[author]Guo Chunzhen
[content]
Guo Chunzhen
Professor of Law School of Xiamen University, doctoral supervisor, and researcher of the Inner Party Law Research Center of Xiamen University.
Keywords: generative AI; Generative pre training model (GPT); Integrated governance; Legal 3.0
Introduction
Generative AI refers to an artificial intelligence system that can autonomously generate new content such as text, images, audio, etc. The Generative Pre trained Transformer (GPT) is one of the important models in generative AI, and its new progress from text models to multimodal models has sparked a wave of research and development in the industry. In November 2022, once the Chatbot ChatGPT developed by American Open AI Company was launched, the number of active users would exceed 100 million in just two months, becoming the fastest growing consumer application in history, and its growth rate was far faster than that of applications that had previously shown outstanding user growth rate. But this record was quickly broken again, as Microsoft embedded ChatGPT into its Bing search engine, and a month later, Bing's daily active users exceeded 100 million. Around the generative AI industry, Microsoft, Google, Meta, Amazon, and domestic companies such as Baidu and Alibaba have all quickly followed suit, hoping to occupy a place in this new technology field.
Generative AI not only brings enormous room for productivity improvement, but also brings a series of impacts and risks to the development of related industries and the protection of citizens' rights and interests. On the one hand, generative AI is not only a production tool, but also a production factor. It can not only be applied as a technological means in the production process, but also carry out a certain degree of innovation in this process, replacing certain human work, and even bringing revolutionary progress to productivity. On the other hand, precisely because of its efficient and high-quality performance in content production, it has brought potential huge impacts and caused psychological panic to the development of industries including secretarial, education, software design, image design, etc. At the same time, during the training process, it needs to be fed the relevant legal issues related to the sources of massive data, as well as the potential discrimination and "blackening" issues in the content it produces, which pose potential risks to data security, algorithm security, network rumors, data sovereignty, national security, etc. Therefore, how to understand and treat generative AI from a legal perspective, and how to treat its industrial development and risk prevention rationally from a legal perspective are important issues currently facing us.
1 The emergence and development of generative pre training model GPT
Discussing the issues caused by technology from the perspectives of law and law does not mean that legal research is playing a solo role in a down-to-earth manner. When weighing whether and how technology should be governed, understanding the basic content and characteristics of technology is a basic respect for technical issues, and also a prerequisite and foundation for making legal research meaningful and enabling law to play a role. Therefore, it is necessary to provide a brief introduction to generative AI.
With the explosive growth of emerging technologies and resources such as large-scale datasets, computing power, and algorithms, generative AI represented by GPT is rapidly developing. The Transformer model based on self attention mechanism is the foundation of GPT technology, which is a deep learning model for processing sequence data. The Transformer model is a deep learning model used for processing sequential data, and is one of the key technologies that have made significant progress in generative AI. It solves the problems of traditional models in dealing with long text, parallel computing, information loss and long-distance dependency. It is stronger than traditional Natural language processing models in semantic feature extraction, long-distance feature capture, task comprehensive feature extraction, parallel computing and operation efficiency, and leads the revolution in the field of Natural language processing. Open AI has developed GPT, GPT-2, GPT-3, and ChatGPT based on the Transformer model and iteratively upgraded them to GPT-4. Generative AI has rapidly evolved from text models to multimodal models such as text and images.
In June 2018, Open AI published a paper introducing a Generative Pre trained Transformer (GPT) model based on Transformer. GPT improves the shortcomings of traditional Natural language processing models that are difficult to process long texts and require a large number of labeled data. It obtains the initial parameters of the model through unsupervised pre training, and adapts to different downstream tasks through supervised fine-tuning. GPT can generate coherent and grammatically correct text according to given prompts, and perform some simple Natural language processing tasks. However, due to the fact that the training dataset of GPT only has 117 million parameters, its generation ability, quality and coherence of generated text, and length of generated content are often limited, lacking complexity and diversity.
In February 2019, Open AI launched the GPT-2 model. Compared to GPT, GPT-2 lacks much innovation and design in its structure, but instead uses more parameters and a larger dataset. The parameter count of GPT-2 is 1.5 billion, and due to the significant increase in model size and parameter count, GPT-2 can better handle complex natural language tasks. In addition to understanding ability, GPT-2 can generate long, coherent, and diverse texts, not only able to read abstracts, chat, continue writing, and make stories, but also able to generate misleading news, spam, or phishing content. At this point, GPT-2 demonstrated universal and powerful capabilities and achieved optimal performance on multiple specific language modeling tasks at that time.
In May 2020, Open AI launched the GPT-3 model. GPT-3 has a larger model size and number of parameters, including multiple versions, with the largest version containing 175 billion parameters and 96 Transformer encoder layers. Compared to GPT-2 and GPT-1, the significant improvement of GPT-3 not only comes from a larger model size and number of parameters, but also from more advanced architecture, more training data, and more optimization techniques. GPT-3 can almost complete most tasks of Natural language processing, such as question answer search, reading comprehension, semantic inference, Machine translation, article generation and automatic question answering, etc. At the same time, the model performs well in many tasks, reaching the best level at that time in the "French English" and "German English" Machine translation tasks. The automatically generated articles almost make people unable to distinguish whether they are written by people or machines (the accuracy rate is 52%), and can also automatically generate code according to the task description. GPT-3 can convert web page descriptions into corresponding code, imitate human narratives, create custom poetry, and generate game scripts.
In February 2022, Open AI announced the Instrument GPT model. This model is a new round of enhancement optimization based on GPT-3. The main difference is the introduction of human feedback Reinforcement learning (RLHF), also known as GPT-3.5. In November 2022, Open AI launched the new AI Chatbot ChatGPT based on the Instrument GPT, which also uses the human feedback Reinforcement learning training model. Users only need to request ChatGPT to achieve functions such as article creation, code writing, and answering questions. In addition to being able to interact with humans in language, ChatGPT can also perform various relatively complex language tasks, including automatic text generation, automatic question answering, automatic summarization, and even write and debug computer programs. Although ChatGPT performs well in Natural language generation, it still has some limitations. For example, the generated text may lack common sense knowledge; Susceptible to prejudice and discrimination; There are issues with dialogue coherence; Vulnerable to adversarial attacks, resulting in incorrect or offensive output.
In March 2023, Open AI released the latest version of GPT-4. GPT-4 is a multimodal model, which means it can not only understand the "modal" information of text, but also accept images as input and generate subtitles for classification and analysis. GPT-4 performs well on various professional and academic benchmarks, is more creative and collaborative than previous models, and can generate, edit, and work with users to complete creative and Technical writing tasks, such as creating songs, writing scripts, or learning users' Writing style. In addition, it can also handle text with over 25000 words, allowing for long content creation, extended conversations, and document search and analysis. Some researchers believe that GPT-4 possesses common sense knowledge from around the world and can make inferences based on it, which can be regarded as an early version of AGI (General Artificial Intelligence). In the testing conducted by the Microsoft team, GPT-4 passed Amazon's simulated interview with a full score, surpassing all human participants in the testing and can be hired as a software engineer.
2 The Disputes surrounding GPT and Their Roots
As a technological reality, generative AI is developing exponentially and influencing our world; As a high-tech application, it has the potential to benefit humanity in areas such as economy, education, and daily life, but it may also bring huge risks to individuals, society, and government. Although legal scholars are contributing a large amount of theoretical exploration, current affairs commentary, legal and policy proposals, and calling on businesses and international organizations to inject ethics into artificial intelligence, due to the lack of sufficient knowledge and technological capabilities in artificial intelligence, they are often ridiculed as "blind men and elephant like" research, making it difficult to undertake the task of solving these problems. However, it is still necessary to conduct preliminary research on the different theoretical perspectives of generative AI.
2.1 Disputes surrounding GPT
When discussing how to face generative AI represented by GPT and its future prospects, there are many controversies in both academia and industry. Overall, there are three main attitudes: support, opposition, and neutrality.
Many scholars and experts highly evaluate GPT as a disruptive breakthrough and believe that it will make a difference in empowering and empowering. In the 2022 "Annual Breakthrough" released by Atlantic magazine, Derek Thompson regards ChatGPT as an incredible tool, and believes that it may change our views on the way we work, think and the true meaning of human creativity. Bill Gates regards it as the most important innovation at present and the most revolutionary technological progress since 1980. It will change our world and make it more efficient. It not only provides optimization that has a huge impact on reading and writing, but also can effectively improve efficiency and significantly improve outcomes in the healthcare and education fields. It can even create vaccine proteins that can cure or prevent diseases, provide safer automated driving, and minimize cognitive errors in all fields. Yu Guoming believes that ChatGPT utilizes human feedback signals to directly optimize language models, producing texts that match human common sense, cognition, and needs as much as possible, bridging the "ability gap" and "usage gap" in information integration and use among people at different levels and regions, and is another significant empowerment and empowerment of humanity. At the same time, it breaks through the differences in the ability of different individuals to use and integrate resources, enabling everyone to engage in social content production and dialogue based on semantic expression and resource mobilization capabilities above the social average, thereby achieving empowerment and empowerment of vulnerable groups.
Scholars opposing GPT mainly discuss from the perspectives of knowledge growth and human subjectivity, power control, and legal risks. Scholars represented by American philosopher and linguist Noam Chomsky criticized from the perspective of knowledge growth and human subjectivity. Chomsky views ChatGPT as a form of high-tech plagiarism, a way of avoiding learning, and believes that students instinctively use technology to avoid learning is a "sign of educational system failure". Alice Wu and others emphasized that ChatGPT and AI Chatbot simply repeat general information, and their rise may weaken our creativity. Wang Min'an believes that the text generation of ChatGPT is an abstract and generalized Structuralism activity, whose purpose is to find the most balanced meaning, which will lead to less and less personalized and living things and may bring about what Foucault called "death of human beings" in a sense. Jiang Qiping believes that generative AI represented by ChatGPT is difficult to use for cutting-edge professional judgment due to its lack of unique insights. The viewpoint of criticizing ChatGPT from the perspective of power control and legal risk emphasizes that ChatGPT and AIGC are actually extensions of traditional power and further strengthening of the centralized monopoly model of the Internet. Every industry and company related to AIGC stands behind a giant, even the giant itself. If the fundamental value of blockchain technology is "decentralization", we hope to break this monopoly and reconstruct a new distributed network, allowing ordinary people to regain their data sovereignty. So, ChatGPT and AIGC completely ignore human sovereignty and recklessly capture global data for training, ultimately creating their own "super brain". Oligopolistic companies that control such super brains not only have external influence and control over consumers or users, but also have more direct control over people's decision-making through the "super brain". In Europe, the main legislators of the AI bill, Benifei and Tudorachi, proposed that AI systems generating complex texts without human supervision should be included in the "high-risk" list. In addition, Mark G. Murphy further pointed out that ChatGPT and Open AI represent an unconscious lack of responsibility, which poses a threat to social bonds and does not provide any solutions for the isolation and basic isolation and confrontation we still face.
At present, both domestic and international academic and practical circles hold a more neutral stance on GPT. This stance takes a dialectical perspective and considers the revolutionary impact, technological priority, and drawbacks of ChatGPT technology. Ted Lieu, the Democratic representative of the Science Committee of the United States House of Representatives, said in a recent commentary published by the New York Times that he was both "excited about AI" and "scared by AI, especially AI that is not under control and supervision." Sam Altman, CEO of Open AI, which developed ChatGPT, also believes that, The best scenario for artificial intelligence is' incredibly good ',' I can imagine having incredible richness and systems that can help us break the deadlock, improve all aspects of reality, and lead the best life for all of us', but he worries that the worst scenario is' everyone will be out of the game 'and' even more worried about accidental misuse in the short term ', So he believes that "it is impossible to exaggerate the importance of artificial intelligence security and calibration work". Michael J. Ahn and Yu Che Chen pointed out that ChatGPT may completely change the way we interact with technology, with the potential to help institutions achieve higher efficiency. But this may make people heavily dependent on the use of technology, leading to a decrease in our ability to remember specific facts and think critically, further reducing the position and role of civilians in decision-making.
In China, some scholars believe that ChatGPT, as a major disruptive application in the field of generative AI, has broad commercialization prospects and will drive rapid changes in many industries, with the potential to quickly land in areas such as customer service, healthcare, education, and home care. However, ChatGPT is currently in its early stages of development, with prominent issues such as immature development of key core technologies, incomplete algorithm models, insufficient understanding ability, insufficient flexibility in answering questions, and excessive reliance on corpora. It is still a long time before large-scale commercial applications can be achieved. Some scholars also believe that ChatGPT may not only make teachers and students rely too much on it and lead to more assimilation and simplification of future students' thinking, but also technically return to classical critical thinking represented by Socratic dialogue. Some scholars believe that if human civilization based on the brain structure of Homo sapiens and the understanding and communication of natural language can be called "Anthropocene Civilization", then the emergence of ChatGPT and AIGC in a sense foreshadows the arrival of "post Anthropocene Civilization", which will lead humanity to an uncertain future where hope and fear coexist.
2.2 Reasons for disputes surrounding GPT
There are always reasons behind any viewpoint that support it. Although it is often difficult or even impossible to argue for a causal relationship between them, analyzing them from the perspective of correlation and logic reveals logical and logical connections between viewpoints and influencing factors. In the dispute surrounding GPT, the views that support its development are temporarily ignored. The views that question, criticize, or even deny it can be analyzed from the perspectives of cognition, economic interests, and rights protection.
2.2.1 Cognitive roots of GPT disputes
Cognition is the understanding and perception of things, which is the prerequisite and foundation for understanding things. This understanding may be an understanding of the essence of things, or it may be an understanding of the surface of things that arises from prejudice or preconceptions. At present, GPT represents the highest level in the field of Natural language processing. However, due to the limitations of modern science and technology, many people do not understand it, and this leads to hesitation and panic. This attitude can be divided into two categories: technology panic and singularity panic.
Technology panic is not only caused by a lack of understanding of a certain type of technology, but also includes panic about the unknown due to a lack of understanding of the technology and its impact, especially when it directly involves one's own important or even fundamental interests, and it is difficult to predict what kind of impact one's own interests will be in the future. For example, due to continuous technological progress, GPT has performed better than humans in many fields and can assist or even replace human work, which naturally brings panic to those engaged in these industries.
Unlike technological panic, the academic community focuses more on singularity panic due to concerns that humans themselves may be replaced by artificial intelligence. Singularity is the despairing endpoint spoken by technology prophets, and it was John von Neumann, the father of computers, who first used the concept of "singularity" in the field of technology, He pointed out that "with the accelerated progress of technology and changes in human life patterns, we have seen some fundamental singularities in human history. Once these singularities are surpassed, the human affairs we are familiar with cannot continue." American mathematician computer expert Vernor Vinge further pointed out that the new super intelligence will continue to self upgrade and progress in technology at an incredible speed, This will mark the end of the human era. Public figures such as Stephen Hawking and Elon Musk expressed concern about the possibility that AI will lead to the extinction of mankind.
Alan Turing also pointed out the shortcomings of the view that "AI cannot think and is unconscious": "If you are sure whether a machine can think, the only way is to become that machine and feel the thinking activities of this machine." It is also based on this concept that he proposed the famous "The Turing test Test" standard, Artificial intelligence imitates humans in typing conversations. If a machine causes an average of over 30% misjudgment from each participant, meaning that human interlocutors cannot effectively distinguish between artificial intelligence and human intelligence, then the machine passes the test and is considered to have human intelligence. Although Turing's refutation is suspected of "the battle between Hao and Liang", his idea of understanding intelligence and consciousness in a weak sense and observing the operation results of artificial intelligence from the perspective of function also has its rationality - human beings have never had wings like birds, but they can still make aircraft according to the principle of air Pneumatics like birds, moving from one space to another at an extremely fast or even faster speed. Douglas Hofstadter's judgment also responded to Turing's view: "What we know is our own consciousness is just the physical activity of the human brain... The key is the mode of organization, not the nature of the components." Therefore, some scholars believe that Causal inference ability has brought AI closer to people, Therefore, the issue of subjectivity no longer constitutes a principled obstacle to our understanding of whether a singularity will arrive. Also based on the functional perspective, although AI is not composed of protoplasm and cannot be metabolized, and the computer system is not a continuous system like the brain nervous system, this does not prevent AI from replacing the human brain in a functional sense. For example, the protoplasmic horse can drag the train on the track at a faster speed, and the non protoplasmic Steam engine can completely replace the work of horse pulling the train and do it better. The evolution of artificial intelligence has never relied on protoplasmic properties, but on iterative upgrades of algorithms. Therefore, it is believed that although it is an empirical question whether the singularity will actually come, it is logically possible from the perspective of Causal inference.
2.2.2 Economic considerations for GPT disputes
In a sense, there are always implicit or explicit subjective and objective interests behind any viewpoint or demand. The interests behind the theoretical dispute over GPT are at least influenced by factors such as public welfare differences and market competition share.
Open AI was initially established as a non-profit organization, but as its market potential and value were gradually explored, it shifted towards a profit model, which raised doubts about it and GPT. Open AI is a non-profit organization founded by Sam Altman, Elon Musk, Amazon Web Services (AWS In 2021 and 2023, three additional investments of billions of dollars were made to Open AI, followed by Sequoia Capital, Bedrock Capital and Kosra Venture Capital, making Open AI raise 11 billion dollars through six rounds of financing. This reflects the forward-looking and non pursuit of short-term profits attitude of companies such as Open AI and Microsoft, which is directly related to their abundance of funds. However, there are also views that the transformation of Open AI from a non-profit institution to a for-profit institution deviates from its founding spirit. The company's' primary fiduciary responsibility is to humanity ', with the goal of' creating value for everyone rather than shareholders', and turning to a profit model has led it to abandon its original lofty mission statement.
A recent public letter on the risks of generative AI research also has market and interest considerations behind it. Industry insiders in the United States have issued an open letter with a thousand signatures, advocating that all artificial intelligence laboratories should immediately suspend training for at least six months. The letter wrote: "We should develop a strong AI system only when we are sure that its effect is positive and the risks are controllable." Although this is on the surface a prudent attitude due to the uncertainty of AI risks, and the open letter was signed by more than 1000 industry executives and experts, including Mask and Apple co-founder Steve Wozniak, It seems to indicate that they are considering this matter more from a technical security perspective. But for various viewpoints, social movements, and people's actions, we should consider who benefits them the most and who benefits the most. Musk once invested in Open AI, but withdrew midway through it. As GPT rapidly iterated and upgraded with increasingly powerful features, he no longer held relevant shares. At this point, it will take time for him to reinvest in related industries. Except for those who are worried about their work being replaced by generative AI, industry professionals who call for a suspension are basically not leaders in the field of generative AI.
From the perspective of Political economy, the call for suspending or even prohibiting large-scale AI research is similar to the digital Enclosure. Driven by capital, platform enterprises start the Enclosure, effectively access other services through the free infrastructure they provide, and constantly extend the scope of the platform system, forcing non platform enterprises to use the platform to continue operating, thus ultimately making users never leave the closed ecosystem. From GPT to GPT-4, as the Technology roadmap has been clarified, it is not unlikely that similar products will reappear after a large amount of human, material and financial resources have been invested. Google, Baidu and other companies have successively released artificial intelligence competitive products similar to ChatGPT, but the high training costs and top scientific and technological talents naturally form a competitive barrier, so only those technology giants have the ability to enter the market. These technology giants combine generative AI with their existing businesses, constantly strengthen their existing business advantages, extend this advantage to the related markets under the effect of anchoring effect, constantly expand their Sphere of influence, and seek a place in the market competition of generative AI through this industrial strategy of Enclosure.
In short, the attitude of questioning, denying and even advocating suspension of generative AI should not be understood only from the perspective of technological development and security, but also from the perspective of Political economy to observe the interest demands and behavioral motivations behind the attitude.
2.2.3 Consideration of Rights Protection in GPT Disputes
The training dataset of GPT carries the risk of infringing on the rights of others. The iterative upgrade of artificial intelligence requires a large amount of data "feeding". The reason why Open AI can iterate generative AI from GPT to GPT-4 within a few years is directly related to the amount of data input. When the data size exceeds a specific threshold, it triggers the transformation of algorithm models from quantitative to qualitative. ChatGPT is a typical example of the application of pre trained basic models. Pre training technology trains general models by using a large number of techniques and tasks, which can be easily fine-tuned in different downstream applications. As the data size increases, the more parameters the pre trained model has, the higher the accuracy of the algorithm output. Open AI continues to expand the parameter size of pre trained models, with only 117 million parameters in the GPT training dataset and 1.5 billion parameters in GPT-2. GPT-3 has a larger model size and parameter count, including multiple versions, with the largest version containing 175 billion parameters and GPT-4 reaching 100 trillion. GPT uses massive text datasets from the Internet for training, mainly from WebText datasets composed of diverse text collections such as news articles, websites, and online forums. These datasets include approximately 8 million publicly available web pages collected from the Internet, as well as other datasets including books, articles, Wikipedia, etc., to enhance the diversity of the training dataset. However, Open AI has not publicly disclosed the exact sources of ChatGPT and GPT-4 training datasets, and the answers provided by GPT lack credibility verification, making it impossible to know whether the content provided is true or false, or fabricated out of thin air, all of which may constitute improper infringement of the legitimate rights and interests of others.
The generated content of GPT carries the risk of copyright infringement. On the one hand, the iterative upgrade of GPT requires capturing massive datasets from the internet for training, and then generating code, text, music, and images by identifying and copying relationships and patterns in the data. These data are themselves created by humans and in some way protected by copyright, but AI companies often use crawler technology to collect datasets on the internet and input them into training models. Is there a risk of infringing on the copyright of others when using these datasets? AI believes that the use of these datasets (at least in the United States) is constrained by the principle of fair use, which encourages the use of copyrighted works to promote free expression. However, many lawyers and analysts believe that this will inevitably infringe upon copyright and may face serious legal challenges in the near future, especially when rights holders have no way of knowing that their works have been crawled and used. At present, some foreign news media accuse Open AI of using its dataset to train GPT without its permission and without paying any fees. News groups such as New York Post, Wall Street Journal are preparing to claim compensation from Open AI, Microsoft and Google through legal means. On the other hand, GPT is a generative language model that can generate new content by learning existing corpus, text rules, and grammar constructs. Its essence is to recombine existing knowledge within a certain abstract limit. If the content generated by GPT is similar to existing copyrighted materials, the generated content may infringe on the copyright of existing works. Even deductive works that are adapted, organized, or translated from existing works of others, without the authorization of the existing copyright owner, may still constitute infringement of the copyright of others. GPT is a Natural language processing technology based on machine learning, without human creativity, and its generated content is mainly based on existing corpus data and statistical models, which often leads to similar input and output, and will cause uncertainty risk to others' existing copyright.
GPT may pose challenges to privacy and personal information rights. GPT is based on a large language model and requires massive data resources to run and improve. Open AI provides approximately 300 billion datasets collected from the internet for this model. Firstly, these training datasets face issues with source legitimacy. Open AI's unauthorized use of personal data clearly constitutes an infringement of the privacy of the parties involved, especially when sensitive data can be used to identify the parties and their family members. Even publicly available data is used beyond the context and reasonable expectations of the initial disclosure. Moreover, Open AI has not paid for the data it retrieves from the internet, and neither individuals nor network owners have received compensation. The pre trained model adopted by GPT does not require manual intervention or supervision, which enables GPT to automatically crawl data resources on the Internet, potentially obtaining data information from illegal sources, infringing on others' intellectual property, privacy, and even constituting the crime of illegally obtaining computer information system data. Secondly, the aggregation of these training datasets may reveal users' private or sensitive information. Open AI aggregates and recombines massive amounts of training data, which may seem insignificant when viewed separately. However, after combination and analysis, it may reveal sensitive facts or privacy of individuals. Modern data analysis techniques can infer extensive information about individuals, especially those that have already been disclosed, from the data footprint. Individuals often cannot make meaningful judgments about the costs and benefits of disclosure, This puts the public's privacy and personal information rights at constant risk. Of course, these training datasets come from a wide range of sources and sometimes contain sensitive data themselves. Finally, these training datasets face data security issues. Data security is a bottom line requirement of generative AI, requiring companies such as Open AI to take technical and management measures to ensure data security. However, in practice, the frequent occurrence of personal information leakage incidents not only threatens users' privacy and personal information rights, but also erodes the public's trust in related technologies. On March 20, 2023, an error in the Redis py open source library on the Redis client led to a serious Data breach event in ChatGPT, which enabled some ChatGPT users to see not only other users' chat content, but also other users' names, billing addresses, credit card types, expiration dates, etc. Just because of the Data breach incident and the concern about the legal basis of using personal data to train Chatbot, the Italian data regulator temporarily banned ChatGPT in the country.
In short, the development of generative AI relies on the massive training datasets crawled from the internet. These datasets may not only infringe on the existing copyright of others, but also bring risks to their privacy and personal information rights. These risks of infringement of rights have led many experts and scholars to question and deny artificial intelligence.
3 Legal 3.0 with emphasis on technical solutions and its integrated governance
As a high-tech achievement representing the development of internet technology and digital technology, GPT not only improves productivity and promotes social development, but also brings a series of social risks, legal risks, and potential legal issues. Even if legislation is not immediately adopted to respond, it is still necessary to engage in proactive thinking and research. Technological issues not only need to be understood and resolved within the legal framework, but also require a certain understanding and grasp of technological issues in order to better respond. Due to the characteristics of the knowledge structure of legal professionals and the characteristics of the legal discipline itself, it is easy to focus on normative analysis in the field of GPT and similar scientific and technological issues when governance is carried out. However, there is a lack of grasp and focus on how to carry out technical governance. At this point, the legal 3.0 theory that focuses on technical solutions can open a door for us in terms of governance ideas.
3.1 Legal Focus on Technical Solutions 3.0
The lag of law and the rapid development of reality create tension between the two, which is particularly evident in contemporary society with rapid technological updates and iterations. In addition to constantly formulating new laws to alleviate this tension, another approach is to attempt to address the problems brought about by technological development by providing a broader understanding of the law and incorporating science, technology, and government regulations into the legal framework. The law has a lag and often lags behind the development of reality, especially the reality of rapid technological advancements. Regarding this, John Perry Barlow once jokingly said that compared to the real world "changing at a dizzying pace", "the law is adjusted in a constantly improving manner, and its rhythm is second only to geological changes in terms of solemnity. A common solution is to alleviate this tension by continuously enacting new laws. However, even if we have confidence in the effectiveness of legal provisions and legal doctrines, existing legal provisions may "address" the legal issues brought about by new technologies under certain circumstances, but they may not necessarily "solve" these problems, making it even more difficult to call them "effective" or "satisfactory" solutions to these problems. The governance of privacy protection, personal information protection, and related issues brought about by the development of the Internet and GPT is an example. Therefore, a more flexible solution is to have a broader understanding of the law, bring science and technology and government regulation into the scope of the law, and incorporate emerging things and phenomena including new technologies into the existing legal provisions through Statutory interpretation, so as to solve new problems brought about by new things in a constant and changeable manner. Therefore, in addition to responding to new things and problems through national legislation in terms of rule improvement, we also need to try to continuously upgrade our understanding of the law and achieve more effective responses at lower legislative costs.
Brownsword's Law 3.0 theory is an attempt to develop legal theory to dynamically respond to social development. In Brownsword's theoretical framework, the understanding of legal concepts and meanings has been continuously upgraded, from Law 1.0 to Law 3.0, corresponding to the technical solutions of common law, written law, and alternative legal rules in the UK field. From the perspective of functional comparison, Law 1.0 and Law 2.0 in the continental legal system and Chinese context correspond to Legal Doctrine and Legal Policy, respectively. The operating scenarios are courts and broad legislative bodies (including government agencies with the power to formulate regulations and rules), while Law 3.0 also refers to technical solutions that replace legal rules. Law 3.0 focuses on both rule modification and technical solutions, and is a two-pronged approach and approach. Laws and regulations need to be updated or modified with the adjustment of legal purposes or policies, and the institutions and resources supporting these rules need to be maintained and upgraded to ensure that the rules align with legislative purposes on paper and in practice. In addition to rules, possible technical solutions should also be sought to supplement or replace rules, making the technology itself a part of the solution. These measures can be integrated into the design of products or workflows, and even into wearable devices and even human beings themselves.
It should be emphasized that Law 3.0 is not only a broad field of legal interest, but also a special mode of dialogue and thinking. The reason why it is a field is because it is not the terminator of Law 2.0 or Law 1.0, but an upgraded version that coexists and is compatible with them; The reason why it is said to be a special dialogue and thinking mode is that, on the basis of this upgraded coexistence, Law 3.0 constantly "talks" with Law 2.0 and Law 1.0. Through dialogue, the "new version" dances with shackles in the limited space of the "old version", and solves new problems brought about by social development and technological progress in a more refined and targeted way.
3.2 Integration of Technology and Norms in Law 3.0
The nature of the theoretical dialogue and upgraded version of Law 3.0 is essentially a fusion of laws, government regulations, and technology. From the perspective of legal application, coherence means dynamically and smoothly maintaining internal consistency. It includes the integration between beliefs based on the subjective perception and introspection originating from the objective world, as well as the integration between different beliefs and beliefs of the same subject. Coherence often appears in the study of jurisprudence and its closely related Hermeneutics, and becomes one of the core features of jurisprudence. In the academic concept of legal doctrine, "legal concepts, principles, systems, and the legal norms formed by them" is a structured network. In this sense, legal doctrine can also be referred to as "legal network structure". In the academic purport of Hermeneutics of law, coherence is not only reflected in the subject's grasp of the existing characteristics of the object itself, but also regarded as the attribute that the subject of legal reasoning and interpretation actively endows with legal materials when understanding them.
The theory of Law 3.0 goes beyond the integration of legal doctrine. In legal doctrine and judicial theory, coherence has three levels: compliance with norms and precedents, integration within the system, and integration outside the legal system. Law 3.0 goes beyond this understanding of coherence. This is mainly reflected in: firstly, in terms of the regulatory system, the norms in the Law 3.0 framework include national laws formulated by legislative bodies, relevant rules formulated by regulatory agencies during the law enforcement process, and a series of technical solutions and specifications to solve real-world problems. Secondly, the governance subjects involved in Law 3.0 theory not only include the judicial organs pointed out by legal doctrine and judicial theory, but also include the administrative regulatory subjects in modern administrative law that emphasize the administrative process as the center, the private subjects that regulate technology, and all relevant subjects that use technology for technological governance. Thirdly, compared with jurisprudence and judicial theory, which are concerned about precedents, rules and principles and the procedures in their operation, the theory of Law 3.0 emphasizes result orientation, and pays more attention to understanding and evaluating practice, technology and norms from the results. Taking the personal flying of UAVs as an example, if the UAVs near the airport cannot take off without affecting flight safety through RF interference, this technical solution can replace the rule design of setting a No-fly zone. At this point, technology and rule operation have the same or even better effect. Finally, the Law 3.0 theory emphasizes observing and analyzing the integration of technology and norms (including laws and government regulations) from a functional perspective. In the sense of functional substitution, the idea that architecture and design can be used to protect people and property is as ancient as pyramids and locks. Under the premise of complying with legal concepts and purposes, institutions, resources, and technological solutions that support the operation of rules can coexist, engage in dialogue with each other, and even replace each other. It is worth emphasizing that the architecture here is not only the process and product of planning, designing and building buildings and physical structures in the sense of architecture, but also a set of rules and methods describing functions, organizations and computer system implementation in current Computer engineering and Internet technology, that is, the structure and behavior design of software systems is realized through program code.
Law 3.0 transcends the judicial oriented Hermeneutics of law and the modern administrative law in the sense of government regulation, and becomes a "new" law that accommodates and emphasizes technical solutions on the basis of both. It is also in this sense that Law 3.0 is a theory that emphasizes dialogue and integration: it emphasizes the dialogue between judicial oriented Hermeneutics of law, administrative process centered regulatory law and technical solutions, and focuses on the integration of national laws, administrative regulations and technical solutions. This integration is a normative integration that maintains the inherent consistency and unity of legal norms at different levels formulated by different entities. It is also a comprehensive integration that integrates technical solutions into norms, thereby integrating technical solutions with norms at different levels and the principles and values that guide these norms. If in a modern administrative country, highly complicated and technical regulatory matters and various risks in the modern society make it difficult for the original legislature to effectively respond to the legislation, so the administrative organ has to make rules to fill the gaps in the law, fill the blueprint defined by the legislature, and use its professional administrative and technical bureaucrats to deal with highly professional problems, then Law 3.0 is absorbing the achievements of system development On the basis of adhering to the integration of values, professional upgrades have been carried out to efficiently address the problems brought about by social and technological development with more professional technical solutions and lower governance costs.
The change of rules, governance ideas and plans is the embodiment of Social change. If legal doctrine, with legal interpretation as its core, is the governance of relatively simple societies during the pre industrial and early stages of industrialization, with problem-solving primarily focused on the downstream judicial process, then legal policy, with administrative processes and corresponding administrative regulations as its dominant approach, is the governance of complex societies during the industrial period. Due to the increasing complexity of social life and the increasing uncertainty of social risks, it is necessary for professional institutions and professionals to do professional things based on specialized legal systems, and solve problems in the middle and upper reaches by focusing on administrative processes and regulations. Law 3.0, while respecting the aforementioned two versions of laws and regulatory measures, emphasizes technical solutions to address and solve problems, which is a new approach to addressing more complex and technologically challenging issues in the industrialized and information society. It should be emphasized again that Law 3.0 does not require technical solutions to replace administrative regulations and legislation, but rather to address and solve problems with low-cost technical solutions as much as possible through integrated interaction with the previous two versions while respecting legislation and administrative regulations.
4 Framework for integrated governance of GPT
Integrated governance integrates technical solutions, government regulations, and national legislation, combining technology with institutions, experts, and the public, providing a relatively comprehensive and systematic approach and framework for the governance of high-tech technology. When regulating a certain technology, experts proficient in that type of technology may propose efficient and targeted governance strategies, or they may "cluster and be foolish", or they may be captured by technology and the capital behind it, and formulate public policies in a "professional" manner that favor a certain or certain industrial institution. Integrative governance aims to integrate regulators who gather experts with legislators who represent public opinion, while incorporating technical solutions into the governance toolbox. Efforts are made to address technical issues through technology while also keeping technology in the cage of the system, achieving a "two handed approach" between technology and the system.
4.1 Constrain "magic" with "magic"
In the world of the internet, there is a saying that "defeats magic with magic", which means treating someone in their own way or giving them back in their own way. However, when faced with artificial intelligence represented by GPT, it is difficult to use the phrase 'defeat'. Perhaps, from the day artificial intelligence was born, it, like other technologies, has formed a symbiotic relationship with humans. Just like the internet technology we have seen, humans and artificial intelligence grow together. While creating artificial intelligence, humans are also being shaped by it. However, in the face of increasingly powerful artificial intelligence, humans can find various tools from their own governance toolbox to constrain it.
From the current achievements in technological development, the technological architecture that constrains GPT includes at least blockchain and artificial intelligence. In addition to using artificial intelligence technology to detect and constrain it, blockchain has become another reliable option. Blockchain helps improve the security of generative artificial intelligence. If we take measures to hand over important personal information, important decision-making basis, and key data generated artificial intelligence energy systems to blockchain smart contract management, it will ensure the safety of generated artificial intelligence for humans. This not only alleviates the public's security anxiety about generative artificial intelligence, but also recognizes that artificial intelligence is actually safer. More importantly, if the rapid iterative development of GPT itself represents the efficiency of technological updates, then its efficient processing of natural language and substitution for existing work reflect its potential and ability to improve the overall efficiency of society. Accompanying and even prioritizing efficiency is fairness. In the perspective of law, although there are different understandings of fairness, there is a basic consensus to consider it as the primary value. If high-tech represented by GPT represents high efficiency, then blockchain reflects the fair orientation of technology. Regarding this, some netizens have made wonderful comments. After two hundred years of the Industrial Revolution, the God of Technology has once again increased its leverage on the balance of "efficiency" and "fairness". While unleashing the spirit of strong artificial intelligence in the bottle, it has also handed over the spell book to control this spirit to humans, which is blockchain.
A narrow blockchain is a chained data structure that combines data blocks in a chronological order in a sequential manner, and guarantees a distributed ledger that cannot be tampered with or forged in a Cryptography way. Generalized blockchain technology is a new distributed infrastructure and computing paradigm that uses block chain data structure to verify and store data, uses distributed node consensus algorithm to generate and update data, uses Cryptography to ensure the security of data transmission and access, and uses smart contracts composed of automated script code to program and operate data. Although blockchain's unique models of tokens, mining, and coin speculation have led many people to question blockchain, believing that blockchain is just a gimmick created by capital and a tool for speculation and speculation, its unique functions and value make it a technology that cannot be ignored. An important product brought about by blockchain technology is token. As a convenient and credible accounting tool, it can be combined with smart contracts to measure work contribution, reward work effectiveness, and represent personal identity and ownership, both online and offline. From the perspective of Political economy, the value of commodities originates from the undifferentiated human labor condensed in them. In the current society, human labor methods are becoming increasingly diverse and diversified. In daily life, people's browsing of web pages, using small programs, posting videos, displaying social circles, and even replying to others' comments constitute a form of labor. Even though these labor have relatively low value, they still bring positive effects to the improvement of network services and the quality of network products. These labor behaviors can be recorded in the form of tokens. Through tokens, the benefits brought by the development of artificial intelligence technology can be more equitably enjoyed by the public and users, while providing new ways for those affected by the development of artificial intelligence to find jobs and receive compensation. For example, some users provide graphics cards and thus provide computing power, some users "feed" artificial intelligence data to improve and improve its performance, and some users write code according to the tasks published by the platform, which may develop true shared artificial intelligence. Computing power, model construction, and data feeding are all completed by users through a decentralized framework, and tokens are obtained based on their contributions, Token can be exchanged for corresponding AI usage permissions and corresponding services. Therefore, blockchain provides a new type of labor plan, while also providing the possibility for those who provide labor to enjoy corresponding services or receive corresponding rewards, making the achievements of artificial intelligence technology more equitably enjoyed by more people.
4.2 Injecting ideas into "magic"
When an organization, technology, or other subject advocates for self-restraint, people often doubt the effectiveness of such constraints. This suspicion is supported by experience and logic: historically, even if the "emperor of the system" can self regulate in the short term, it is difficult to institutionalize and standardize in the long term, and ultimately, it is also difficult to escape the consequences of exceeding power and collapse. Logically speaking, any power has an impulse to expand until the cost and benefits of power expansion reach a marginal balance. As a result, people naturally doubt the motivation, motivation, and execution of artificial intelligence technology and the self-restraint of institutions that master these technologies.
For this question, we can understand and respond from both internal and external perspectives. From an internal perspective, an institution that looks forward to long-term benefits has the motivation to regulate and restrain itself. Regulation is the act of restricting the activities of individuals who constitute a specific society and economic entities that constitute the economy. The subject of restriction can be administrative agencies or private entities that set behavioral standards or patterns for themselves. The results of regulation not only contribute to the realization of public interests, but also contribute to the long-term interests of private entities. For example, the environmental supervisor system of state-owned enterprises in China, the voluntary certification system of food enterprises, and the industrial self-discipline specifications of the Internet Society. In this sense, self-regulation is a "subject responsibility", that is, the obligation of the subject to take proactive actions and inactions to do their job well. From an external perspective, many self-regulation not only involves self-restraint, self-interest, and the pursuit of social welfare, but also external factors such as supervision from society and the state. It is a form of self-regulation under the gaze of the law, the state, and society. External 'gaze' is an important factor that motivates and enforces internal self-restraint. Self regulation is actually a form of self-discipline influenced by public and social power, so we don't have to worry too much about the effectiveness of self regulation, as it is self-discipline under heteronomy. It is precisely in this sense that self-regulation is the understanding of human existence from the perspective of the "interactive relationship" and "interdependence relationship" between individuals when individual rationality pursues the maximization of their overall interests, thus also practicing public rationality.
Building 'ethical' artificial intelligence. When the GPT was asked about racial discrimination, its answers were very Political correctness, which was due to the injection of ethical norms and the setting of prohibitive rules. Overall, ChatGPT has not experienced any systematic ethical violations. Scholars have evaluated the "Axiloma Principles of Artificial Intelligence", the Montreal Declaration released by the Forum on the Social Responsibility Development of Artificial Intelligence, the second edition of "Ethical Consistent Design: Prioritizing Human Welfare through Autonomous and Intelligent Systems" published in 2017, and the European Commission's "Artificial Intelligence, Robotics, and Autonomous Systems" After 47 principles in six documents including "Five General Principles of Artificial Intelligence Code" and "Partnership with Artificial Intelligence" proposed in the report "Artificial Intelligence in the UK: Ready, Willing and Capable?" released by the Artificial Intelligence Committee of the House of Lords, the consistency and overlap between these six sets of principles are impressive, and these principles are often used in bioethics as benevolence, non malice The four core principles of autonomy and justice are interlinked. Correspondingly, when conducting research on the Large Language Model (LLM), the academic community has systematically identified six risk areas: discrimination, exclusion, and toxicity, information hazards, misinformation hazards, malicious use, human-machine interaction hazards, automation, access, and environmental hazards, and studied the violations of basic human ethics in these areas. The results show that in the latest language model system, there is no indication that these hazards will occur. In addition, on April 8, 2019, the European Union formulated and released the "Trusted AI Ethics Guidelines", which regards human dignity as the core value of artificial intelligence development. After examining the main regulatory ideas of the 2018 Toronto Declaration, some scholars consider algorithms as a special form of legislation. These phenomena seem to indicate that algorithms are embedded with relatively stable ethics when designing and learning ChatGPT. Of course, this does not mean that its ethical requirements and regulation are at ease. In fact, when someone asked ChatGPT to write a poem praising Trump, it refused on the grounds of technological neutrality, and when asked to write a poem praising Biden, it quickly gave a crude but highly affirmative hymn. Similarly, when people asked it how the United States should deal with China's Weather balloon when it flew into the United States, and how China should deal with the Weather balloon when it flew into China, it gave diametrically opposite answers. This means that even though existing artificial intelligence or GPTs are subject to basic ethical constraints, they are still influenced by ideology and international relations. Therefore, we should always pay attention to the ethical constraints of artificial intelligence and inject ethical concepts into the application of magic.
4.3 Using "Market + Rules" to Assist the Development of "Magic"
Integrative governance means using technology, regulation, and legislation as governance tools to solve real-world problems. From a functional perspective, whether it is legislation by the legislative body, regulation by administrative organs, or precedents or guiding cases by judicial organs, all have the attribute of rules, which can guide, evaluate, predict, educate, and enforce the behavior of an unspecified majority. In a sense, the fundamental role of the market in resource allocation has almost all other functions except for legal enforcement. Therefore, in the gaps between technology, regulation, and legislation, there exists a market position and room for its effectiveness.
From the perspective of resource allocation and economic development, the market naturally has a fair attribute. Of course, when market entities develop to a certain extent, there will also be phenomena of distorting the market through monopolies, unfair allocation of resources, and unreasonable distribution of vulnerable individuals. In terms of the development of artificial intelligence, as it is in the stage of development and has immeasurable potential for economic development and social progress, when the risks and adverse effects it brings are not yet sufficient and obvious, it is not urgent to restrict it through laws, but to fully develop it through the market and fully unleash its potential for energy upgrading. When the Steam engine train replaces the carriage, we should not restrict the train industry for the sake of the work of the coachman; When the popularity of computers makes it possible for everyone to type, we should not ban the development of the computer industry for the sake of Audio typist. On the contrary, the development of the train and computer industries has not only caused some people to lose their jobs, but also created more positions that represent the direction of productivity development and higher pay. It is in the process of industrial upgrading that productivity continues to improve, society continues to develop, and people themselves continue to be liberated. Similarly, when artificial intelligence begins to burst into vitality and have an impact on certain industries and people's psychology, we can cautiously approach and pay attention to the various positive and negative impacts it brings, but do not rush to limit it. In this sense, the attitude towards generative AI can refer to the "illegal rise" in the early stages of the development of the internet industry. At that time, the Internet, as a new mode of production, continuously connected various types of productive resources online and offline, mobilized and matched within society, creatively generating new methods for effectively utilizing resources, and triggering sustained changes in production methods.
The development of the market and industry will react to the development of artificial intelligence technology. Watching the changes and seeing the bidding process unfold, and maintaining a tolerant attitude towards it before discovering "obvious and immediate dangers", may be the best strategy. From the current development status of artificial intelligence, it can be seen that this industry is a heavy capital industry that requires a large amount of manpower, material resources, and financial resources to be artificially "fed" in order to optimize its functions. However, there are not many enterprises and countries that can invest such resources. Therefore, the number of regulated entities is limited. When potential threats and dangers arise that require regulation, the breadth of regulation will reduce the difficulty of regulation. At the same time, artificial intelligence products such as ChatGPT and ERNIE Bot can also reduce the technical risks of other links in their industrial chain through self-regulation and regulation at a lower cost within the framework required by the existing legal rule system. Although the performance of artificial intelligence continues to improve and performs well in certain office work, academic support work, and even audio and video production work, there is no need to worry too much about this. For example, some paintings and audio produced by artificial intelligence have high quality, which is the result of repeated training using high-quality works. When there is no updated or more creative content in terms of information feedback or data collection, it may only result in low-level or high-level repetition. Therefore, products produced by artificial intelligence do not pose a threat to truly meaningful innovation for humanity.
In addition to using the market and its underlying logic and rules to guide artificial intelligence, legal rules cannot be absent. In fact, there is no shortage of existing legal rules in China. For example, the existing legislation on copyright, privacy, personal information rights, network security, and national security basically covers the legal fields involved in the development of the artificial intelligence industry. Although there are certain so-called "gaps", such as whether personal information has a rightful status and how data property rights should be determined, this is more due to the lack of social consensus on industrial development and rights attributes, and the current law is not suitable or urgent to intervene. It is a deliberate "blank" that needs to be confirmed in law when the time is ripe and consensus is formed. Overall, China's existing regulatory framework, including national legislation, government regulations, and technical solutions, is in a "basically sufficient" stage.
Therefore, it seems that we should not be overly anxious and anxious about regulating artificial intelligence. Instead, we should focus on the potential risks of the development of the artificial intelligence industry while promoting industrial development and technological progress, and not force systematic solutions to problems once and for all - in fact, this is also impossible, Instead, we adopt a step-by-step and gradual upgrading approach: for risks that have already been discovered or have strong practicality, we first try to determine whether they can be solved through technical solutions, because for certain rights, code can provide dual protection with real laws and social norms, and even be more effective than law; For risks or behaviors where technical solutions are difficult to fully work, try to solve them through standards, policies, government specific and abstract administrative actions. This can not only reduce legislative costs, but also mobilize expert teams from professional bureaucrats to provide targeted solutions to problems; For the rules and solutions that are still difficult to solve and have basically formed a social consensus, we will consider adopting specialized legislation through the legislative body.
Epilogue
When ChatGPT emerged, the industry and academia excitedly declared that it was another 'iPhone moment', viewing it as a great invention just as great as the iPhone, and believing that it was a great revolution in solving the interaction between humans and artificial intelligence, allowing people to use artificial intelligence as easily and smoothly as they could with an iPhone. In the more than ten years since the existence of the iPhone, we have also seen the positive impact of various fields such as smartphone and mobile internet theory, applications, and industry models on society and the world. It not only provides high-quality communication methods, but also provides an efficient and high-quality mobile internet platform, convenient and safe life.
In fact, when the iPhone first emerged, not many people believed that it would bring significant risks to our lives, and there was an urgent need to regulate it. Similarly, generative artificial intelligence represented by GPT is in the ascendant, and the legal issues involved in its production and operation can be effectively addressed within the existing legal framework, such as privacy, copyright, personal information, etc. There are some areas of legal ambiguity, such as the nature of information ownership and the attributes of data property rights, which have not yet formed a social consensus. Regardless of whether GPT is generated or not, this ambiguity still exists. Therefore, in terms of the legal issues involved in GPT, we should not and do not need to regulate it through legislation until a social consensus is formed. For the practical, urgent, and legal issues it brings, we can try to implement integrated governance, which prioritizes the use of "magic" to constrain "magic" in technological solutions, injects value concepts into "magic" through internal and external means to build "virtuous" artificial intelligence, creates a good market for generative artificial intelligence to compete fairly and generate high-quality algorithm models, When necessary, guide and constrain the development of generative artificial intelligence through government regulations and national legislation.