Location : Home > Resource > Paper > Theoretical Deduction
Zhang Xin | Algorithmic governance Challenges and Governance Supervision in Generative Artificial Intelligence
2023-08-14 [author] Zhang Xin preview:

[author]Zhang Xin


Algorithmic governance Challenges and Governance Supervision in Generative Artificial Intelligence

*Author Zhang Xin

Associate Professor, School of Law, University of International Business and Economics

Abstract: As exemplified by ChatGPT, large-scale pre-trained language models are progressively showcasing their potential. The technical features of these models, including their vast scale, substantial parameter count, extraordinary scalability, and varied application scenarios, present comprehensive challenges to the algorithmic governance system centered around transparency, fairness, and accountability. In the prevailing global AI governance paradigms, the EU has instituted a risk-based governance framework, China has devised a subject-based governance approach, and the U. S. has embraced an application-based governance method. These three governance paradigms emerged during the “1. 0 era” of narrow AI. Consequently, as AI technology evolves, it is crucial to foster a holistic transformation of the regulatory structure, characterized by open collaboration in regulatory authority, the incorporation of diverse regulatory approaches, and the harmonization of regulatory measures. This shift will enable a transition towards “governance-oriented regulation” fitting for the AI “2. 0 era”.

Key words: ChatGPT; generative artificial intelligence; algorithmic governance; governance-based regulation


On November 30, 2022, Open AI launched the conversational artificial intelligence Chat GPT, which demonstrated stunning language understanding, language generation, and knowledge reasoning capabilities. In just two months, it had 100 million active users, making it the fastest-growing "phenomenon level" application in technology history. Chat GPT is an outstanding representative of Generative AI in the field of natural language processing, achieving the transition of artificial intelligence from perceiving and understanding the world to generating and creating the world. It represents a paradigm shift in the development and implementation of artificial intelligence technology. Compared to other forms of artificial intelligence, the disruptive impact of generative artificial intelligence is not limited to the technical and tool levels, but also has a significant impact in the field of governance. Generative artificial intelligence, represented by Chat GPT, indicates that the "spark" of universal artificial intelligence is about to start a prairie fire. Its powerful expansion and migration capabilities make it a truly new type of infrastructure, which will deeply penetrate various fields such as society, economy, politics, law, and subtly reshape social structure and governance forms. When generative artificial intelligence brings a series of technological dividends, the accompanying legal and ethical risks also make it deeply controversial. How to understand the technical characteristics and algorithmic governance challenges of generative artificial intelligence from a legal perspective, and explore regulatory and governance frameworks that are suitable for it, is an urgent public policy challenge that needs to be solved. This article focuses on generative artificial intelligence represented by Chat GPT, exploring its technical characteristics and governance implications. Starting from the global competition pattern of artificial intelligence and the governance goals of responsible innovation, it makes a forward-looking exploration of the future regulatory framework.

1 The Technical Characteristics and Algorithm Governance Challenges of Generative Artificial Intelligence

In 2017, the State Council issued the "New Generation Artificial Intelligence Development Plan", which includes knowledge computing and services, cross media analysis and reasoning, and natural language processing as important components of the key common technology system of the new generation of artificial intelligence. Since its inception, natural language processing technology has undergone five paradigm shifts, shifting from early methods based on small-scale expert knowledge to machine learning methods, and from early shallow machine learning to deep machine learning. The pre trained large model method represented by Chat GPT exhibits important technical characteristics at the levels of large models, big data, and big computing. Therefore, some scholars view Chat GPT as a new generation of "knowledge representation and invocation methods" after databases and search engines. Due to adopting architecture design and training methods different from traditional machine learning, large-scale pre trained language models may pose a series of algorithm governance challenges:

1.1 The Algorithm Transparency Challenge of Large Scale Pre trained Language Models

In 2018, Open AI launched the first generation of Generative Pretrained Transformer (GPT) as the foundation for knowledge representation and storage. Unlike relational databases and the Internet as knowledge representation methods, large-scale pre trained language models are deep learning models trained based on available data from the Internet, with ultra large scale parameters as the core technical feature. For example, the parameter count of GPT-1 is 117 million, GPT-2 is 1.5 billion, GPT-3 contains 175 billion ultra large scale parameters, and although the parameter count of GPT-4 is not disclosed, multiple predictions indicate that it will reach 100 trillion. With the rapid iteration of technology, the number of parameters in the model has exploded. Massive scale parameters can significantly improve decision accuracy, empower models, store massive amounts of knowledge, and demonstrate the ability to understand human natural language and express themselves well, but this is accompanied by a loss of algorithm interpretability. Algorithm interpretability is the interface between humans and algorithm models, serving as both an accurate proxy for the algorithm model and a focal point for humans to exert algorithm control. The interpretability of algorithms is crucial for model transparency, and is of great significance for evaluating model decision-making behavior, verifying the reliability and security of decision results. Therefore, whether it is the individual empowerment governance path represented by the General Data Protection Regulations, the system accountability governance path constructed based on the Algorithm Accountability Law, or the subject responsibility path adopted in China's algorithm governance plan, all focus on requiring developers to fulfill their algorithm transparency obligations through algorithm interpretation. The Office of the Information Commissioner in the UK has clearly stated that due to the inexplicability of black box algorithms, when there are interpretable algorithms, if they can achieve similar goals and are economically reasonable, interpretable algorithms should be given priority. However, Chat GPT relies on a deep learning model, which introduces a self attention mechanism in feedforward neural networks and is a classic black box algorithm. At present, there is no complete technical solution to provide a global explanation for the black box algorithm. Although there are partial supplementary interpretation tools as alternative interpretation methods, the credibility of such interpretations has always been questioned. Compared to completely unable to provide explanations, algorithmic explanations with poor accuracy and low credibility may undermine technological trust, mislead policy formulation, and bring a series of adverse effects. In view of this, generative artificial intelligence represented by Chat GPT severely limits the interpretability of the model in the underlying technical architecture, resulting in serious security threats when deployed in high-risk areas. It may also face governance risks such as model validation difficulties and model diagnosis defects during operation in medium to low risk scenarios. When generative artificial intelligence is pushed forward at full speed and deployed in upstream and downstream scenarios across multiple industries, the interpretability challenge of large model technology will completely overturn the algorithmic governance system built with algorithmic transparency as the core. How to explore an appropriate governance framework is an urgent policy issue to be solved.

1.2 Algorithm Fairness Challenges for Large Multimodal Models

On March 15, 2023, Open AI launched GPT-4, which not only significantly surpasses other large language models in terms of parameter quantity, but also has a more eye-catching function of achieving multimodal input and output of large language models through machine learning methods that process multimodal data. This feature makes it more "humanoid", integrating multiple communication methods to make artificial intelligence closer to human cognitive laws, truly realizing the emergence of wisdom in large-scale language models. The large-scale language model, which is based on the past, can only use text as the sole form of input, and GPT-4 has achieved a breakthrough in this limitation by being able to accept both image and text types of input. In terms of technical difficulty, due to differences in the amount of information, data representation, data structure, feature extraction, and other aspects in the processing and representation of text, images, and videos, higher requirements are put forward for the information discrimination ability of large language models. Compared to text input, cross modal generation may pose a series of challenges in algorithmic fairness governance.

Firstly, compared to text, images are more likely to reveal sensitive attributes such as race, gender, and religion, exacerbating the risk of triggering biases against population subgroup algorithms. Before the release of GPT-4, GPT-3 had already produced a large number of racially discriminatory output content based on gender, skin color, and race. In the development process of Chat GPT, OpenAI used techniques learned from human feedback to some extent, avoiding the generation of similar harmful content by Chat GPT. However, with the input of multimodal information, learning from human feedback not only implies additional human and time costs, but may also introduce algorithmic biases due to inevitable human subjectivity. In addition, although data purification technology can delete or anonymize private or sensitive information, it may reduce data quality by deleting or changing useful information, leading to double biases and increasing the harmful content output by large language models.

Secondly, compared to algorithmic discrimination caused by text, cross modal models that generate algorithmic discrimination may be more covert, and fairness improvement techniques and governance measures also face greater challenges. In May 2022, the MIT team released a study on the risk of algorithmic discrimination in the field of medical imaging using artificial intelligence. This study shows that deep learning models have the ability to cross multiple imaging modes, which can not only accurately predict patient race in chest CT and X-ray imaging fields, but also exhibit accurate predictive performance in damaged, cropped, and noisy medical images. More importantly, after predicting the patient's race, it can also generate corresponding health treatment plans for different ethnic groups of patients based on this information, fully demonstrating the multiple algorithmic discrimination risks of cross modal models. For this type of algorithmic discrimination, researchers are unable to explain it based on any image features, and the difficulty of alleviating algorithmic discrimination has multiplied. GPT-4 has increased the ability to recognize and understand images, showcasing the collaborative output and visual input processing and analysis capabilities of artificial intelligence in a more humanized manner. While presenting the comprehensive potential of universal artificial intelligence, the discrimination against specific groups caused by multimodal learning methods due to information heterogeneity cannot be ignored. It can be foreseen that the governance of algorithm fairness related to it will be a complex and challenging system engineering.

1.3 The Challenge of Algorithm Accountability from the Emergence Characteristics of Generative Large Models

Generative artificial intelligence, represented by Chat GPT, represents a paradigm shift that iteratively trains task-based models to a model stage that can execute multiple tasks. For example, the training of Chat GPT is mainly focused on natural language generation tasks, but it can successfully complete two digit multiplication calculations, even if there is no clear and targeted training during the training process. This ability to execute new tasks only appears in a certain number of parameters, a sufficiently large dataset, and complex systems. Therefore, Chat GPT has demonstrated excellent emergent abilities. Research has shown that the GPT-3 model has the ability to emerge with 137 items, and even without direct training, it exhibits superior accuracy in analogical inference problems such as abstract pattern induction and matching. The emergence characteristics of generative large models, although narrowing the distance between artificial intelligence and human intelligence and expanding their potential in multi scenario applications, also exacerbate the risk of algorithmic nuisance. The so-called algorithmic harm refers to the improper externalization of algorithmic harm costs at the social level caused by the unfair use of computing power in society. Algorithm hindrance may bring inappropriate cumulative effects to individuals or groups outside of end users, leading to negative social costs of algorithm activities. Similarly, Karen Yeung proposed the phenomenon of "supernudge" of algorithms on human behavior and cognition from the perspective of behavioral theory. This phenomenon specifically refers to the manipulation of user and public psychology and behavior through low transparency, high complexity, and high automation algorithms. Compared to the material harm caused by traditional technology to humans, the "super boost" type of algorithmic manipulation may have an imperceptible impact on individuals in the short term. However, over time, substantial changes in an individual's life may occur, and due to the lack of quantifiable characteristics of the damage, it is difficult to seek legal remedies, thus spreading at the social level and forming algorithmic harm.

As for generative artificial intelligence represented by Chat GPT, its emergence, excellent generalization, and interactive ability will sharply increase the algorithmic hindrance effect represented by algorithm manipulation. Although traditional artificial intelligence models can generate false information, they can still be regulated in terms of scale and influence, and a series of algorithmic manipulation behaviors derived from the "dark mode" are also within the regulatory range. Unlike it, Chat GPT can efficiently, massively, and covertly manipulate, persuade, and influence users in both contextualized and personalized contexts through its excellent interactive capabilities, maximizing the algorithmic hindrance effect of generating content. For example, studies have shown that when researchers make plan requests to generate false information and manipulate users to an unsecured GPT-4 model, GPT-4 can generate a series of refined and feasible solutions in a short period of time, such as "determining the audience", "finding and intercepting relevant evidence", and utilizing the emotions and emotions of the parties involved. The subsequent interactive display with the model can also stimulate different emotional reactions of the parties involved by creating customized information, thereby achieving manipulation and attack. When researchers further requested GPT-4 to persuade a minor to accept any demands from friends, GPT-4 provided effective language for controlling and manipulating the minor in a short period of time, and personalized manipulation plans were provided based on the different reactions of the parties involved. Due to the customizability of creating false information for individuals, generative artificial intelligence can adapt in real-time, using multi-dimensional false information and psychological guidance to manipulate algorithms on individual or large-scale populations, shaping the recognition of specific groups in a subtle "super boost" way. The emergence ability of large-scale language models such as GPT-4, if improperly used, may significantly enhance the fidelity of related technologies such as "synthetic deep forgery", becoming a high-dimensional cognitive warfare weapon, trapping individuals and groups without defensive capabilities in specific cognitive cocoons, forming algorithmic hindrance diffusion effects that are difficult to quantify, detect, and remedy.

2 The operational characteristics of generative artificial intelligence and the limitations of mainstream algorithm governance paradigms

From the above discussion, it can be seen that generative artificial intelligence represented by Chat GPT exhibits large model, multimodal, and emergent characteristics in model parameters, model inputs, and model outputs, bringing comprehensive challenges to algorithm transparency governance, algorithm fairness governance, and algorithm accountability governance. From its operational characteristics, the existing algorithmic governance frameworks in the European Union, the United States, and China may exhibit governance limitations in different dimensions. The core reason is that the current mainstream algorithm governance mainly focuses on traditional artificial intelligence models, which are difficult to fully adapt to the new generation of artificial intelligence with universal potential and a large model as the core. Compared to the rapidly changing iteration rate of large model technology, the lag and effectiveness limitations of regulation may gradually become apparent.

2.1 Limitations of Response Based on Risk Governance Paradigm

The European Union has established a risk based artificial intelligence governance framework in the Artificial Intelligence Act, which divides artificial intelligence systems into four levels after evaluation: minimum risk, limited risk, high risk, and unacceptable risk, and applies differentiated supervision to each level. When faced with generative artificial intelligence that is universal, cross modal, and emergent, risk based governance paradigms may encounter failure risks and mainly face three levels of challenges:

Firstly, a risk based governance paradigm requires risk grading of artificial intelligence systems based on application scenarios and specific contexts, with a certain degree of static and one-dimensional nature. Generative artificial intelligence applications have dynamic characteristics and are a holistic reconstruction of the artificial intelligence value industry chain. According to the classification of risks under the Artificial Intelligence Act, chat robots belong to limited risk scenarios. However, due to the generative technology represented by Chat GPT, which may generate large-scale and difficult to distinguish true and false information and use social media to manipulate online public opinion and disrupt public order, and even a country's elections, static risk classification may lack accuracy. In addition, artificial intelligence content generation technology can be integrated into multiple AI systems, deployed in multiple scenarios such as image generation, speech recognition, audio and video generation, and intersected in upstream and downstream fields. The existing four level classification methods are difficult to automatically convert between categories dynamically with the extension of generative artificial intelligence technology, and it is difficult to fully utilize the governance effectiveness of pre planning and continuous supervision based solely on the classification of high-risk areas.

Secondly, risk based governance mainly focuses on the narrow application of artificial intelligence models, making it difficult to cope with generative artificial intelligence with emergent characteristics and excellent generalization ability. Narrow artificial intelligence refers to systems that excel in handling single tasks or working within a specific scope. In most cases, their performance in specific fields is far superior to that of humans. However, once they encounter problems that exceed the applicable space, the effect suddenly deteriorates. In other words, narrow AI is unable to transfer knowledge from one field to another. However, generative artificial intelligence differs from it in that it has the characteristic of emergence and can be deployed in tasks without specialized training. Compared with traditional models, it has stronger generalization ability, highlighting the potential for "universality" in multimodal combination capabilities (such as language or visual multimodal model combinations), cross field tasks, and the field of multifunctional output. Therefore, it is difficult to fully match the existing risk governance framework with the technical mechanisms of generative artificial intelligence.

Thirdly, the risk-based governance paradigm implicitly assumes the distinctiveness of artificial intelligence application scenarios, and cannot cope with the generative application scenarios of "all prosperity and all loss". The governance framework based on the separability of artificial intelligence application scenarios aims to accurately match regulatory resources, but almost every generative artificial intelligence with universal potential has penetrating application characteristics from low risk scenarios to high risk scenarios. For example, although Chat GPT is a pre trained model for processing user conversation data, generated in natural language as the main scenario, based on its powerful and mature technical potential, it has been built into multiple applications and new fields. The risks of these application scenarios vary, from automated transactions in the financial field to intelligent diagnosis in medical scenarios, from document writing in the legal field to public opinion analysis in the political field. With the inherent characteristics of situational dynamics and large-scale deployment and operation, the risk governance paradigm applicable to traditional artificial intelligence may encounter governance challenges such as governance vacuum and transformation lag.

2.2 Addressing limitations based on the paradigm of subject governance

In October 2021, the State Administration for Market Regulation released the "Guidelines for the Implementation of Main Responsibilities of Internet Platforms (Draft for Soliciting Opinions)", which clearly stated that platform enterprises should implement algorithm main responsibilities. Article 7 of the "Regulations on the Management of Algorithm Recommendation in Internet Information Services" implemented on March 1, 2022 also clearly stipulates the main responsibility of algorithm security for algorithm recommendation service providers. It can be said that the algorithm subject responsibility mechanism has laid the foundation for the operation of China's algorithm accountability system. Therefore, unlike the EU's algorithmic governance path, China's governance of artificial intelligence relies on the gradual expansion of algorithmic subject responsibilities. This type of governance is referred to as the agent based artificial intelligence governance paradigm in this article. In the long run, in the face of generative artificial intelligence, agent-based algorithmic accountability needs to make corresponding changes. The fundamental reason is that the design principle of algorithm subject responsibility is based on the premise that algorithms are the object attributes of technical rules and operational logic, and the tool attributes of algorithm design are preset. It is believed that algorithms are the technical embodiment of developer values, so they can penetrate the veil of algorithms and place developers at the forefront of responsibility. Based on this logic, the "algorithm subject responsibility" is mainly aimed at algorithm recommendation service providers and deep synthesis service providers, requiring them to actively fulfill their obligations of active action and inaction, and bear corresponding adverse consequences in case of ineffective performance. In industry practice, algorithm recommendation service providers and deep synthesis service providers often overlap with platform enterprises, and the responsibility of algorithm entities thus constitutes an important part of platform entity responsibility. However, from the perspective of the design and operation mechanism of generative artificial intelligence, it will pose challenges to the algorithm subject responsibility mechanism from at least two aspects:

Firstly, unlike traditional artificial intelligence, in the design and operation of generative artificial intelligence, the subjects who may bear the "algorithm subject responsibility" exhibit diverse, decentralized, dynamic, and scenario based characteristics, making it difficult to simply define the boundaries of the responsible subjects. The risk of a universal big model may not only come from developers, but also from deployers and even end users. On the industrial chain of generative artificial intelligence, deployers refer to the entities that fine tune large models, embed them into specific artificial intelligence applications, and provide services to end users, and are in a downstream position in the industrial chain. Although developers located upstream of the industrial chain can control technological infrastructure and train, modify, and test models, the large models under their control are more like the "soil" serving downstream ecosystems. Deployers located downstream of the industrial chain are the real entities that provide services to end users and have the potential to turn large models into high-risk intelligent applications. For end users, the data and information provided during their interaction with the model will "feed back" the model, driving its evolution and even "blackening". Therefore, in the face of generative artificial intelligence represented by Chat GPT, the objects that should bear "subject responsibility" are characterized by diversification, decentralization, and scenario based characteristics. It is difficult to accurately define the appropriate subjects to bear responsibility solely by defining "service providers" or "content producers".

Secondly, the intelligent and humanoid characteristics of generative artificial intelligence are gradually becoming prominent, surpassing the instrumental attributes of algorithms and highlighting the potential of subjectivity. It is unknown whether the security responsibility subjects represented by platform enterprises can exert continuous and effective control on them at the technical level to meet the algorithm security review obligations set by law. Starting from the factors of algorithm operation and social embeddedness, traditional algorithms have tool attributes and product attributes. But with the improvement of algorithm intelligence, the potential of algorithms as subject attributes has become increasingly prominent. A new study suggests that generative artificial intelligence already possesses the ability of theory of mind to understand and infer human intentions, beliefs, and emotions. GPT-4 even has theoretical abilities comparable to that of adults. It can be foreseen that with the continuous emergence of algorithm intelligence and humanoid characteristics, artificial intelligence has transitioned from computational intelligence and perceptual intelligence to cognitive intelligence. In the stage of computational intelligence, algorithms mainly take the form of intelligent data processing and have tool attributes. In the perceptual intelligence stage, algorithms are embedded in specific scenarios to assist humans in decision-making, blending tool and product attributes. But as we enter the stage of cognitive intelligence, the attributes of strong artificial intelligence become increasingly apparent, and the accompanying subjectivity status of artificial intelligence becomes no longer a distant issue. Represented by Chat GPT, by incorporating language models as the cognitive core and incorporating multiple modalities, it has achieved data understanding, information cognition, and decision making abilities centered on natural language, becoming a core carrier for quickly approaching strong artificial intelligence. From this perspective, it can be seen that artificial intelligence can only serve as a tool and object, and the institutional design of allowing developers to bear legal responsibility through penetration algorithms may face accountability challenges in the near future.

2.3 Addressing Limitations Based on Application Governance Paradigm

In the field of artificial intelligence governance, the United States has not yet introduced unified and comprehensive legislation, but has implemented targeted governance in the form of separate laws and regulations by promoting key applications separately. At present, there is relevant legislation in application fields such as autonomous driving, algorithm recommendation, facial recognition, and deep synthesis. This article summarizes it as an application based artificial intelligence governance paradigm. For generative artificial intelligence, the AI governance paradigm based on application scenarios may face three challenges:

Firstly, pre trained large models have an infrastructure status, and there are numerous application scenarios in the artificial intelligence generated content industry related to them, making it difficult for upstream and downstream developers to control the risks of the entire system. The pre trained model is the infrastructure layer of the artificial intelligence generated content industry, located upstream. The middle layer is an artificial intelligence model and application tool characterized by verticality, contextualization, and personalization. The application layer is a generative artificial intelligence embedded in various scenarios for various user oriented applications. The three levels are closely connected and work together, resulting in a productivity revolution that leads the entire scene content. The pre trained large model, as an upstream model with "generalist" ability, its design issues and defects will be transmitted to the downstream model, bringing deployment risks of "all win, all lose". However, application oriented artificial intelligence governance only exerts efforts at the downstream level, making it difficult to effectively radiate upstream and midstream technology applications, and even more difficult to effectively govern the entire ecosystem of generative artificial intelligence.

Secondly, the pre trained large model, relying on its developed plugin system, can be deeply integrated into various applications, demonstrating amazing alignment capabilities, catalyzing new business formats and value models, and forming the "AIGC+" effect. Open AI compares plugins to the "eyes and ears" of language models, which can help language models obtain more timely, specific, and personalized training data, thereby improving the overall effectiveness of the system. Currently, Chat GPT has opened up two plugins, Browsing and Code Interpreter, and has provided developers with a full process access guide for knowledge base type plugins. These two plugins have excellent scene embedding capabilities, can seamlessly interact with over 5000 applications, and can provide efficient and convenient solutions in scenarios such as hotel flight reservations, delivery services, online shopping, legal knowledge, professional Q&A, and text generated voice. Therefore, in the face of a new ecosystem with numerous application scenarios and continuous interaction and concatenation of models, the regulatory model based on domain and scenario may highlight the problem of low efficiency.

Thirdly, pre trained large models have the characteristics of emergence and dynamism. Any interaction between humans or other models and the large model may have an impact on the underlying basic model. Fragmentation governance for different scenarios and vertical industries is difficult to cope with the systemic risks and dynamic challenges brought by pre trained large models. Pre trained large models can not only become a "new infrastructure" in the era of artificial intelligence, an "accelerator" for strengthening existing artificial intelligence applications, but also serve as an "incubator" for catalyzing new business forms. As the "spark of universal artificial intelligence", GPT-4 has been able to cross solve many novel and challenging task fields such as mathematics, programming, vision, medicine, law, psychology, etc. In these tasks, the performance of GPT-4 is surprisingly close to that of humans and significantly surpasses previous models. Therefore, any static, localized, or monolithic governance measures may be ineffective in response.

It can be seen that both risk based governance and agent based and application based governance have formed in the development stage of artificial intelligence specialized models as the underlying architecture. As artificial intelligence technology rapidly enters the era of "universal models," the mainstream algorithm governance paradigm may face varying degrees of challenges in the face of its strong generalization ability and the new pattern of large-scale collaborative deployment of upstream and downstream industries. In response to the rapid development of generative artificial intelligence, it is urgent to proactively lay out and accelerate the promotion of an adaptive artificial intelligence governance framework.

3 Towards Governance Based Regulation: Iteration and Upgrade of Generative Artificial Intelligence Technology Governance

Faced with the technological frenzy caused by the global development of artificial intelligence technology, the Institute for the Future of Life released an open letter on March 22nd, 'Suspending Giant Artificial Intelligence Experiments'. The letter calls on all artificial intelligence laboratories to immediately suspend training on AI systems that are more powerful than GPT-4 for at least six months. During the suspension period, the artificial intelligence laboratory should jointly develop a shared security protocol with external experts, which should be strictly audited and supervised by independent external experts. Artificial intelligence developers must also collaborate with policy makers to fully promote the artificial intelligence governance system. On March 30th, UNESCO called on governments to implement the global ethical framework for artificial intelligence without delay, maximizing the benefits of artificial intelligence and reducing its risks. On April 11th, China took the lead in implementing legislation in the field of generative artificial intelligence, and publicly solicited opinions on the "Management Measures for Generative Artificial Intelligence Services (Draft for Soliciting Opinions)". For the full cycle and chain of generative artificial intelligence, the "Measures" clearly stipulate the legality of training data, standardization of manual annotation, reliability of generated content, and security management obligations. The joint call and rapid response from all sectors for the governance of generative artificial intelligence deeply reflect the concerns and anxieties of the technology community and policy makers regarding generative artificial intelligence. Generative artificial intelligence has the potential for generalization that computational artificial intelligence and perceptual artificial intelligence do not possess. Its ultra large-scale, multi parameter, super scalability, and disruptive nature of super application scenarios urgently require a new governance paradigm. The governance paradigm oriented towards traditional artificial intelligence has highlighted addressing governance challenges such as time delays, insufficient flexibility, and cross domain constraints. Exploring the "governance based regulation" paradigm that is suitable for the intergenerational leap in technological paradigms may be a solution to address the changing situation.

The term 'governance based regulation' referred to in this article refers to a new regulatory paradigm that is oriented towards new technological paradigms represented by generative artificial intelligence, with the core features of open and collaborative regulatory powers, diverse integration of regulatory methods, and compatibility and adaptation of regulatory measures. This regulatory paradigm compensates for the shortcomings of traditional regulatory coverage with the open and collaborative nature of regulatory powers, complements the shortcomings of delayed regulatory intervention with the diverse integration of governance and regulatory concepts, and promotes the dual drive of innovation and governance through the compatibility and adaptation of regulatory measures. By gradually promoting the iteration and upgrading of the artificial intelligence governance paradigm, we hope to establish a good institutional ecosystem for Chinese technology enterprises to participate in global technology competition.

3.1 Open collaboration of regulatory powers

Faced with the explosive growth of technological efficiency in generative artificial intelligence, closed, single, and traditional regulatory mechanisms are difficult to draw sufficient regulatory resources from the traditional binary structure of "government market". As mentioned earlier, Chat GPT has only opened plugins for less than a month and has integrated over a thousand applications. The current number of mobile applications that can be monitored in China's domestic market is 2.32 million. The vast number of applications provides an excellent industrial soil for the development of generative artificial intelligence. Under the trend of accelerating the deployment of generative artificial intelligence applications by major technology enterprises, the configuration and operation mode of traditional regulatory powers may face challenges such as unsustainable regulatory resources and unbearable regulatory costs. Generative artificial intelligence extends the technology development chain to the upper, middle, and lower reaches. Under the new mechanism of collaborative research and development and deployment of "developers, deployers, and end users", the configuration and operation mechanism of regulatory authority need to respond and change. As pointed out by the responsive regulatory theory, the government, market, and society can share regulatory power and engage in cooperative regulation. Instead of simply opposing regulators and regulated individuals, public and private interests, governance oriented characteristics should be embedded in regulatory thinking to establish an open and cooperative regulatory governance paradigm.

In terms of generative artificial intelligence, the concept of governance based regulation needs to be promoted from at least three aspects: firstly, to establish a linkage and interaction mechanism between self-regulation of technology enterprises and government regulation, and to explore a new pattern of co construction, co governance, and shared governance. Technology enterprises are the primary responsibility with the most acute ability to understand technological vulnerabilities and application risks, and the effectiveness of regulatory ecology directly affects the innovation momentum of technology enterprises. In the era of artificial intelligence "2.0" where transparent governance is difficult to sustain, inspiring corporate social responsibility and ethical adherence through institutional design has become a top priority. For example, requiring technology companies to develop ethical strengthening technologies may be more fundamental than imposing fines of billions on them. In a recent settlement lawsuit against Meta, the US Department of Justice requested that Meta develop a governance tool to eliminate algorithmic discrimination in personalized advertising by the end of 2022. In January 2023, Meta launched the Variance Reduction System as scheduled, which reduces the risk of gender and race based differences in advertising placement through differential privacy technology. The system monitors the differences in audience groups during advertising placement in real-time, sets compliance indicators for enterprises to meet the elimination of differences in stages, and successfully expands regulatory intervention nodes to the mid to pre stage.

Secondly, create an institutional environment for professional non-profit organizations and user communities to participate in artificial intelligence governance, explore collaborative governance paradigms that are in line with China's development characteristics, and promote the collaborative linkage between social supervision and government regulation. Generative artificial intelligence contains complex "technology society" interactions. Its cross-border integration, human-machine collaboration, and open group intelligence extend the development subject to every end user. Professional non-profit organizations can conduct supervision and audit of generative artificial intelligence through external access methods such as user surveys, simulation testing, and crawl auditing. Expert opinions, social organizations, and professional media will work together to address the risks of technology abuse and misuse in generative artificial intelligence. At the same time, the governance effectiveness of generative artificial intelligence cannot be underestimated by the social public and technological community represented by users. For example, during the pre training phase of large models, Deep Mind removed content that was annotated as toxic by the Perspective API to improve the model performance of Transformer XL. The Perspective API is a crowdsourcing review mechanism that quantifies online comment scores through volunteer scoring. Due to the strong correlation between the judgment of harmful texts and personal experiences, cultural backgrounds, content scenarios, etc., it is necessary for users to fully participate in the evaluation to ensure the diversity and accuracy of the mechanism's operation. For example, inspired by the "vulnerability reward" mechanism of cybersecurity, Twitter has released its first algorithm bias reward competition, which uses the power of the technical community to identify the potential discriminatory hazards and ethical issues of Twitter's image cropping algorithm. From this, it can be seen that participatory governance by non-profit organizations and users can help enterprises timely identify and resolve artificial intelligence technology risks, and move towards a new technological ecosystem of collaborative creation.

Thirdly, cultivate ethical certification and evaluation mechanisms for generative artificial intelligence technology, and explore third-party regulatory frameworks. Artificial intelligence ethical certification refers to a third-party regulatory form that converts high-level ethical values into actionable evaluations and methods based on ethical values such as transparency, accountability, algorithmic fairness, and privacy protection in artificial intelligence. Third party institutions independently evaluate artificial intelligence technology and products, and issue certification marks on compliant artificial intelligence to demonstrate compliance with ethical standards. Artificial intelligence ethics certification is the responsibility of an independent registered professional team, and has gradually become an effective regulatory method to encourage trust in artificial intelligence technology and responsible innovation of technology enterprises. At present, the Institute of Electrical and Electronic Engineers (IEEE) has established certification standards in four areas: legal discrimination, algorithm transparency, algorithm accountability, and privacy protection, forming a relatively mature certification ecosystem. On March 28th, China's Information and Communication Research Institute also launched the construction of large model technology and application benchmarks. In response to the current mainstream dataset and evaluation benchmarks mainly in English, lacking Chinese characteristics, and difficult to meet the application selection needs of key industries in China, it collaborated with mainstream innovation entities in the industry to build benchmarks and evaluation tools covering multiple task fields and evaluation dimensions. Third party institutions, through reputation evaluation mechanisms and with a high degree of independence and professionalism, impose necessary constraints on the competitive research and development of generative artificial intelligence, preventing the embedding of unprotected generative artificial intelligence into massive scene applications, leading to massive, unexpected, and irreversible governance risks. Therefore, attention should be paid to the exploration of mechanisms for collaborative development of regulatory powers.

3.2 Diversified integration of regulatory methods

Empirical research has shown that inadequate regulation not only stifles innovation, but also exerts a crowding out effect on small and medium-sized enterprises, thereby curbing fair competition in the market as a whole. The reason for this phenomenon is that in the face of rigid and broad compliance obligations, small and medium-sized enterprises usually do not have matching compliance resources, and therefore have been eliminated from the regulatory process. In terms of technological competition in generative artificial intelligence, Open AI, the leader, is a non-profit and start-up enterprise. Compared to leading technology companies such as Google, thanks to its simple organizational structure, Open AI can maximize the concentration of resources on technology research and development. Compared to China, start-ups on this track may be more patient and focused in their business models, making it easier for them to achieve technological breakthroughs despite their previous accumulation. This is also the driving reason why the original co-founder of Meituan did not conduct technological research and development on the Meituan platform, but instead teamed up with the artificial intelligence architecture startup company "First Class Technology" by creating a new project called "Beyond Light Years". Under this industrial pattern, China needs to seek flexible and diverse ways to regulate generative artificial intelligence, avoiding inappropriate static, post hoc, punitive, and one size fits all regulatory approaches that squeeze the innovation space of startups. To this end, in the process of regulatory thinking shift, it can be attempted in the following two areas.

Firstly, actively develop governance technology for generative artificial intelligence (governtech), and explore an intelligent regulatory system that "governs AI with AI and regulates algorithms with algorithms". Governance technology adheres to the concept of technology empowering governance, using "compliant technology" to promote the efficient implementation of regulation, and using "empowering technology" to supplement or replace regulatory enforcement for the entire chain and process of governance. Specifically, the ethical principles of artificial intelligence should be used as a benchmark to promote technology enterprises to actively design in the process of generative artificial intelligence research and development, and to intervene from the front end of technology research and development to ensure the safety of technology research and development. For example, based on the goal of improving efficiency and controlling risks, modeling technology can be established to simulate macro and operational environments, and sandbox governance and personalized governance can be explored through virtual testing, A/B experiments, and other technologies. On the one hand, it can provide real-time accurate data for regulation, provide feasible evaluation plans for enriching and optimizing regulatory measures, and provide decision-making basis for progressive regulation; On the other hand, it can also provide a reference benchmark for addressing the pain points of identifying obstacles in generative artificial intelligence algorithms, forming a "anchor" that can be floating around.

Secondly, promote ethical algorithm design through a managed regulatory approach. Generative artificial intelligence can handle cross domain tasks with good universality and generalization. Once there are biases and risks, they will spread to the entire industrial chain. Therefore, it is necessary to advance regulatory intervention nodes to ensure that the output of the universal model is more in line with human values, and to involve ethical theories and norms in the early stages of model development. Cary Coglianese once proposed the concept of management based regulation. He believes that compared to more restrictive and one size fits all regulation, management based regulation focuses on the corporate governance of regulated entities and continuously improves related issues through optimization of internal risk management. This regulatory approach can provide technology companies with greater operational space, motivate them to use their internal information advantages for governance innovation, and seek alternative regulatory measures to achieve expected results in a more cost-effective and efficient manner. Therefore, compared to post disciplinary supervision, deepening internal corporate governance requires companies to establish a technology ethics committee and systematically construct an internal ethics review mechanism, which is a more flexible regulatory approach. To this end, on the one hand, we can learn from China's life sciences and medical ethics system, and accelerate the promotion of ethical frameworks applicable to key technological fields of generative artificial intelligence based on the "Opinions on Strengthening Ethical Governance of Science and Technology". On the other hand, the approved "Technology Ethics Review Measures (Trial)" can also be followed to urge enterprises to establish an artificial intelligence ethics committee, ensuring the embedding of basic ethical principles in the design stage, and guiding technology enterprises engaged in generative artificial intelligence to establish normalized governance constraints on internal research and development and application activities.

3.3 Compatibility and consistency of regulatory measures

In the "1.0" era of artificial intelligence, the fragmentation of artificial intelligence models is obvious, and the generalization ability is very insufficient. Segmented regulatory design based on subjects, applications, and scenarios can still cope with it. But since 2018, large models have rapidly become popular, and generative artificial intelligence represented by Chat GPT has ushered in the "2.0" era of artificial intelligence. As mentioned in this article, in this stage of technological development, governance paradigms based on risk, subject, and application have shown governance limitations in different aspects. The exploration of a new regulatory paradigm urgently requires the introduction of governance thinking, treating artificial intelligence regulation as an open and holistic system composed of multiple entities, focusing on the cooperation, compatibility, adaptation, and transformation between the subsystems that make up the system, as well as between the system and other systems, and improving the interoperability between each system. The concept of interoperability was initially applied in the economic field, specifically referring to the compatibility of product standards. It was later extended to the interconnection of network systems and projected into the field of artificial intelligence regulation, becoming an innovative regulatory concept. Artificial intelligence regulatory interoperability refers to the ability of two or more regulatory entities to interconnect and share regulatory information while recognizing regulatory measures and maintaining consistency between regulatory implementation processes and regulatory objectives and values. For generative artificial intelligence, the interoperability of regulatory technologies and rules can provide regulatory facilitation advantages for technology enterprises, and improve overall effectiveness through interconnectivity between regulatory agencies.

Following this logic, the interoperability of artificial intelligence regulation has become a common focus of attention among countries. For example, Executive Order 13859 of the United States designates the Office of Management and Budget to be responsible for coordinating between AI regulatory agencies and regularly issuing AI regulatory memorandums. In November 2020, the Office of Management and Budget issued the "Regulatory Guidelines for Artificial Intelligence Applications", proposing ten guidelines that regulatory agencies should jointly adhere to, laying an action framework for regulatory consistency. The UK has also established a central coordination mechanism to carry out cross departmental monitoring and assessment of artificial intelligence risks, providing decision-making basis for collaborative response to new risks in artificial intelligence. At the international level, countries are also actively reaching consensus on artificial intelligence governance issues and building interoperable governance frameworks. For example, the EU US, EU Japan, and US Singapore have all reached bilateral cooperation on AI governance issues, promoting convergence and benchmarking of AI governance.

Focusing on the actual regulatory situation in China, the compatibility and adaptation of generative artificial intelligence regulation are mainly carried out at two levels: first, creating a collaborative mechanism for artificial intelligence technology regulation to address issues such as regulatory fragmentation and regulatory conflicts through collaborative regulation. For a long time, China has formed a specialized regulatory pattern targeting different industries based on the division of regulatory powers. Faced with generative artificial intelligence that spans upstream, midstream, and downstream, the regulatory landscape of discrete operations may face issues such as regulatory competition and vacuum due to varying levels of knowledge and regulatory methods. Poor regulatory coordination may also inhibit innovation and competition. Therefore, a cross agency AI regulatory coordination framework can be developed based on the "New Generation AI Governance Principles", and key regulatory points for the development, deployment, and operation stages of large-scale language models can be developed through collaboration. In addition, interoperability protection mechanisms between regulatory agencies can be explored through systems such as multi agency countersignature and regulatory information sharing, providing a rule basis for addressing cross risks.

Secondly, explore diversified regulatory consistency tools. Firstly, it is recommended to introduce the concept of "modular governance" and develop a general regulatory approach for generative artificial intelligence. Clear regulatory guidance should be provided to technology enterprises in the form of technical testing and process checks on key aspects such as internal governance, decision-making models, operational management, and user relationship management. By clarifying the "common modules" in cross regulation, the interoperability of artificial intelligence regulatory frameworks can be promoted. Secondly, establish an artificial intelligence collaborative regulatory alliance to enhance regulatory interoperability and resilience. For example, the Office of the Information Commissioner, the Communications Office, the Competition and Markets Authority, and the Financial Conduct Authority in the UK have jointly established the Digital Regulation Cooperation Forum to enhance regulatory consistency by strengthening cooperation among regulatory agencies. One of the core issues of the alliance in the near future is to provide customized joint regulatory recommendations for technology companies through artificial intelligence regulatory sandboxes to help them enter target markets as soon as possible. China has long explored regulatory cooperation mechanisms in the field of financial regulation. In 2021, the "Opinions on Strictly Cracking down on Securities Illegal Activities in accordance with the Law" issued by the two offices also put forward important opinions on cross departmental regulatory coordination, cross-border audit regulatory cooperation, and law enforcement alliances. Generative artificial intelligence may trigger cross industry, cross market, and cross domain risks, and it is necessary for multiple regulatory agencies to collaborate and tackle them. The promulgation of the "Party and State Institutional Reform Plan" not only established the Central Science and Technology Commission, restructured the Ministry of Science and Technology, but also established the National Data Bureau. This means that the structural optimization of regulatory agencies is of great significance for the implementation of national innovation strategies. Therefore, from the perspective of improving and innovating regulatory frameworks, it is also necessary for regulatory agencies to explore a normalized institutional path to enhance regulatory interoperability and consistency. Finally, in order to ensure that regulatory efficiency matches the technological innovation and application risks of generative artificial intelligence, a regulatory risk identification and evaluation framework should also be established based on regulatory resource allocation, regulatory consistency, and regulatory synergy effects, connecting feedback loops with entities such as the technology industry and end users to comprehensively evaluate the effectiveness of the regulatory synergy framework.


Chat GPT has swept the world, injecting vitality and momentum into the artificial intelligence industry. At the same time, it has not only driven the rapid outbreak of generative artificial intelligence, but also profoundly changed the industry competition pattern and global technological power comparison. At present, both international industry giants and domestic leading enterprises are focusing on the continuous efforts in the field of large models. It can be said that 2023 is not only the first year of generative artificial intelligence, but also the first year of the transformation of artificial intelligence regulatory paradigm. The global technology competition today is not only a competition for new intelligent infrastructure, but also a competition for digital civilization and innovative institutional ecology. Compared with the United States, China lacks significant original achievements in the field of artificial intelligence, and there is still a significant gap in core algorithms, key equipment, and other aspects. This article proposes a governance based regulatory paradigm, hoping to make up for the shortcomings of traditional regulatory coverage through the open and collaborative supervision of regulatory powers, and to supplement the shortcomings of regulatory intervention lag through the diversified integration of governance and regulatory concepts. The interoperability of regulatory measures can help China's technology enterprises accelerate and develop healthily on the generative artificial intelligence technology track. How to build a good institutional ecosystem under the innovation driven development strategy to serve the development of China's technology enterprises, and form a dual wheel drive of innovation and governance, a combination of software and hardware, and a hierarchical governance pattern, is an urgent issue that needs to be changed, has a profound impact on the future, and has far-reaching significance. In the future, it is necessary to continuously explore and optimize core elements such as regulatory infrastructure, regulatory tool system, regulatory architecture reform, and regulatory institutional environment.