Navigating the Ethical Landscape of AI Innovation: Challenges and Opportunities

4.1 Biased AI: The Non-Neutrality of Technology Determined by Social Attributes

The social attribute of technology refers to the fact that the generation and application of technology is never an isolated or neutral process but is shaped and developed in a specific social background, cultural context and power structure (Weber 2019). Technology is not only a tool to satisfy human needs, but also a product that carries and reflects social values, institutional arrangements and power distribution (Howard and Borenstein 2018). As a controversial social technology, AI is designed, trained and deployed in close interaction with social factors.

Currently, with the gradual evolution of AI from weak AI, strong AI to super AI, its algorithmic structure and processing power have been significantly enhanced, and the “autonomy” of the system is also increasing. This transformation has sparked legal and ethical discussions across various fields. For example, in the recruitment process, the application of AI may raise concerns regarding fairness, privacy protection, and other related issues (Dattner et al. 2019). However, despite continuous iterations in AI technology, its fundamental data foundation (e.g. gender ratio in historical employment data) and training goals (e.g. efficiency-first recruitment model) are still set by humans (Shin 2024). Therefore, AI’s “autonomous learning” is not truly autonomous, but rather a conscious or unconscious projection of human bias: when training data reflects inherent societal discrimination (e.g. male dominance in the computer industry), the algorithm not only fails to correct for these discriminations but instead automatically reinforces them, resulting in systemic discrimination and algorithmic bias. Machine learning models are essentially “probabilistic replicators” of historical data, and their core goal is to find patterns and maximize prediction accuracy, not to actively challenge the social power structure behind the data. In other words, when “efficiency first” becomes the underlying logic of the algorithm design, the model will consider the group characteristics in the existing data (e.g. the correlation between gender and occupation) as “objective laws” to be reinforced, thus maintaining or even exacerbating the existing data bias. This “technological black box” effect translates bias from implicit social structures into explicit decision-making outcomes (e.g. a preference for men in computer positions), creating a vicious cycle of “discriminate-reinforce-reproduce”.

The existing governance system’s response to “biased AI” reveals the cognitive limitations of the technocentric paradigm. The current mainstream anti-discrimination regulation approach still sees algorithmic bias as a fixable technological fault but ignores its nature as a digital mapping of social structures. Therefore, numerous governance challenges remain in achieving an effective response to algorithmic discrimination.

4.2 AI Algorithm Disclosure: Limits of Transparency

Transparency of AI systems usually refers to the degree to which the relevant subjects can understand the algorithm’s decision-making logic, data input and output mechanisms, and potential risks (O’Neill 2012). With the wide application of AI in key fields such as healthcare, architecture, and public safety, its “black box” characteristics have become more and more prominent, and people are often confused by the logic of algorithmic operation and decision-making basis, which triggers a strong demand for transparency, and pushes the academic community and the public to continuously call for “algorithmic disclosure”. The public and academics continue to call for “algorithm disclosure”. However, the disclosure of AI algorithms faces multiple contradictions and challenges.

On the one hand, the essence of AI transparency is the human subject’s cognitive breakout of the technological black box. Although the disclosure of algorithms is regarded as an important means to deal with the “black box” problem, in addition to the perceived deliberate concealment, the opacity of algorithms is fundamentally rooted in the complexity of the technology (Ding 2022). At the same time, the purpose of algorithmic disclosure is to allow consumers, developers and stakeholders to understand the internal mechanisms of AI models. However, for algorithms, especially deep learning models, with their complex structures and highly automated operations, even if a certain degree of disclosure can be achieved technically, due to the high threshold of professional knowledge, the huge amount of information, and the fact that the data operations rely on correlation rather than causation logic (Schölkopf 2022), consumers and general stakeholders are still unable to form a substantive understanding of the logic of the algorithm, and even more so, the original purpose of algorithm disclosure cannot be realized (Alsaigh et al. 2023). It even makes “technological democracy” a mere formality, reduced to a symbolic display of symbols. On the other hand, in practice, the demand for algorithmic disclosure is often alienated as a tool for responsibility avoidance. Take the COMPAS algorithm applied in the United States judicial system as an example, when the decision-making results of the system are questioned about the existence of bias, the designer uses the name of “technical objectivity” to transfer the ethical responsibility through the “black-box defense” (Engel et al. 2024), and attributes the responsibility to the model’s own non-interpretability. This phenomenon reflects the demand for transparency. This phenomenon reflects a technological authorization and instrumentalization of the demand for transparency, which essentially weakens the effectiveness of ethical accountability.

The deeper crisis lies in the fact that when the public blames algorithmic bias on “code defects”, the fundamental role of human subjectivity is overlooked. The value penetration in data labeling, the power game in feature engineering, and the design preference of the model objective function, the “human intervention” in these key links is covered by the black box effect (Burrell 2016). As a result of this cognitive inversion, technical governance often stays at the surface level of parameter disclosure and algorithm publicity, failing to touch on the substantive reconstruction of the power structure and value mechanism behind.

4.3 Data Utilization and Privacy Protection: Value Trade-offs Dilemmas

Data has been described as a “new type of production factor” in the AI era, and its efficient circulation and in-depth utilization constitute the core support for algorithm iteration and model optimization (Cheng and Gong 2024). In this context, data security has transcended the traditional scope of technical protection and has been upgraded to the most fundamental and universal normative requirements in the AI ethical system. The right to privacy is a product of the conflict between interests and balance, and its connotation is constantly transmuted with the exponential growth of computing power and the paradigm revolution of data utilization mode (Romanosky and Acquisti 2009). Nowadays, the technical rationality represented by data utilization and the ethical value centered on privacy protection show an increasingly significant antagonistic tension, which not only permeates the operational aspects of data collection, processing, and sharing, but also runs through the complete life cycle of AI systems from design and development to deployment and application (Hewage et al. 2024).

On the one hand, the precise evolution of AI systems relies on the continuous “feeding” of massive data: medical diagnostic models need to be supported by millions of medical records data, financial risk control system relies on real-time consumption behavior trajectory, and automatic driving algorithms need to devour massive road condition information (Regan 2003). This perpetual demand for “data fuel” makes technology giants constantly break through the traditional data collection boundaries, and realize the “ultra-granularity” capture of personal information through sensor networks, biometrics and other technologies. On the other hand, uncontrolled data collection and commercialization have given rise to the risk of privacy infringement, data abuse, and even digital “surveillance capitalism” (Saheb 2023), with the individual’s right to self-determination and human dignity facing systematic erosion. The essence of this dichotomy lies in the fact that privacy protection takes the protection of individual rights as its core demand, while the utilization of data points to the fact that the realization of public well-being relies on the efficient circulation of data and the transformation of its value. The two are in structural opposition to each other, with the values of rights-based and efficiency prioritization.

This paradox is further exacerbated by the ambiguity of data property rights. As a mainstream regulatory paradigm, “personal data empowerment” attempts to empower individuals with data rights from the perspective of the object on which algorithms are based, so as to strengthen individuals’ knowledge and control over their personal data (Ding 2022). In practice, however, this approach has revealed three challenges: first, when users are confronted with 10,000-word privacy agreements, the actual right to choose is virtually non-existent. In such cases, users either passively accept the terms and conditions in order to access the service or face exclusion from the service. When the technical architecture reduces “informed consent” to a binary option, the individual’s control over the flow of data deviates from the original purpose of empowerment (Chomanski and Lauwaert 2024). Second, the definition of personal data is controversial, especially in the context of the widespread use of anonymization and de-identification technologies, and the question of whether data belongs to the category of “identifiable individuals” has become a focal point (Song and Mittal 2021). For example, behavioral data such as browsing records and location information may re-identify individuals after a multi-dimensional intersection, leading to the blurring of the boundary between “non-personal data” and “personal data”, making it difficult to clarify the scope of empowerment, and providing grey space for enterprises to circumvent regulation (Podda 2024). Third, data infringement is often covert, technical and difficult to identify, which makes it difficult for users to obtain sufficient evidence even if they realize that their rights and interests have been damaged (Schneider 2023). Moreover, the evidentiary burden, the delayed responsiveness of existing remedial mechanisms, and the high costs associated with individual rights protection further hinder the effectiveness of personal data empowerment mechanisms. As a result, these mechanisms frequently fail to fulfill their intended protective function, leading to the dilemma of formalized rights without substantive enforcement.

5.1 International AI Legal Governance

5.2 Domestic AI Legal Governance

With the rapid development of AI technology, China’s legislative practice around AI governance has been deepening, gradually building up a multi-level and three-dimensional regulatory system covering algorithmic safety, ethical norms, data governance and industry applications. On the whole, it shows a governance pattern by strategic planning and domain-specific advancements, forming a governance path with Chinese characteristics in which institutional norms and technical standards are synergistically linked.

In 2017, the State Council issued the Development Plan on the New Generation of Artificial Intelligence,^[7] which puts forward the strategic goal of “three steps”, setting off a new boom in artificial intelligence, and explicitly mentions the need to “strengthen the research on legal, ethical and social issues related to artificial intelligence, and establish a legal, regulatory and ethical framework that guarantees the healthy development of artificial intelligence”. It also explicitly proposed to “strengthen research on legal, ethical and social issues related to AI, and establish a legal, regulatory and ethical framework to ensure the healthy development of AI”. Translation: Guiding Opinions on Strengthening Overall Governance of Internet Information Service Algorithms^[8] makes it clear that the goal of algorithm governance is to establish a comprehensive governance pattern with a sound governance mechanism, a perfect regulatory system, and a standardized algorithm ecosystem; and the Opinions on Strengthening the Governance of Ethics in Science and Technology^[9] further advocates that ethics should come first, and calls for the ethics of science and technology to be carried out throughout the entire process of scientific research and technological development. In addition, documents such as A Next Generation Artificial Intelligence Development Plan,^[10] Ethical Norms for New Generation Artificial Intelligence,^[11] and the Guiding Opinions on Accelerating Scenario Innovation to Promote High-Level Application of Artificial Intelligence for High-Quality Development of the Economy^[12] collectively clarify the direction for our country to grasp the opportunities for the development of AI and promote high-level application, and provide basic guidance and value guidance for the development of the technology through the reinforcement of ethical norms.

Sub-domain breakthroughs and precise response to key risks: In key governance areas, China has implemented specialized regulations to address risks. In 2021, Provisions on Administration of Algorithmic Recommendation in the Internet Information Service^[13] addressed issues such as big data-driven price discrimination, and established mechanisms for algorithm registration, human intervention, and user autonomy. In 2023, the Administrative Provisions on Deep Synthesis in Internet-based Information Services^[14] focused on deep synthesis services, enforcing content labeling and strengthening review mechanisms to ensure information authenticity and maintain public trust. The Interim Measures for the Management of Generative Artificial Intelligence Services^[15] issued in the same year clarified the specific measures and basic norms to promote the healthy development of generative artificial intelligence, proposed to adhere to the principle of equal emphasis on development and safety, and the combination of innovation and governance in accordance with the law, and to adopt an inclusive, prudent, classified and hierarchical regulatory approach, and to encourage technological innovation. Although it focuses on the sub-field of generative AI, its classification and grading governance model also provides an important first demonstration for the overall regulation of AI in China in the future.

Overall, China’s AI governance system is committed to realizing a governance balance between “safe and controllable” and “innovation-driven” (Cheng 2023), emphasizing the equal importance of development and safety, and the combination of innovation and governance in accordance with the law, as well as the ethical principles throughout the entire process of the development of AI technology, in an effort to In addition, it has incorporated ethical principles throughout the entire process of AI technology development, and strives to provide solid ethical and institutional safeguards for the healthy and sustainable development of AI technology by guiding the development of AI towards goodness while safeguarding technological progress (Ding 2025).

6.1 Systematic Choice of Ethical Governance Pathways

In the realm of AI governance, two major theoretical orientations dominate academic discourse: the oppositional theory and the system theory (Dubber et al. 2020). The oppositional theory perceives technological development as inherently at odds with ethical constraints, advocating for strict regulatory measures to mitigate potential risks (Mäntymäki et al. 2022). In contrast, the system theory underscores the symbiotic relationship between technology and social structures, emphasizing the need to achieve a dynamic balance between innovation and ethical values.

Building upon prior discussions on AI’s social attributes, this study particularly highlights the necessity of embracing a systemic governance approach in the ethical dimension. Technological advancement and ethical oversight are not isolated forces; rather, they are deeply embedded within specific social structures, cultural contexts, and political systems, continuously influencing and shaping one another. Consequently, AI governance should break away from the binary of regulator and regulated. Instead, it should follow a more nuanced path grounded in social systems analysis (Neuberg 2003), shifting the logic of governance from confrontation to collaboration.

China’s AI governance practices in recent years have exemplified the importance of “systemic synergy”, implementing a regulatory framework that spans the entire AI lifecycle. This holistic approach integrates all relevant stakeholders – including individuals, enterprises, and organizations involved in AI management, research and development, supply, and usage – into a unified governance framework (Wang and Ding 2023). This demonstrates the state’s commitment to comprehensive coordination and regulation across the AI ecosystem. The emphasis on multi-stakeholder involvement and end-to-end governance lays a strong institutional foundation for constructing a systematic ethical governance model that ensures technological development aligns with societal values and interests (Schmarzo and Borne 2020).

6.2 Double-Helix Governance Model

In order to respond to the current dilemma of AI governance, this study tries to put forward the Double-helix Model for AI Governance of the “Technology-driven Innovation Axis” and the “Ethical-legal Constraint Axis”. The aim is to maintain the momentum of technological innovation while building a governance system with social responsibility and institutional constraints. The model does not regard technology and ethics as opposing dimensions but rather emphasizes that the two are constantly shaping each other in dynamic synergy, so as to realize the simultaneous enhancement of governance effectiveness and technological progress through spiraling interaction.

1. The “Technology-driven Innovation Axis”

This axis focuses on promoting independent innovation, industrialization and application of AI technology, as well as the sustainable improvement of economic benefits, and its core driving force covers key elements such as algorithm breakthroughs, data sharing, arithmetic enhancement and talent cultivation. As a highly complex socio-technical system, the development of AI not only relies on the technical support of machine learning models and software components, but is also deeply embedded in specific social organizations and institutional environments, and its evolution has a wide range of impacts on different cultures, societies, and political contexts, which urgently requires multidisciplinary collaborative systematic analyses to comprehensively grasp the social changes triggered by the technological advancement of AI.

In addition, the power axis of technological development should also actively respond to key issues such as talent cultivation, algorithmic discrimination and the “technological black box”. Realizing a full understanding of the AI decision-making process and effective supervision, establishing model monitoring and accountability mechanisms instead of blind trust in AI capabilities have become the core propositions of AI governance. Advancing interpretability research, enhancing the understanding of model logic, and gradually improving fairness algorithms are expected to provide more solid technical support and practical pathways for ethical and legal governance.

1. The “Ethical-legal Constraint Axis”

The second axis aims to establish value guidance and behavioral boundaries for the development of AI technology through ethical norms, legal systems and regulatory mechanisms, with the core mission of preventing the abuse of the technology and safeguarding social justice and basic human rights (Tsamados et al. 2022). Especially, in the context of increasing digitization, this governance axis should always adhere to the digital humanism stance, emphasizing that the development of AI should always serve the dignity and well-being of human beings, and ensure that technological progress does not deviate from the core value of “people-centeredness” (Wallach and Marchant 2018).

In terms of governance paths, it is possible to promote the embedding of ethical principles into the entire life cycle of AI systems, and to realize preventive regulation of high-risk technologies through unified ethical assessment standards and risk review mechanisms (Schmarzo and Borne 2020). At the same time, it is necessary to strengthen the embedded supervision of human beings in the key decision-making process, ensure that they have the substantive right to supervise and intervene in the operation of the AI systems, and effectively enhance the transparency and accountability of algorithms through the establishment of an independent review organization and a professional evaluation team. In addition, to address the technical and legal challenges of data rights and circulation, flexible mechanisms such as “scenario-based identification” (Ding 2022) and “data holding rights” (Gao 2023) can be explored to promote the rational use of data resources on the basis of safeguarding individual rights. Through prudent and flexible institutional design, the “Ethical-legal Constraint Axis” is expected to provide flexible and powerful institutional support for the innovative development of AI while protecting basic human rights and social values.

6.3 Dynamic Interaction: Realizing the Security and Development Spiral

Given the evolving nature of AI, governance strategies should be adjusted prudently in response to different periods and contexts, ensuring an optimal allocation of regulatory resources while striving for a balanced governance ecosystem. Rather than adhering rigidly to static regulatory norms, AI governance should remain adaptable and responsive, continuously refining its approach in alignment with technological progress and societal feedback. By fostering a governance framework that emphasizes “Technology for the Greater Good” (Cheng et al. 2023), AI can be guided toward steady and responsible development. This approach aspires to ensure that AI serves as a tool for advancing social well-being while supporting the creation of a more intelligent, equitable, and sustainable governance system.

On this basis, the AI governance model is supposed to facilitate the simultaneous enhancement of security and development through dynamic interaction, as illustrated in Figure 1.

Figure 1:

Dynamic interactive double-helix model for AI governance.

In the face of ethical challenges and opportunities in the process of AI innovation, the proposed Dynamic Interactive Double-helix Model for AI Governance provides a feasible path for balancing security and development by dynamically interacting the “Technology-driven Innovation Axis” and the “Ethical-legal Constraint Axis”. The former promotes AI innovation, industrial upgrading and economic growth, while the latter pursues fairness, transparency, responsibility and social well-being. The synergy between the two helices fosters the establishment of a governance system that is both inclusive and robust. This model enables AI to develop fully in a safe manner, while simultaneously guiding society towards a smarter, fairer, and more sustainable future. Moreover, it is also in line with the evaluation model of establishing a green, harmonious and sustainable network content ecosystem (Cheng et al. 2023).