiding, not fatigued, not intoxicated, not distracted); increase mobility and autonomy for vulnerable groups (e.g., the elderly); improve efficiency and reduce pollution
Cons / Risks: Potential violations of privacy and freedom; unclear allocation of responsibility
Definition: Situations where a vehicle cannot completely avoid colliding with an obstacle
Ethical delegation: Beyond operational control, moral choices (e.g., collision angle, passenger protection priority) are delegated to the system
Scenario 1: Two cyclists crossing the road (A wearing a helmet / B without); the vehicle can only hit one but may choose the target
Scenario 2: Two pedestrians crossing ahead; options — collide with pedestrians, or swerve to avoid them but risk severe passenger injury
Consequentialism (Utilitarianism)
Principle: Moral value is judged by consequences; choose the option that benefits the majority or minimizes harm
Application: In inevitable crashes, “fairly minimize risk and damage”
Scenario 1 inference: The system might be programmed to hit the helmeted cyclist (expecting less harm) — a controversial choice that seems to “punish” compliance
Deontological Ethics
Principle: The morality of an act depends on moral rules, rights,
or duties, independent of outcomes
Application: For instance, a mother driving her child daily — prioritizing the child’s protection can be viewed as a separate moral duty
Tension: It may conflict with the rights of others and universal social rules (e.g., does “Passengers first” harm pedestrians’ rights?) Challenges in Assigning Responsibility
Potential responsible parties: Manufacturers, design engineers, vehicle owners
Source of difficulty: No one has “complete and direct control” of the system — responsibility chains are fragmented or overlapping
Academic approaches: Multiple frameworks coexist, depending on ethical and institutional choices; the “policy vacuum” must be addressed
Exploring “Autonomous Weapon Systems (AWS)” from an ethical perspective: First, clarify its definition and categories — weapon platforms with autonomous navigation, identification, and attack capabilities (Definition and categories) → Then reveal three driving factors: operational efficiency, economic costs, and strategic advantages, which collectively promote its rapid development (Three drivers) → Further distinguish the key differences between remotely piloted drones and fully autonomous weapons, pointing out that the chain of responsibility tends to be vague in the latter (Key distinction: Remotely piloted drones vs. Autonomous weapons) → Subsequently, introduce international debates on use and regulation: proponents emphasize reducing human errors, while opponents warn of out-of-control risks and ethical hazards (Debate on use and regulation) → The use of autonomous weapons is not completely prohibited; the focus of the debate lies in how to restrict and manage them (Regulatory perspectives and policy positions) → At the theoretical level, deontology emphasizes human dignity and moral responsibility, advocating that “life-or-death decisions” should be reserved for humans (Deontological argument (Ethics of duty)) → However, real cases show the dilemma of the responsibility chain: engineers, commanders, and algorithm designers may all be partially responsible, but no one assumes full responsibility (The chain of responsibility dilemma (example)) → In contrast, consequentialists believe that if autonomous weapons can reduce casualties and mistaken killings, they may have positive value (Consequentialist argument) → Finally, the concept of “Meaningful Human Control” is proposed, advocating the retention of sufficient information and intervention opportunities to ensure the continuity of ethical and legal responsibilities (Meaningful human control)
Autonomous weapon systems: Capable of navigation, investigation, and attack without continuous human supervision
Categories: Aerial systems, robotic sentries, loitering munitions, and other similar platforms
Technological evolution: Group intelligent weapons — large numbers of small, inexpensive platforms collaborating to execute complex missions
Operational: Unmanned or automated systems can act faster and operate in high-risk environments
Economic: Reduced need for human operators—–lowers operational costs
Strategic: Enhanced military capability and altered geopolitical dynamics
Remotely piloted drones: Human-operated, maintaining a command-and-control chain that allows clear attribution of responsibility
Autonomous weapons: Systems capable of independently selecting and engaging targets — human control is diminished, and responsibility chains become blurred
Support (Ronald Arkin): Believes that properly designed autonomous weapons could reduce human battlefield misconduct and improve adherence to international law
Opposition (Noel Sharkey): Argues that such systems lack verifiable safeguards; unpredictable machine behavior could exceed human oversight
2021/03: The US AI National Security Commission: Does Not Support Complete Ban on Autonomous Weapon Systems Capable of AI
Debate focus: Whether and how to restrict or regulate autonomous weapons
Basic obligation: Following the principle of human dignity (THL); emphasize the moral responsibility chain and uphold respect for humanity
Dignity requires: The decision to take life should remain a uniquely human moral act
Proportionality principle: Only attack targets that bring about a military advantage; it is necessary to distinguish between those who are still in active combat and those who have surrendered.
Challenge: How can AI identify “intentions of surrender”? How can it avoid making incorrect perceptions?
Case of Mistakenly Attacking a School Bus: There are numerous potential responsible parties (engineers, managers, procurement decision-makers, consultants, manual writers, commanders, operating soldiers…)
The essence of the problem: No one has a “significant causal contribution”, leading to a gap in attribution
Supporters argue: Replacing human soldiers could reduce casualties and psychological harm; strict launch conditions can be set
Critics/warn:
Background of proposal: Rethinking responsibility in uncertain military environments
Connotation: “Meaningful” means having sufficient information and time for human intervention
Current status: How to operationalize it remains controversial, but the focus has shifted from “defining autonomy” to “establishing control and supervision”
Regarding the ethics of digital healthcare, we first point out the new opportunities and risks brought by AI in the medical field: Algorithms promote precision diagnosis and public health, but also bring challenges related to privacy, bias, and trust (Emerging opportunities and challenges) → Then sort out the four major application scenarios of digital health: health monitoring, auxiliary diagnosis, treatment decision prediction, and public health governance, revealing that AI runs through the entire medical process (Four types of applications in digital health/digital medicine) → Focus on AI in diagnosis, emphasizing that its potential lies in improving accuracy, reducing the burden on doctors, and expanding preventive coverage (AI in diagnosis: potential and objectives) → However, there are three core ethical focuses at the same time: privacy and data protection, algorithm representativeness and fairness, transparency and interpretability (Three major ethical focuses) → Therefore, it is necessary to rely on regulations and international guidelines as guarantees. Currently, initiatives from GDPR to OECD/WHO all emphasize human control, privacy minimization, and responsible application (The current situation of regulations and guidelines) → Finally, summarize: Digital healthcare is reshaping medical practice and the doctor-patient relationship. In the future, we should adhere to the principle that “medicine is a social practice” and maintain an ethical balance between technological innovation and patients’ rights (Summary and outlook)
IT/AI-enabled diagnosis and treatment (data-driven medical decision-making)
Moral Challenges: Global Public Health Policies, Transparency and Credibility, Privacy Protection
Health monitoring (wearable devices / mobile apps / home-based equipment)
Auxiliary diagnosis (image and clinical data analysis)
Predictive algorithm for treatment decisions
Public health applications (epidemiological monitoring, policy tools)
Machine learning is highly effective in clinical/imaging analysis, and sometimes outperforms humans.
Objective: Improve accuracy, reduce burden on specialists, and expand prevention coverage
Privacy and Data Protection
Based on large-scale collection/storage/analysis/distribution
Informed consent is challenged in the era of big data (as algorithms can infer undisclosed information)
The applicability/representativeness of algorithms in medicine
Training requires representative data
Biomedical data is biased towards white men (in North America/Europe) and harms disadvantaged/underrepresented groups, violating justice
Transparency and Explainability
ML has a lot of “internal correlation” learning, and the parameters are not explainable (it is a “black box”)
Challenges Arise in Determining Doctoral Duties, Patient Informed Consent, and Legal and Ethical Responsibilities
GDPR (2018): Granting the right to refuse fully automated decision-making
The current guidelines are mostly abstract; the OECD/WHO has proposed more specific suggestions:
Protecting privacy and minimizing negative impacts
Digital solutions support rather than replace clinical practice
Human control must always be ensured
Digital medicine is reshaping medical practice and medical perspectives
The tension between prediction and explanation (correlation vs. causation) remains unresolved
The future should be guided by the principle that “medicine is a shared social practice”, and patient rights (both clinical and ethical) should be protected in two dimensions
Turning to the ethics of AI and the environment, this section starts with the “duality” of AI and the environment, pointing out that AI can not only promote sustainable development but also may increase resource consumption and environmental burdens (Overview: The duality of AI and the environment) → Then reveal the contradictions within the sustainability framework: the digital revolution supports the UN SDGs, but harms vulnerable groups due to insufficient data representativeness (The framework of sustainability and its contradictions) → Subsequently, explore the measurement dilemma and the metaphorical trap of “cloud computing” — the cloud is not intangible, but a physical infrastructure that consumes enormous energy and land resources (The measurement dilemma and the metaphorical trap of “The cloud”) → Therefore, ethical evaluation should go beyond quantitative calculation and shift to value measurement, integrating multi-dimensional consequences such as environment, education, and health (Ethical assessment: Beyond quantification to value) → Further discuss distributive justice: the unequal flow of data centers and electronic waste makes poor areas bear environmental and health costs (Justice and distribution) → Facing enterprises’ “green AI” commitments and carbon neutrality goals, we also need to be alert to the risk of “greenwashing” and advocate for transparent and verifiable sustainable actions (Initiatives, Goals, and ‘Greenwashing’ risks) → Finally, summarize: AI sustainability is a systematic issue involving technology, ethics, and politics, requiring multi-level responsibilities and global cooperation to achieve truly green intelligence (Summary)
Opportunities: Smart grid, satellite data monitoring of deforestation, contributing to SDGs, etc.
Risk: The processing of massive data and the disposal of equipment cause environmental and social impacts.
Core: The Global Connection between the Well-being of Current/Future Generations and Ecosystems
2030 Sustainable Development Goals (SDGs)：The digital revolution can play a crucial role, but the lack of representative data has disadvantaged vulnerable groups. This reflects an inherent contradiction.
High energy consumption and significant emissions; “The Cloud” requires physical infrastructure (land occupation, energy consumption, emissions)
The energy consumption of Alil is significantly reduced (for instance, the training of large-scale NLP models can exceed several times the emissions of an entire vehicle throughout its lifecycle).
Consequentialist approach: Assess the costs and benefits (including comprehensive consequences such as environmental, educational, health, and energy consumption aspects)
Limitations:
Direct and Indirect Influences & Relative Pollution
It is necessary to assess absolute and relative pollution (compared with social/environmental benefits)
Note the indirect effect: For instance, remote working reduces the need for company vehicles but increases the demand for personal vehicles at home.
Resource Utilization and Community Impact: The Effects of Data Centers on Land/Energy and Local Political and Social Structures
Disposal of justice; Electronic waste mostly flows to poor countries, with insufficient regulation enforcement - health/environmental hazards
Governance requirements: Decisions related to urban space must be transparent, democratic, and subject to public supervision.
One of the trustworthy AI standards: Environmental Sustainability (High-Level Expert Group on AI of the European Union)
Practical challenges: Difficulty in measurement and implementation; most goals are voluntary - with limited impact
Corporate carbon neutrality/net zero commitments often lack detailed implementation plans - there is a risk of “greenwashing”
Improving digital pollution and enhancing the sustainability of AI are comprehensive issues involving technology, ethics, and politics
It is necessary to establish common moral values and implement multi-level responsibilities (individuals, countries, and the international community).
It is necessary to adopt a global perspective and go beyond the simplistic narrative that “energy consumption can be reduced simply by optimizing technology”
Moving on to the ethics of cybersecurity, this section first defines basic concepts and “the three elements” — Confidentiality, Integrity, and Availability, which constitute the core pillars of information security (Basic definitions and “The three elements of CIA”) → Then point out that security issues are not only a technical bottom line but also ethical issues, because protecting data means protecting human rights and social order (The connection between safety and ethics) → Subsequently, discuss responsibility and liability: if risks arise due to failure to take protective measures, corresponding moral and legal responsibilities should be borne (Responsibility and liability) → Further reveal value trade-offs: strengthening integrity and monitoring may sacrifice confidentiality, while enhancing privacy may reduce availability (Value trade-offs (the tension between the CIA)) → At the social level, this trade-off extends to the conflict between the government’s security power and citizens’ privacy and freedom, such as ethical reflections after anti-terrorism monitoring (Social-level trade-off: Government power and privacy) → At the same time, it raises the moral controversy of “ethical hackers” and vulnerability disclosure: how to achieve legitimacy between revealing risks and violating privacy (“Ethical hackers” and disclosure ethics) → In the AI context, it is necessary to combine cybersecurity with the four principles of trustworthy AI to ensure the parallel implementation of harm prevention, respect for autonomy, fairness, and interpretability (Cybersecurity in the AI context and the principles of trustworthy AI) → Finally, summarize: Cybersecurity is both a technical issue and an ethical issue. A balance should be achieved between the three CIA elements and human-centric values to build a responsible and trustworthy digital society (1.2.5.8 Summary).
Confidentiality：Access is granted only to the authorized person
Integrity：Prevent unauthorized tampering
Availability：The authorizer can access it at any time and it will not be denied
Bottom line: Unauthorized access in itself is inappropriate.
On a deeper level: Safety measures will shape technology and social order
Who is responsible for the security vulnerability? If no protective measures were taken, should they bear some responsibility?
Ethical Focus: The Responsibility of Actors for the Risks of Others
The solution for enhancing integrity (global monitoring) may sacrifice confidentiality
At the same time, a solution that ensures high confidentiality and integrity may reduce availability (complex verification)
Authorization penetration can be used in security testing, but the method of disclosure raises ethical concerns
Four types of arguments and counterarguments:
Preventing harm: If the well-being and rights cannot be protected, the plan is ethically inappropriate; it needs to be balanced with usability
Respecting human autonomy: Excessive reliance on autonomous AI leads to degradation of human skills and impairment of recognition and decision-making abilities
Fairness: Automated intrusion detection may undermine fairness due to bias or errors
Explainability: The results of black-box ML are difficult to explain, thus meaningful human control may fail
Shifting focus to AI’s role in human decision-making, AI has begun to replace humans in various human decision-making scenarios (Application scenarios and method evolution) → Its operation follows an automated reasoning process of “learning-modeling-prediction” (The learning–modeling–prediction process of machine learning) → This process brings efficiency in recommendation systems and high-risk decision-making, but also triggers ethical controversies (Recommendations and controversial applications) → The fundamental issue lies in the fact that algorithms are not neutral; instead, they embed biases and limitations (‘Algorithmic positive bias’ and model limitations) → Therefore, we need to establish normative principles to ensure that AI decisions conform to social, legal, and ethical standards (Normative conclusion and transition)
Application fields: Automated recruitment, business crime prediction tools, medical diagnosis and treatment recommendations, etc
Automated Management Decision System: Processes data related to employees, integrating human and machine decision-making
Historical comparison
The core of learning: Adapting to new situations from experience, without explicit instructions
Model Construction: By training with large-scale data, a mathematical model can be derived, where the input data will be mapped until the most likely answer emerges
Classification as a learning paradigm
Decisions are classified: Many real-life decisions can be expressed as classification problems
Simple decision-making: Recommendation system (for recruitment interviews, etc.), based on the “similarity principle” to learn user preferences
Controversial decision-making case
Myth: Mathematics/Statistics are not equivalent to neutrality and objectivity
Reality: A model is an approximation of reality, involving assumptions/simplifications/errors
Representativeness and Bias: If the generation mechanism of the training data is flawed, the model will mistake spurious correlations for genuine associations
Overfitting and failure of generalization
Phenomenon: The training set performs well, but the test/new scenarios perform poorly
Metaphor: Focusing only on depicting a small part of the city map, resulting in the rest of the area being useless
Now focusing on privacy ethics in the AI era, this section starts with privacy issues in the age of AI. First, it explains why privacy has become a core ethical focus against the backdrop of growing intelligence and the popularization of surveillance (Why discuss privacy in the era of AI) → Then distinguish the two connotations of privacy: “freedom from interference” at the decision-making level and “data access control” at the information level (The two connotations of privacy) → Further draw on Deborah Johnson’s ideas to put forward key reflections: What is the value of privacy? What will we lose if we lose privacy? And what kind of beings will individuals be shaped into in a society under constant surveillance (Key questions (Inspired by Deborah Johnson)) → Subsequently, analyze the characteristics of information flow in contemporary organizations: large-scale, diverse in type, easy to spread, and difficult to delete. Misinformation will be amplified rapidly, leading to systematic privacy risks (Characteristics of information flow in contemporary organizations) → Then through common arguments against privacy (such as “those with nothing to hide don’t need privacy” or “data sharing benefits everyone”) and their rebuttals, explain that privacy is not only about secrets, but also about fairness, dignity, and autonomy (Common arguments against privacy and their rebuttals) → Finally, summarize: In an information society dominated by AI, privacy is the cornerstone of safeguarding individual freedom and social justice (Summary)
Decision/Constitutional-level Privacy: The freedom to make decisions without interference in private matters
Privacy at the information/infringement level: Control over “access to one’s own information” (such as social disclosure)
What is the value of privacy? What will be lost if privacy is lost?
In a surveillance society, how will we be shaped as individuals?
Scale: The electronic records of employees make it easier to collect/store/share, which leads to an explosive increase in the volume of information
Information type: Includes new types of personal data such as transaction generation information
Distribution: Internal organizational data can be bought/sold/misappropriated/stealed
Persistence: Insufficient deletion incentives in digital storage
Error amplification: Errors within the organization spread rapidly and widely, making it difficult to fully trace them
“Only departments with secrets need privacy.”
The employees have given up their privacy, resulting in it becoming worthless
Collecting data is beneficial for both organizations and individuals
This section continues to deepen the discussion on privacy ethics, pointing out that privacy is inseparable from fundamental values such as autonomy, equality, and democracy (The interweaving of privacy and fundamental values) → Secondly, from the individual perspective, privacy is the “individual well-being” of employees. It ensures the diversity of relationships between people and enables individuals to freely express themselves in different scenarios (Privacy as an “individual well-being” of employees) → From the organizational perspective, privacy is also a kind of “social well-being”. Priscilla Regan points out that in organizations with a “panopticon-style” surveillance system, continuous observation will weaken autonomy and democratic participation, leading to institutional obedience and the loss of critical ability (Privacy as “social well-being” within an organization (Priscilla Regan)) → Therefore, multi-level response strategies must be adopted: establish a multiple responsibility system involving individuals, organizations, and regulators through industry ethical guidelines, transparent policies, and technical approaches such as “Privacy by Design” (Response strategies and multiple responsibilities) → Finally, summarize: Privacy is not only an individual right, but also a common value for maintaining organizational health and social democracy, which should be jointly protected through the coordination of systems and technologies (Summary)
Rachels (1975)：Privacy maintains the diversity of relationships (asymmetric information shapes different types of relationships)
Problem: Often finds itself in a disadvantaged position when balancing against social welfare (such as in the case of the “Patriot Act”)
Professional associations: Ethical guidelines and industry standards, promoting fair handling
Regulation and Policy: Shift from default consent / only opt-out to active opt-in; Transparent Policy
Technical Design: Privacy by Design, determining the minimization of data collection/transmission from the very beginning
This section explores “the moralization of technology”. First, it points out that technology itself is not neutral; instead, it carries the social power structure and value orientation, and design is politics (Technology embodies morality and politics) → Then put forward the concept of “technological mediation”: technology is not a mere tool, but a medium that affects human behavior, perception, and judgment (Technological mediation) → This leads to the connotation of “technological moralization”, that is, actively embedding moral goals in the design stage so that technology can guide humans to behave in a more ethical manner (The connotation of technological moralization) → However, this process has two major underlying problems: first, the result of the interaction between technology and humans is highly uncertain; second, the values in the program are often invisible and difficult to perceive (Two major underlying problems) → Examples show that although moralized technologies such as drunk driving lock systems and speed bumps are well-intentioned, they often encounter three types of backlash — restricted freedom, degraded moral ability, and lack of democratic legitimacy in value embedding (Examples of moralized technologies and three types of backlash) → Therefore, it is necessary to explore how to design ethical AI in a democratic way: emphasize transparency, public participation, and value consensus to prevent negative effects caused by “unexpected mediation” (How to design ethical AI in a democratic manner) → At the same time, clarify the division of labor among multiple parties: designers need to predict technical consequences, users should understand the embedded values, and policy makers need to shift from ex post supervision to ex ante co-creation (Roles and division of labor) → Finally, summarize: Moralized design is a process jointly shaped by humans and technology. Only through open, interpretable, and collaborative ethical design can responsible AI innovation be achieved (Summary)
Winner：The “racist overpass” designed by Robert Moses - the low bridge with height restrictions to exclude buses, determined who could access the beach
Latour：Artificial objects are the bearers of morality (speed bumps “make the moral decision for you” to slow down)
The technology is not a neutral method, but rather a mediator that influences actions and perceptions
Case: Obstetric Ultrasound
Separate the fetus from the maternal representation and consider it as an “independent living entity”
Incorporating pregnancy into medical guidelines: The fetus becomes a “potential patient”, and defects become “preventable suffering”
Definition: The deliberate development of technologies for the purpose of shaping moral actions or decisions
The relationship with proactive responsibility: Preempting negative outcomes in advance and facilitating positive effects
AI relevance: Al has a significant impact on humans and society. The proactive approach is particularly crucial during the design stage
Uncertainty: The introduction of technology and human-machine interaction is uncertain. AI is often regarded as an “experimental technology”
The three types of invisibility problems (Jim Moor)
Alcohol lock for cars (locks the car when alcohol is detected) - Even without concerns over accuracy and privacy, the sales resistance remains high
The three main reasons why people are opposed to technologically moralized systems
How to Democratize Ethical Design: Transparency, public discourse in the process of value embedding, and avoiding “unexpected mediation”
Counter example: Rebound effect of energy-efficient light bulbs (installed in more locations, leading to increased total energy consumption)
Insight: Design = materialization of morality; there is a need to re-examine the moral responsibility of designers
Designers: Should not only focus on “desired morality” but also need to predict mediation effects
Users/Citizens: Should know who is embedding which values
Policymakers: Should not only conduct ex-post regulation but also proactively participate in co-shaping technology
Focusing on “Trustworthy AI”, this section first introduces its background and conceptual positioning — derived from Europe’s Ethical Guidelines for Trustworthy AI, emphasizing that AI should not only be technically reliable but also trustworthy in ethical and social dimensions (Background and conceptual orientation) → Then summarize the four core ethical principles: Human Agency and Autonomy, Prevention of Harm, Fairness, and Interpretability, which form the value foundation for evaluating Trustworthy AI (Four ethical principles (European guidelines)) → Seven implementable ethical requirements are derived from these four principles, including Human Oversight, Technical Robustness and Safety, Privacy and Data Governance, Transparency, Diversity and Fairness, Social and Environmental Well-being, and Accountability Systems, providing an implementation framework for AI governance (Seven implementable ethical requirements (Derived from the four principles)) → Finally, point out the key insight: Although technical performance (such as accuracy and robustness) is a necessary condition, it is not sufficient to ensure AI’s trustworthiness; true trust also depends on the implementation of ethical principles and the balance of social responsibilities (Key insight: Trustworthiness is not solely determined by technical indicators)
Issue Prominence：Trustworthy AI is a focal topic in both academic and public discourse. It underpins the Ethics Guidelines for Trustworthy AI in Europe(2019) , which further serves as the ethical foundation for the AI Act (a legally binding instrument that provides unified supervision)
Difference between “Trustworthiness” and Human Trustworthiness:
Human agency & autonomy
Prevention of harm
Fairness
Explainability
Human Agency and Oversight
Technical Robustness and Safety
Privacy and Data Governance
Transparency (to be elaborated later): Transparency in data, systems, and business models
Diversity, Non-Discrimination, and Fairness (to be elaborated later): Focus on bias identification: mitigation and assessment
Social and environmental well-being
Accountability system
Internal and external audits can assess the system; incidents/ negative impacts should be publicly reported
Clarify the responsible entities and pay attention to vulnerable groups; Improve the remedial mechanisms
Technical stability/safety is a necessary but insufficient condition
Large Language Model (LLM/ChatGPT) Challenges：
Conclusion: The trustworthy AI framework enables us to balance ethical and social dimensions during the design and deployment stages, rather than focusing solely on performance
Focusing on “Algorithmic Bias and Fairness”, this section first reveals the facts and causes of bias in AI: Algorithms are not neutral; instead, they inherit and amplify social inequalities and structural biases embedded in training data (Facts and causes of bias in AI) → Then classify the types and sources of bias, including data bias (incomplete samples, cultural exclusivity) and label bias (incorrect annotation, crowdsourcing errors) (Types and sources of bias) → Next, elaborate on the multiple definitions and paradigms of fairness: From the perspective of distributive justice, it includes both “equality” (equal opportunities) and “impartiality” (compensation based on differences); at the technical level, it is further divided into two dimensions: “individual fairness” and “group fairness” (Definition and paradigms of fairness) → Further explain the three technical links of fairness intervention: the data preprocessing stage (cleaning and resampling), the model training stage (introducing fairness constraints or regularization), and the result post-processing stage (threshold adjustment or re-ranking), which together form a full-process path for algorithmic fairness (Involvement stage: Three types of technical approaches) → Finally, summarize: AI fairness is not only a technical challenge but also a reflection of social values; only through the parallel implementation of institutional governance and technical means can a trustworthy AI system that balances efficiency and justice be established (Summary)
Technology is not neutral: AI often inherits and amplifies biases in data and algorithms
Example: Recruitment training data reflects historical reality (most managers are male) → The model favors men, exacerbating gender bias
Statistical Perspective
“Biased Algorithm”: Systematic error in automated processes, producing unfair results for a certain group (e.g., systematically excluding applicants under 25)
Data Bias
Label Bias：
Amplification Effect When Targeting Humans: In arease Face Recognition, spreads across domains like advertising, security, soft biometrics, human-computer interaction, and healthcare; classification + inference decisions impact life outcomes
Fairness = non-discrimination (intuitive at the top level), but there is no single definition; multiple frameworks coexist
Distribution Principles
Object Levels
Preprocessing Fairness (Data Layer): Cleaning/Resampling/De-sensitization/Causal Re-weighting, etc.; Example: Before training the recruitment model, remove or correct the gender signals that cause discrimination
Fairness in the processing (model layer): During training, fairness constraints/regularization are introduced to achieve multi-objective optimization of accuracy and fairness
Post-processing fairness (result layer): Adjusting thresholds, reordering, or setting quotas for the output; for example: If the number of female hires is significantly lower than that of males, the prediction will increase the selection rate of females to the target range.
This section explores the ethical implications of “explainability and opacity”. First, it clarifies the core concept: Explainability refers to whether the process of a model from input to output is understandable, and its ethical significance lies in ensuring traceable responsibility, identifiable bias, and establishable trust (Concepts and ethical importance) → Further clarify its value positioning:Explainability is an “instrumental value”, which aims to serve higher-level intrinsic values such as safety, fairness, and autonomy, rather than being an end in itself (Value positioning: Instrumental value, not instrument itself) → Then demonstrate the trade-offs between explainability and other values through contextual cases, such as excessive information in autonomous driving may interfere with safety, and “black-box models” in medical triage, although accurate, undermine informed consent—indicating that the degree of interpretation depends on risks and contexts (Examples of contextual trade-offs) → Finally, put forward a practical orientation: The “dynamic balance” between explainability and performance should be achieved based on risk levels and usage scenarios. By combining mechanisms such as auditability, traceability, and human-in-the-loop, AI can be both understandable and effective (Practical orientation: How to ‘strike a balance’)
Explainability: Whether the process and key factors from input to output can be explained.
Black Box Problem: The input-output pathway of complex models (e.g., deep neural networks/LLMs) is opaque, even experts struggle to interpret it
Why It Matters:
Legal/Ethical Risks: High-stakes decisions require justifiability and accountability; black boxes undermine responsibility attribution and error correction
Bias Hard to Detect: Opacity combined with historical biases makes errors difficult to locate, endangering the safety of underrepresented groups
Trust Eroded: Unexplainable processes are perceived as untrustworthy, triggering suspicion and resistance
Instrumental Value: Helps achieve intrinsic values (e.g., autonomy, safety, fairness, trustworthiness)
Inevitability of Trade-offs: Explainability needs to be contextually balanced with other values. One cannot neglect intrinsic values like safety/effectiveness “in the name of explainability”
Autonomous Driving
Hospital Triage (ICU Allocation)
Clarify usage scenarios and risk levels (higher risk → higher requirements for explanation and auditing)
Integrate auditability, traceability, interrogability, and human-in-the-loop mechanisms
Recognizing that explanatory improvements do not equal a panacea, additional measures such as bias management, data governance, and security safeguards are still necessary
This section summarizes the paradigm shift in AI ethics. First, it explains the overall context: AI ethics has shifted from focusing on single issues (such as privacy or fairness) to systematic reflection, requiring collaborative governance at the technical, social, and institutional levels (Overall context) → Then point out the three types of key actors and their responsibilities in the paradigm shift: designers and engineering teams, users and citizens, and policy makers (Three types of actors and their responsibilities in paradigm shift) → Finally, put forward the methodological positioning: Ethics should not be regarded as a “list of answers”, but a continuous process of reflection and creation, helping society enhance its judgment and governance capabilities in the complex and uncertain AI ecosystem (Methodological positioning of ethics)
Expansion of the Issue Spectrum: From traditional issues (privacy, responsibility) to new topics (sustainability, digital medicine, etc.)
Conclusion Highlights: AI ethical issues cannot be solved by technology alone; a socio-technical perspective is needed (technical solutions must be embedded in broader ethical and social frameworks)
Designers/Engineering Teams: Should not directly embed “desired moral forms” into technology; instead, they need to anticipate technological mediation and employ moral imagination
Users and Citizens: Should recognize “who is embedding which values into AI” and foster public supervision and participation
Policymakers/Regulators: Should not only conduct ex-post supervision but also co-shape technology, participate from the outset, and promote public discourse
This section marks the starting point of the AI Introduction course. Firstly, it points out the contemporary significance of AI and its latest representative achievements—from deep learning to large language models, AI has become a core driving force behind technological, economic, and social transformations (The significance of AI in the era and its recent representative achievements)→ Then it explains the difficulty in defining AI: the boundaries of AI continue to evolve with technological advancement, so this course regards AI as an interdisciplinary “dynamic concept” and focuses on its technological logic and social impacts (The difficulty in defining “What is AI” and the positioning of this course)→ Next, it provides an intuitive definition—“Judging intelligence by behavior”, which means evaluating intelligence by observing whether a machine can demonstrate human-like cognition or decision-making in complex environments (Intuitive definition and the method of ‘Judging intelligence by behavior’)→ Subsequently, it introduces the “AI effect” and Larry Tesler’s view: once a technology is widely mastered, people no longer consider it part of AI, reflecting the relativity and social constructiveness of “intelligence” (The ‘AI effect’ and Larry Tesler’s view)→ Finally, it previews the subsequent structure: the course will take the four historical stages of AI development as the main thread, sort out the evolutionary path from symbolism to machine learning, and then to modern generative AI, laying a historical context for understanding AI ethics and social impacts (Subsequent structure: Historical stages of four periods)
Regarded as one of the most fascinating technologies of this century and will dominate at least until 2050
Representative Milestones and Systems:
Nowadays, everyone talks about AI, but it is not easy to give a strict answer
This course: Traces the evolutionary history of definitions and their impact on society; here only outlines concepts and proposes an analytical framework
Intuitive understanding: Endowing machines with the ability of “intelligence” (distinguishing from the natural intelligence of humans/animals)
Dilemma: Lack of a unified definition of “intelligence” → Adopt “observable behavior + objective function” for judgment
Measure and optimize with formal indicators (relying on mathematics/logic):
This section unfolds around the “intellectual origins of AI”. Firstly, it traces the early goals and focus areas—humans have long attempted to simulate reasoning, computation, and intelligent behaviors through mechanical means (Early goals and focus areas)→ Then it introduces the three major milestones in mechanical computing in the 17th century, such as Pascal’s Adding Machine and Leibniz’s Step Reckoner, demonstrating the initial attempts to “replace human cognitive operations with machines” (Three milestones in mechanical computing in the 17th century)→ Entering the 19th century, the ideological foundation before Turing had emerged: Babbage’s Analytical Engine and Ada Lovelace’s concept of programming made “programmability” the core idea of intelligent machines (Emergence of programmable ideas in the 19th century)→ By the mid-20th century, the birth of electronic computers and the theoretical foundation of Turing’s Universal Machine marked the possibility of “general intelligent computing” in the true sense (Mid-20th century: From electronic computing to general computing theory)→ During the same period, McCulloch and Pitts proposed the first generation of artificial neuron models (1943), igniting another intelligent paradigm inspired by biological cognition (The first spark of neural networks (1943))→ Finally, it elevates to the conceptual level: these technological veins collectively lead to an ultimate question—what is the essence of intelligence? Based on what criteria should we judge “whether a machine is thinking”? (Elevated discussion: The essence of intelligence and criteria for judgment)
Humans have been attempting to create machines that can “replicate” intelligence for hundreds of years
Early exploration of machines and theories focusing on “computing power”
1623 Schickard: Mechanical calculation within six digits (requiring human intervention)
1642 Pascal: Mechanical adder-subtractor (multiplication and division realized through repeated addition and subtraction)
1674 Leibniz: Recursive mechanism for addition/subtraction/multiplication
Babbage’s “Difference Engine”: Solving polynomials with the finite difference method (materialized by the Scheutz father-son team)
Babbage’s “Analytical Engine”: A general computing concept that surpasses the difference engine
Turing: Proposal of the Turing Machine — simulating any mathematical reasoning with a 0/1 symbol system
Church & Turing: Digital computers can simulate any formal reasoning; the Turing Machine is a universal model
Convergence with neurobiology, information theory, and cybernetics → the “electronic brain” concept
This section focuses on the ideological logic of the Turing Test. Firstly, it introduces its “problem restatement strategy”: Turing transformed the originally intractable question “Can machines think?” into a more operable one—“Can machines behave like humans in conversations?” (Problem restatement strategy)→ Then it explains the three-person setup of the “Imitation Game”: the interrogator, the human subject, and the machine act as judges and imitators respectively, testing the machine’s intelligent performance through text-based conversations (‘Imitation game’ Three-person game setup)→ Next, it points out the key transformation: shifting from “What is the essence of intelligence?” to “How is intelligence recognized through behavior”, that is, evaluating intelligence through observable behaviors rather than internal mechanisms (Key transformation)→ Finally, it discusses the variants and impacts of the Turing Test, from CAPTCHA to dialogue system evaluation. It has not only shaped AI research methods but also triggered ongoing debates about intelligence, consciousness, and machine ethics (Variants and impacts)
Original Question: “Can machines think?” → “Thinking” is difficult to define
Turing’s Proposal: Substitute with an operable alternative question
Roles: A (Male), B (Female), C (Interrogator, gender-neutral)
Interaction: C cannot see A/B and judges gender only through text-based questions
Goal: A induces C to make a wrong judgment, while B helps C judge correctly
This section returns to the formal birth of AI as an independent discipline. Firstly, marked by the 1956 Dartmouth Conference, it illustrates that AI evolved from scattered explorations into a research field with a common agenda (Dartmouth conference (1956, The formal birth of AI))→ Then it explains the proposal of the concept and naming of “Artificial Intelligence”, which signifies the expansion of research goals from “automatic computation” to the grand vision of enabling machines to demonstrate human-like intelligence (Concept and naming)→ Finally, it briefly reviews two major early achievements of AI, such as symbolic reasoning and problem-solving programs, early game-playing and theorem-proving systems, laying the paradigmatic foundation for symbolic AI in the subsequent decades (Two major early achievements)
Organizers: Minsky, McCarthy; Initiators include Shannon, Rochester
Format: Intensive “brainstorming” lasting 6-8 weeks
Participants: Solomonoff, Selfridge, More, Samuel, Newell, Simon, and about ten others
Proposal Declaration: “All aspects of learning and intelligence can be precisely described and simulated by machines”
McCarthy established the term Artificial Intelligence (deliberately distinguished from “cybernetics”)
1956 “Logic Theorist” (Newell & Simon)
Proved 38 out of the first 52 theorems in Principia Mathematica
Provided more concise proofs for some theorems
1959 Samuel’s Checkers Program: Reached the level of an amateur master
This section sorts out the periodic ups and downs in AI development. Firstly, it points out that early research made unfulfillable promises due to over-optimism, leading to a serious disconnect between expectations and reality (Over-optimism and failed promises)→ Then it analyzes the reasons for failure, including structural issues such as insufficient computing resources, limited algorithmic capabilities, and misjudgment of the essence of intelligence (Reasons for failure)→ These setbacks ultimately led to funding withdrawal and confidence collapse, contributing to the first “AI winter” in 1974 (Funding withdrawal and the ‘First AI winter’ (1974))→ However, with the rise of expert systems in the 1980s, AI ushered in a short-lived “spring”, driving practical commercial applications through knowledge engineering and rule-based systems (Expert systems bring a ‘Spring’ (1980s))→ Yet as maintenance costs soared and systems became rigid, AI entered the second winter between 1990 and 2000 (The second AI winter (1990–2000))→ Finally, in the 21st century, relying on improved computing power, accumulation of big data, and breakthroughs in machine learning, AI moved towards revival and diversification, entering a new stage of modern industry and research (Revival and diversification after the 21st century)
1958 Simon & Newell: Computers will become world chess champions within a decade; discover and prove important theorems within a decade
1965 Simon: Machines will be able to do everything humans can do within two decades
1967 Minsky: “AI problems will be largely solved within a generation”
1970 Minsky: Build a machine with “average human intelligence” within three to eight years
Application difficulty was underestimated: Solvable in small examples, but difficult to handle at real scale
Insufficient computing power: Severely limited by the hardware of the time
Data scarcity: Fields like computer vision/NLP require massive data, which was difficult to construct/learn at that time
Governments (UK), Defense Advanced Research Projects Agency (DARPA), National Research Council (NRC), etc., withdrew funding for “undirected AI research”
Funding plummeted, making projects unsustainable
Architecture: Knowledge base + inference engine
Implementation: Successful enterprise-level applications such as medical diagnosis
The boom lasted about a decade
Maintainability/portability issues of expert systems emerged
Investment and enthusiasm cooled down again
Causes: Improved computing power + focus on “verifiable” specific problems + rigorous scientific paradigms
Result: AI diversified into multiple competitive subfields (each deeply exploring specific problems/methods)
Industry: Flourished comprehensively in large-scale engineering and products
This section distinguishes between different types of AI. Firstly, it introduces Artificial General Intelligence (AGI / General AI), emphasizing that its goal is to enable machines to possess cross-domain, human-level understanding and reasoning capabilities (Artificial general intelligence)→ Then it explains that Narrow AI only performs excellently in specific tasks, such as image recognition or recommendation systems, but cannot be transferred to other scenarios (Narrow AI)→ Finally, it summarizes the significance of the shift in research focus from “pursuing general intelligence” to “developing usable narrow-domain intelligence”: it has brought AI from a philosophical vision to practical applications and driven the rapid development of modern AI (Significance of the shift in research focus)
Goal: Can understand/learn any intellectual task that humans can accomplish
Current Status: No successful instances yet
Only handles single or limited tasks (targeting specific scenarios and metrics)
Representatives: AlphaGo, IBM Watson (integrates multiple narrow-domain technologies)
The shift from “pursuing general AI” to “deeply cultivating narrow AI” was the key to emerging from the first AI winter
1990s – early 2000s: The rise of subfields oriented towards verifiable results and commercial value—bringing greater academic and commercial success
This section deepens the AI definition system. Firstly, it introduces the academic classification of AI concepts, refining “intelligence” into several research directions from different dimensions (Deepening and classification system of AI definitions)→ Then it elaborates on four classic definition paths respectively: targeting “Thinking humanly”, attempting to simulate human cognitive structures; taking “Thinking rationally” as the criterion, emphasizing logical reasoning and formalized knowledge; measuring intelligent behavior by “Acting humanly”; and focusing on “Acting rationally”, pursuing the optimal behavior choice in the environment→ Finally, it puts forward the core concept of rational agent, defining AI as a system that can independently perceive, reason and take optimal actions in the environment, laying a unified framework for subsequent AI systems and modeling methods (Core concept of rational agent)
In recent years, the understanding of the definition of AI has become more in-depth.
The two-dimensional framework proposed by Stuart Russell and Peter Norvig:
Dimension 1: Thinking vs. Acting
Dimension 2: Humanly vs. Rationally
Four types of artificial intelligence formed therefrom:
| Human-Like | Rational | |
| Thinking | Thinking Humanly | Thinking Rationally |
| Acting | Acting Humanly | Acting Rationally |
Representative definition: AI aims to automate activities related to human thinking, such as decision-making, problem-solving, and learning (Bellman, 1978)
Closely related to cognitive science and neuroscience
The goal is to replicate the human thinking process
Distinction: Different from using computational models to study mental abilities (which falls into the realm of rational thinking)
Goal: Establish strict reasoning rules to achieve logical thinking (Charniak & McDermott, 1985)
Highly relevant to the field of computational logic
Limitations:
Not all intelligent behaviors can be generated through logical reasoning
Therefore, thinking rationally cannot fully cover the field of artificial intelligence
Definition: Let computers accomplish tasks that humans are currently better at (Rich & Knight, 1991)
Closely related to the Turing Test: Demonstrate behavioral performance as close to humans as possible
Characteristic: Focuses on the behavior itself rather than the mechanism of behavior selection
Limiting the imitation of human behavior + achieving the ultimate goal of artificial intelligence
Definition (Luger & Stubblefield, 1993):
Artificial intelligence is a branch of computer science
Devoted to automating “intelligent behavior”
Intelligence is measured by formal indicators (quantifiable criteria)
Differences:
Acting Rationally vs. Acting Humanly
Humans may make mistakes, while rational programs can avoid errors
Examples:
Tic-tac-toe, checkers, Rubik’s Cube solving programs
All can achieve error-free “rational behavior”
The most common definition of modern AI is based on rational agents.
Definition of a rational agent:
A function that maps perceptions to actions
Goal: Maximize the expected utility of the objective function
Scope of application: Specific tasks (such as playing chess, image recognition)
Perfect rationality: Always take optimal actions
Bounded rationality: Limited by computational resources, make the best decisions within feasible ranges
Goal of artificial intelligence: Design agents that achieve rational behavior under limited resources
This section focuses on the capability levels of agents. Firstly, it introduces five classifications of agent performance levels, ranging from simple reflexive behaviors to advanced agents with learning and reasoning abilities, forming a complete spectrum of AI behavioral capabilities (Five classifications of agent performance levels)→ Then it explains that the structured degree of different problems affects agent design: some tasks have clear rules and definable states, while others involve open and complex environments that require stronger adaptability (Problem characteristics and structured degree)→ Next, it points out which tasks are most challenging at high-human and par-human levels, such as natural language processing, visual understanding, and complex decision-making scenarios (Tasks at high-human and par-human levels)→ Finally, it discusses development limitations and future outlook: data dependence, computing power bottlenecks, and insufficient generalization ability remain current obstacles, but the long-term direction still points to a more autonomous and generalized agent system (Development limitations and future outlook)
Based on the core idea of rational agents (i.e., maximizing or minimizing indicators), they are classified into five categories by comparison with human performance:
Optimal: Performance reaches the theoretical limit and cannot be improved
Super-human: Superior to all humans
High-human: Superior to most humans
Par-human: Comparable to the average human level
Sub-human: Inferior to most humans
Problems that can reach optimal or super-human levels are mostly well-structured and mathematically describable
Typical examples of optimal agents:
Tic-tac-toe, Connect Four, Checkers, Rubik’s Cube
Heads-up limited Texas Hold’em (long-term optimal returns)
Representative examples of super-human agents:
Othello
Scrabble
Backgammon
Jeopardy!
Go
Heads-up/multiplayer no-limit Texas Hold’em (Chess)
Examples of high-human agents:
Crossword puzzles
Dota 2
Bridge
StarCraft
Tasks for par-human or sub-human agents:
Par-human: Optical Character Recognition (OCR), image classification, handwritten recognition
Sub-human: Object recognition, facial recognition, bipedal walking, speech recognition
This section presents the key leaps of AI technology in a chronological context. Firstly, it reviews early milestones: from ELIZA to expert systems, AI began to perform pattern matching and rule-based professional reasoning (Early milestones: From ELIZA to expert systems)→ Then it moves to major breakthroughs in the 1990s, such as support vector machines, probabilistic graphical models, and reinforcement learning, laying the mathematical foundation for modern machine learning (Major breakthroughs in the 1990s)→ Subsequently, it demonstrates the rise after 2000: big data, parallel computing power, and deep learning drove revolutionary progress in image recognition, speech recognition, and natural language processing (After 2000: The rise and explosion of modern AI)→ Finally, since 2017, AI has begun to conquer the field of “imperfect information”—systems like AlphaZero and AlphaStar have demonstrated superhuman performance in games, strategies, and high-complexity environments, marking AI’s entry into a new capability landscape (After 2017: AI conquers imperfect information domains)
ELIZA (1966) :
XCON (1980) :
IBM Deep Blue (1997)：
Defeated the world chess champion
Though relying heavily on brute-force computing, it marked a historic breakthrough for AI
2005: Succeeded in autonomous driving at the DARPA Grand Challenge.
2011: IBM Watson Defeated Human Champions in Jeopardy!
Could understand natural language problems
Did not rely on logical or algebraic reasoning
Combined classic NLP techniques with statistical methods
2012: AlexNet Sparked the Deep Learning Revolution
2016: Deep learning defeated top human Go players
2017 Libratus: Defeated the world’s top 4 poker players and won $1.8 million
2018 Waymo： Launched self-driving taxis in Phoenix
2019 AlphaStar：Defeated 99% of human players in StarCraft II
After 2020: GPT-3：
Generates texts like stories and poems
The language is fluent and natural, on par with human-level
This section examines AI development from an ecological perspective. Firstly, it points out the explosive growth of the AI research community, with the number of scholars, engineers, and interdisciplinary researchers rising rapidly, forming an unprecedented density of innovation (Explosive growth of the research community)→ Then it uses quantitative indicators such as the number of papers, conference influence, and patents to show the continuous rise in research output, indicating that AI has become a core sector in the global scientific research landscape (Quantitative indicators of research output)→ This is followed by the continuous strengthening of cooperation between academia and industry, as large technology companies and universities/research institutions promote the transformation of technology from papers to products through joint laboratories and cooperative projects (Strengthened academic-industrial cooperation)→ Finally, it emphasizes the importance of AI education and talent cultivation: university curriculum systems, professional degrees, and online education jointly build an AI talent training pipeline, providing a human resource foundation for future technological development and governance (AI education and talent cultivation)
In 1956: Only 10 researchers in the AI field
Today: The research community has expanded exponentially
Number of AI scientific papers (2000-2019)
The number of papers reflects:
The number of researchers
Invested funds and research popularity
This section emphasizes the core position of enterprises in the modern AI ecosystem. Firstly, it points out that AI research has shifted from being academia-led to a pattern where “enterprises are on an equal footing or even leading”, and commercial demands have driven enterprises to increase R&D investment (Rise of AI research in enterprises)→ Then it explains the key role of industry in AI research: mastering large-scale data, computing power resources, and real application scenarios makes it the main engine for promoting technology landing and iteration (Core role of industry in AI research)→ Subsequently, it lists major global AI research enterprises, such as Google, OpenAI, Microsoft, Meta, Amazon, etc., which lead the frontier through foundation models and platform technologies (Major global AI research enterprises)→ At the same time, it also emphasizes the importance of other types of enterprises and interdisciplinary forces, such as companies in medical care, finance, autonomous driving and other fields jointly participating in AI innovation (Other enterprises and interdisciplinary research)→ Finally, it summarizes: the in-depth integration of industry and academia is the key to modern AI development. Only by realizing the cycle of research-application-feedback can AI maintain continuous progress and meet social needs (Importance of integration between industry and research)
After 2010: Major companies established AI/ML labs
Deep cooperation between enterprises and academic researchers
Researchers’ involvement:
Academic conferences
Paper publications
Product innovation and application
AI research has become the core driver of corporate technological innovation
Many AI milestones have been achieved by private enterprise labs
Tech giants: Google, Microsoft, Amazon, Apple, IBM, Meta (formerly Facebook)
Each lab has an average of over 1,000 researchers
Provide awards and funding support to academia
Co-founded the “Partnership on AI”
Focus on the social responsibility and ethics of AI
Baidu joined in 2018, becoming the first Chinese member
By 2019, the alliance had more than 100 members (covering academia, enterprises, and non-profit organizations)
This section outlines the overall framework for building AI systems. Firstly, it clarifies its goals and scope: from classification methods to various characteristics, to comprehensively understand the composition and operation of AI systems (Goals and scope)→ Then it proposes a top-level tripartite classification, dividing AI systems into three categories: technical methods, technical tools, and regulatory and organizational issues (Top-level tripartite classification: Technical methods / Technical tools / Regulatory and organizational issues)→ Technical methods are further subdivided into three major subtypes: Perception technology, Action technology, and Cognitive technology (Three major subtypes of technical methods)→ In the dimension of technical tools, it is divided into modeling tools and basic support platforms (Two major types of technical tools)→ Finally, it points out that the regulatory and organizational issues need to focus on ethical compliance, safety risks, human-machine collaboration, and social impacts, laying a structured foundation for subsequent discussions on AI governance (Overview of regulatory and organizational issues)
Overview of AI-related subfields
Explain classification methods, respective characteristics, and their interrelationships
Other courses: Will develop into each subfield separately
Technical methods (theories, models, algorithms)
Diverse and require further subdivision
Classified by imitated human capabilities: Perception / Cognition / Action
Technical tools (software/hardware)
Support the development and implementation of AI products
Two categories: Methodology and AI platforms
Regulatory and organizational issues
Perception technology
Task: Acquire, process, and understand information from the environment
Representatives: Computer Vision (CV), Natural Language Processing (NLP)
Action technology
Task: Execute actions in the physical world
Representative: Robotics
Cognitive technology
Task: Learning, Knowledge Representation, Reasoning, Planning, Collaboration
Representatives: Machine Learning (ML), Knowledge Representation and Reasoning (KRR), Planning and Scheduling, Multi-Agent Systems
This section introduces the core role of “perception and action” in AI. Firstly, it points out that the two together enable agents to interact with the physical world: perception is responsible for acquiring, processing, and understanding information; action is responsible for performing tasks based on that understanding (Overall role of perception and action)→ Then it moves to the visual perception module, explaining that the goal of Computer Vision (CV) is to gain high-level understanding from images and videos. By converting high-dimensional visual signals into computable representations, it promotes capabilities such as recognition, tracking, and generation based on geometric, physical, and statistical modeling (Perception: Computer vision (CV))→ Finally, it introduces Natural Language Processing (NLP) as another key perception method. Its goal is to process and analyze large-scale natural language data to support tasks such as information extraction, understanding, question answering, translation, and generation, connecting language, knowledge, and intelligent behaviors (Perception: natural language processing (NLP))
Together enable interaction with the physical world
Perception: Acquire—Process—Understand information
Action: Execute physical tasks
Definition and Goals
Gain high-level understanding from images/videos
Convert high-dimensional visual data into computable numerical/symbolic representations
Model-based understanding grounded in geometry, physics, statistics, and learning theory
Subfields
Biometrics; Face/Gesture/Posture recognition
Image/Video generation and retrieval
Video understanding and behavior analysis; motion and tracking
Robotics and autonomous driving vision
Application examples
Automated MRI cancer detection
Autonomous driving vision systems
GAN image conversion
Automated social distancing monitoring
Definition and goals
An interdisciplinary field of linguistics, computer science, and AI
Process/analyze large-scale natural language data to extract information and insights
Support understanding, classification, organization, and generation
Subfields
Dialogue systems/Conversational AI; Text generation; Information extraction
Machine translation; Question answering systems; Speech processing and synthesis
Text classification and sentiment analysis
Application examples
Smart assistants, Machine translation
Document retrieval, Email filtering
This section introduces the “action” technology in the AI system, focusing on robotics. Firstly, it emphasizes its interdisciplinary nature: integrating and developing in collaboration with fields such as computer science, mechanical engineering, electrical engineering, control engineering, mechatronics, biology, and software engineering (Action: Robotics)
Interdisciplinary integration: Computer science +Mechanical/Electrical/Control/Mechatronics/Biological/Software Engineering, etc.
Goals and application scenarios
Assist or replace humans in completing tasks
Hazardous environments (radioactive, explosive handling), manufacturing processes, extreme environments (space/underwater/high temperature), hazardous/radiation environment cleanup
Form and bionics
Diverse appearances; some are humanoid to improve acceptance
Replicate walking, weightlifting, speech, cognition, etc
Development of nature-inspired biorobotics
Subfields
Human-robot interaction (HRI)
Localization, mapping, and navigation (SLAM, etc.)
Manipulation control
Motion and path planning
Application examples
This section systematically sorts out the composition of “cognitive AI technologies”. Firstly, it clarifies its task scope and points out that it is the core module supporting the high-level intelligence of agents (Scope of cognition and AI tasks)→ Then it introduces Machine Learning (ML), whose goal is to enable algorithms to automatically improve performance through data and experience, covering methods such as supervised/unsupervised/reinforcement learning, deep learning, and evolutionary algorithms, and is widely used in prediction, recommendation, maintenance, and anomaly detection (Machine learning (ML))→ Next, it discusses Knowledge Representation and Reasoning (KRR), emphasizing enabling machines to understand and automatically reason about world knowledge in the form of rules, logic, conceptual structures, etc., supporting applications such as expert systems, semantic networks, and automated theorem proving (Knowledge representation and reasoning (KRR))→ Afterwards, it introduces automated planning and scheduling, whose core is to formulate strategies and action sequences for agents, conducting goal optimization and task execution in different environments (modelable vs. dynamic unknown) (Automated planning and scheduling)→ Finally, it presents Multi-Agent Systems (MAS), composed of multiple interacting agents. Through methods such as collaboration, game theory, distributed problem-solving, and mechanism design, it solves complex problems that cannot be handled by a single agent, and is used in economic transaction automation, resource scheduling, and distributed control (Multi-agent systems (MAS))
Covers: Knowledge formation, memory/working memory, judgment/evaluation, reasoning, problem-solving, decision-making
Core subfields: Machine learning, KRR, Planning and scheduling, Multi-agent systems
Definition: Algorithms that automatically improve through experience and data; build models based on training data for prediction/decision-making
Data sources: From perception systems like CV/NLP
Relationship with statistics: Closely related to computational statistics but not identical; data mining emphasizes unsupervised exploratory analysis
Commercial terms: Predictive analytics/Advanced analytics
Subfields/Methods:
Unsupervised/Supervised/Reinforcement learning
Classification/Regression; Clustering
Deep neural networks; Evolutionary learning
Online learning and bandit algorithms; Quantum machine learning
Applications:
Prediction and forecasting (demand/sales volume)
Recommendation systems
Predictive maintenance
Anomaly detection
Goal: Represent world knowledge in a machine-usable form and enable automatic reasoning
Method sources: Draws from human problem-solving research (psychology) and logic
Types of reasoning: Rule application, set/subset relationships, etc
Subfields:
Logic describing; Knowledge engineering; Logic programming
Ontology; Automated reasoning and Theorem proving
Applications: Semantic Web, expert Systems, etc
Goal: Develop strategies or action sequences for agents/robots/unmanned systems
Relationship with decision theory: Optimization under goals and uncertainty
Environmental differences:
Known and modelable: Offline planning and pre-evaluation are feasible
Dynamically unknown: Online adjustment is needed, often combined with iterative trial-and-error and learning
Subfields:
Deterministic planning; Spatio-temporal system optimization
Plan execution and monitoring
Integration of planning/scheduling and learning
Temporal planning
Definition: A system composed of multiple interacting agents
Motivation: Solve problems that are difficult for a single agent to handle
Sources of intelligence: Methodology, functionality, procedural methods, search and reinforcement learning, etc
Subfields:
Adversarial agents
Protocol, argumentation and negotiation
Coordination and collaboration
Distributed problem solving
Mechanism design
Applications:
Automation of economic transactions (microeconomic models)
Physical security resource allocation
Decentralized control
This section outlines the supporting structure of the AI technology system. Firstly, it distinguishes between “technical methods” and “technical tools”: the former focuses on the functions of AI products themselves (such as algorithm and model design), while the latter focuses on the development and deployment processes (engineering and platformization) (Technical methods vs. technical tools)→ Then it emphasizes the engineering methodology of AI projects, pointing out that their high-risk nature requires strict processes, including concept design, blueprint evaluation, PoC verification, and MVP delivery, and draws on the systematic management of software engineering (Methodology (Engineering process and project management))→ Subsequently, it introduces the importance of software and hardware infrastructure: big data and algorithms form the software foundation, while high-computing-power hardware (especially accelerators such as GPUs) significantly determines the efficiency of deep learning training and inference (Infrastructure (Software and hardware))→ It further explains that AI platforms, as key tools integrating software and hardware, provide enterprises with pre-built, drag-and-drop, or low-code tools (such as platforms from Microsoft, Amazon, and Google) to lower development thresholds and continuously expand algorithm capabilities (Artificial intelligence platforms)→ Finally, it discusses AI democratization, whose goal is to enable small and medium-sized enterprises and individuals to carry out AI projects without building their own teams. By opening up modules, data, and cloud infrastructure, it reduces costs, promotes collaboration, and accelerates the popularization of AI, playing a key role in ecological development and industrial implementation (AI democratization)
Technical methods: Focus on the functional realization of AI products
Technical tools: Focus on development and deployment processes (engineering and platforms)
AI projects are high-risk and require strict processes
Typical stages: Conception — Blueprint and Evaluation — Proof of Concept (PoC) — Minimum Viable Product (MVP)
Draw on and customize software engineering practices
Software infrastructure supported by big data and algorithms
Hardware: High computing power demand (especially for deep learning); accelerators like GPUs significantly reduce training/inference costs
Background: Talent shortage leads major companies to launch pre-built/drag-and-drop/no-code AI tools
Representatives: Microsoft, Amazon, Google platforms
Coverage: Mature components for common applications like CV, NLP
Major companies continuously expand the algorithmic capabilities of their platforms
Concept: Enable small and medium-sized enterprises to carry out AI projects without building their own teams
Means: Provide directly usable modules, open data and algorithms, cloud infrastructure
Effects:
Reduce the entry barrier and overall costs for individuals and organizations
Promote community collaboration and talent cultivation in the ecosystem
Accelerate the popularization in academia and business
Timeline-based coverage of resources: Data, storage/computing, algorithms, deployment, market ecology
Importance: Plays a decisive role in the overall success of AI
This section discusses the governance issues of AI at the social and organizational levels. Firstly, it points out the public value driven by technology: big data, cloud computing, and complex algorithms provide broad application potential for governments and public sectors in healthcare, education, transportation, energy management, and other fields (Technology-driven and public value)→ Then it emphasizes risks and responsibilities, proposes the concept of “Responsible Innovation”, and reminds that large-scale data and automated decision-making may amplify discrimination. Therefore, public sectors should proactively prevent risks and build safe and fair AI systems (Risk and responsibility: Promote ‘Responsible innovation’)→ Subsequently, it defines the scope and subjects of AI ethics, covering the moral behaviors of humans in designing/using AI, as well as the ethical issues of the machines themselves; everyone from engineers to managers needs to assume ethical responsibilities, extending to topics such as superintelligent AI (Scope and subjects of AI ethics)→ Furthermore, it analyzes key legal issues, such as privacy protection, division of algorithmic responsibilities, and establishment of regulatory frameworks, which are core legal challenges for global AI governance (Key issues in legal matters)→ Finally, it discusses the impacts on organizations and the labor market: AI spawns new jobs, promotes the evolution of organizational structures, changes management methods, and brings a trend of “coexistence of job replacement and job transformation” in employment, becoming an important issue for future social transformation (Organizational and labor market impacts)
Three major drivers: Big Data, cloud computing platforms, complex ML algorithms
Potential government benefits: Healthcare, education, and transportation in smart cities
Other public applications: Food supply chain, energy and environmental management
Large-scale data and automated decision-making may exacerbate discrimination
Public sectors should foster a culture of responsible innovation and proactively prevent potential harms
Build ethical, fair, and secure AI systems
AI ethics: A subdomain of technical ethics
Two dimensions:
Moral behavior of humans in designing/manufacturing/using/disposing of AI systems
Ethics of machines themselves (machine ethics)
Issue extension: Discussions on superintelligence and the “singularity”
Actors: Data scientists, engineers, domain experts, delivery managers, department heads, etc., all need to prioritize ethics and safety
Privacy protection: Continuous interaction between devices and monitoring of human activities require strong regulation
Liability attribution: Responsibility positioning for algorithmic non-compliant decisions (designers/developers/deployers)
Other legal frameworks that governments around the world are researching
New occupations and structures: New roles such as AI engineers and ML engineers emerge
Experimentation and evolution of organizational structures
Management transformation: Help employers adapt to AI paradigms
Employment impact:
Monitoring and evaluation of machine substitution (not simply “replacement”)
Europe’s IT talent shortage (millions), and AI popularization is expected to alleviate it
This section emphasizes the necessity of formulating AI strategies at the national level. Firstly, it points out AI’s comprehensive impact on a country’s future, covering areas such as economic growth, national defense, public security, and privacy protection. AI can generate trillion-dollar-level global value annually, and national competitiveness depends on R&D investment, human capital, and industrial structure (Comprehensive impact of AI on a country’s future)→ Then it analyzes the global national stratification: the United States and China are in the first tier, the United Kingdom, Germany, France, etc., in the second tier, and countries like Italy in the third tier, reflecting differences in national investment, talent, and technological ecosystems (National stratification and Italy’s current status)→ Subsequently, it elaborates on the necessity and goals of formulating national AI strategies: to improve national AI readiness, promote AI adoption in enterprises and public sectors, coordinate development and risks, and protect citizens from technological side effects (such as discrimination and privacy risks caused by lack of regulation) (Necessity and goals of formulating national AI strategies)→ It further introduces the world’s earliest national strategies (e.g., Canada’s 2017 Pan-Canadian AI Strategy) and explains that countries have since formed differentiated priorities, covering layouts in scientific research, talent, ethics, public governance, industrial collaboration, and other aspects (Pioneers and diversified focuses)→ Finally, it summarizes the differences in AI development paths among the three key strategic blocs: the United States, China, and the European Union (Three key strategic bodies (US/China/EU) and their different development perspectives)
Impact domains: Economic development, national defense construction, public safety and privacy, etc
Economic volume (McKinsey 2018): AI unlocks more than $1.5 trillion in global market value annually
Determinants of value differences: National preparedness (investment, research, human capital, labor market structure)
First-tier: United States, China
Second-tier: United Kingdom, France, Germany, Australia, Canada, Sweden, Norway, South Korea, etc
Third-tier: Italy, Costa Rica, India, Lithuania, Spain, etc
Necessity: AI development is crucial to nations, so national strategies must be formulated
Overall goals:
Enhance national preparedness
Promote the application and development of AI in enterprises/public administration/education
Dual tasks:
Coordinate economic development and benefit citizens
Protect citizens from side effects (unintended consequences), leading to regulatory demands in ethics, law, etc
First national strategy: Canada’s “Pan-Canadian AI Strategy” (2017)
Focus: Investment in R&D/technology and talent
Main goals: Increase researchers; identify centers of excellence; support leadership in economic/ethical/legal issues; support academic communities
Subsequent national strategies: Form differentiated focuses, including research, talent/skills/education, public-private applications, ethics and inclusion, standards and regulations, data and digital infrastructure, etc
United States/China: Postpone regulatory issues (5-10 years) and adopt relatively loose regulation currently
United States: Development is mainly driven by large tech companies
China: Development is mainly led by the government
Europe: Prioritizes regulation before large-scale development
Follow-up arrangement: Three separate sessions will analyze the strategies of the United States, China, and the European Union
This section takes the U.S. national AI strategy as a case study. Firstly, it reviews the starting point and goals of the Obama administration: based on a full understanding of AI’s potential, impacts, and risks, formulate a future-oriented national strategic agenda (Starting point and goals (Obama administration period))→ Then it introduces the first report Preparing for the future of AI, which focuses on ethics and regulation, and puts forward suggestions such as opening up data, standardization, and establishing a special committee to monitor AI development in the long term (Report 1: Preparing for the future of AI)→ The second report National AI research and development strategic plan clarifies priority directions such as long-term scientific research investment, human-machine collaboration methods, and system safety and reliability from the perspective of public funding, becoming the cornerstone of the U.S. national AI strategy thereafter (Report 2: National AI research and development strategic plan)→ The third report AI, automation, and the economy evaluates the impact of automation on industries, governments, and the labor market, focusing on how to achieve a cost-benefit balance through structural adjustments (Report 3: AI, automation, and the economy)→ Later, a strategic shift occurred during the Trump administration: it emphasized deregulation, with large technology companies leading the technical route. At the same time, through the White House-tech giant summit, it established goals such as maintaining global leadership, supporting the workforce, encouraging R&D, and removing innovation barriers, reflecting a more market-driven development model (Strategic shift (Trump administration period))
Focus: Ethics and regulation
Key recommendations:
Data openness and standardization: Make open data and data standards federal priorities (data is the “fuel” for AI)
Establish a special committee: Continuously monitor AI development and report to the government regularly
Support basic and long-term research: Prioritize investment in AI basic/long-term research
Priority directions for public funding investment:
Long-term investment in AI research
Develop effective methods for human-machine collaboration
Ensure the safety and reliability of AI systems
2019 update: Become the foundation of the current U.S. national strategy
Content: Assess the impact of automation on industry, administration, business, education, etc
Focus: Labor market impact and restructuring to optimize the benefit-cost balance and promote AI development
2017: Relax regulation → large tech companies lead technology development
2018 White House – Tech Giants Summit established four major goals:
Maintain global leadership
Support AI’s assistance to human labor
Promote R&D with public funds
Eliminate innovation barriers
Supporting measures:
Establish an AI Priorities Advisory Committee
Establish federal-level industry-academia partnerships
Remove multiple regulatory restrictions
2021 Investment: The White House plans to invest over $8 billion in AI-related R&D (an increase of about 50% compared to 2020)
Priority areas: National defense and security
Applications in other areas: Mainly promoted by enterprises
This section introduces China’s national AI development strategy. Starting with the New Generation Artificial Intelligence Development Plan (2017), it clarifies the three-stage overall goals: by 2020, the industry technology reaches the international leading level; by 2025, it leads in key application fields; and by 2030, China becomes the global AI innovation center (Overall goals and timeline)→ Then it presents the economic scale expectations, showing that China’s AI market value will grow from approximately $23 billion to about $150 billion between 2020-2030, reflecting the country’s long-term strategic investment (Economic expectations)→ Next, it elaborates on the regulatory path and talent strategy: China plans to complete the AI-related specification system by 2030, while emphasizing the dual approach of “global recruitment of high-end talents + local talent cultivation,” and improving the industry’s overall level through internal corporate training and ethical supervision (Regulatory path and talent strategy)→ Finally, it explains the technology priority directions and infrastructure layout, focusing on intelligent IoT (autonomous driving, robotics, recognition systems), accelerating breakthroughs in key technologies such as intelligent sensors and neural network chips, and building large science and technology parks with massive public funding to provide systematic support for AI industry development (Technology priority directions and infrastructure)
Goal: Become a global AI leader
Three-phase milestones:
2020: Industrial technology level reaches the level of global competitors
2025: Become a world leader in some application fields
2030: Become a global AI innovation leader
2020: ~$23 billion
2025: ~$60 billion
2030: ~$150 billion
Regulation: Complete the compilation of norms and rules system by 2030
Talent: Emphasize global recruitment of top talents and local labor market development
Biggest challenge for enterprises: Lack of qualified talents (reported by >50% of enterprises)
Measures: In-company AI training; formulate regulations and ethical guidelines to promote development and application
Priority directions: Intelligent Internet of Things (autonomous driving, robots, recognition systems)
Key technologies: Intelligent sensors, neural network chips, intelligent manufacturing
Major facility: Establishment of an excellence science and technology park in Beijing (~$2.1 billion)
Public funding: ~$8 billion annual public expenditure on AI research and industry
This section takes “human-centered and trustworthy AI” as the core position. Firstly, it emphasizes Europe’s principle for AI development: AI should only be developed when it benefits humans, and risks should be strictly assessed based on a strong regulatory framework (Human-centered and “Trustworthy AI”)→ Then it sorts out the key policy milestones from 2018 to 2021, from the European Artificial Intelligence Strategy to the proposal of the AI Act, gradually translating the vision of “excellence + trustworthiness” into specific systems (Key policy milestones (2018–2021))→ Furthermore, it explains how regulation reversely promotes development: by enhancing user trust, providing legal certainty and unified rules, it expands market space rather than suppressing innovation (Regulation promotes development: Mechanisms and logic)→ On this basis, it introduces a risk-oriented hierarchical regulatory framework, setting prohibitions, mandatory compliance, transparency obligations and regular management requirements for unacceptable risk, high-risk, limited-risk and minimal-risk scenarios respectively (Risk classification and regulatory requirements)→ At the same time, it clarifies the roles of member states in the governance system: establishing national competent authorities and supervisory bodies responsible for implementation, market monitoring, and acting as interfaces with the public and the EU level (Roles of member states and governance framework)→ Finally, it emphasizes building an innovation ecosystem through funding programs, digital infrastructure and public-private partnerships, aiming to spread “excellent and trustworthy” AI applications under strict regulation and a unified market (Investment, infrastructure, and ecosystem)
Principle: AI should only be developed when it benefits humans
Approach: Strictly assess benefits and risks before approving specific applications
Reliance: Central regulatory framework (regulation first)
2018: European Artificial Intelligence Strategy; Member states signed the AI Coordination Plan
2019: Guidelines for Trustworthy AI
2020：
2021: Proposal for the AI Act: Unify AI rules and establish a strict regulatory framework
Position Summary (2021):
Increased user trust promotes demand from enterprises and public sectors
Legal certainty and unified rules allow suppliers to enter broader markets and make products more popular
Conclusion: Without avoiding risks, ensuring trustworthiness, and legal safeguards, AI cannot develop effectively
Unacceptable Risk (Prohibited)
A few applications that violate European values and infringe on fundamental rights
Examples: Government social scoring; exploiting children’s vulnerabilities; subliminal manipulation; real-time remote biometrics for law enforcement in public places (with very few exceptions)
High Risk (Mandatory Compliance)
Systems that adversely affect safety or fundamental rights
Mandatory requirements: Data quality; technical documentation and records; transparency and user information; human oversight; robustness and accuracy; cybersecurity
The list can be adjusted as use cases evolve
Limited Risk (Transparency Obligation)
Minimal Risk (Routine Management)
Other AI systems are developed and used in accordance with existing laws, with no additional obligations
Suppliers can voluntarily follow Trustworthy AI requirements and codes of conduct
National Competent Authorities: Supervise implementation and enforcement, conduct market monitoring
National Supervisory Authorities: Official contact points with the public/stakeholders; represent their countries in the European AI Committee
Funding Channels: Initiatives like “Digital Europe” and “Horizon Europe” (20% of expenditure on digital transformation)
Measures: Encourage sharing of information/data/computing infrastructure; public-private cooperation; promote digital innovation and talent mobility (from lab to market)
Goal: Promote the diffusion of excellent and trustworthy AI applications under strict regulation and a unified market
1.2.5.5 Social-level trade-off: Government power and privacy
After 9/11, the “USA Patriot Act” granted broader powers for electronic surveillance
Lesson: Safety should not take precedence over other values; each case requires a detailed analysis and weighing