Location : Home > Resource > Paper > Theoretical Deduction
Resource
ZHI Zhenfeng | Information Content Governance of Large Model of Generative Artificial Intelligence
2024-04-26 [author] ZHI Zhenfeng preview:

[author]ZHI Zhenfeng

[content]

Information Content Governance of Large Model of Generative Artificial Intelligence



ZHI Zhenfeng
Professor, School of Law, University of Chinese Academy of Social Sciences

Researcher, Institute of Law, Chinese Academy of Social Sciences



Abstracts: Based on great computing power and powerful algorithms, generative artificial intelligence large language models can process massive amounts of data in natural language processing, computer vision, speech processing, and other fields. They can provide services such as content creation, digital humans, conversational search, and code generation. These models have great potential in fields such as autonomous driving, financial risk control, healthcare, and the Internet of Things. As a major change in internet information technology, large models greatly enhance logical reasoning ability, and ‘understanding ability’ of humans. They have become powerful tools for producing creative information content and may greatly change the ecology of internet information content. However, they also pose information risks such as the spread of low-quality information, pollution of original information sources, and impact on social ethics. It is necessary to balance development and security and explore compatible governance methods to address these risks.


Improving the computer's understanding of the integration and intention of human knowledge, expanding the boundaries of human intelligence, and achieving smoother human-computer interaction have always been important directions of information technology efforts. With the explosion of ChatGPT (ChatGPT) launched by OpenAI, a US artificial intelligence research company, many tech giants continue to increase the generative AI race. Google released a multimodal "second-generation channel language model" (PaLM 2) that can "understand" and generate audio and video content after the chatbot Bard, Microsoft's new Bing search engine integrated the multimodal "Generative Pretrained Transformer 4" (hereinafter referred to as GPT-4), and Amazon also announced that it had joined the battle by releasing Titan. Baidu's "Wenxin Yiyan", Huawei's "Pangu", Tencent's "Mixed Yuan Assistant", Alibaba's "Tongyi Qianwen", SenseTime's "Ririxin", Kunlun Wanwei's "Tiangong", iFLYTEK's "Xinghuo Cognition" and other large models from China are constantly emerging. Various types of generative AI large language models (LLMs) have exploded, and the technology application boom has swept the world.

Based on large computing power, the large artificial intelligence model that uses strong algorithms to process massive big data, trains on large-scale unlabeled data, and learns certain features or rules to predict future results, the number of parameters has increased from hundreds of millions to hundreds of billions, realizing the leap from a single task supporting a single modality of pictures, images, text, and speech to a variety of tasks supporting multiple modalities, thus becoming a model library with generalization capabilities and certain general capabilities. The large model has performed well in the fields of natural language processing, computer vision, and speech processing, and has been able to provide services such as content creative generation, digital human, dialogue search, and code generation, and is also very promising in the fields of autonomous driving, financial risk control, health care, and the Internet of Things.

The large model already has the ability to serve "thousands of industries". However, as a major change in Internet information technology, the logical reasoning ability of large models and the "comprehension ability" of human beings have been greatly improved, bringing revolutionary changes in the generation of information content such as text, images, voices, and videos, which will truly bring the production and dissemination of information content into a new era of Artificial Intelligence Generated Content (hereinafter referred to as AIGC), and is triggering a knowledge revolution in human society. By learning the characteristics of objects from huge data, no longer simply comparing and matching, but trying to understand people's thoughts, AI content creation that uses existing text, image or audio files and generates content based on large data sets will not only become a powerful tool for human beings to produce creative information content, but also may greatly change the online information content ecology and bring new risks and challenges to information content governance.

Since the high degree of technical professionalism of generative AI models in the production and dissemination of information content is far from people's existing common sense, the first part of this paper will mainly sort out the typical functions of AI models in content generation as concisely as possible. On this basis, the second part of the paper will demonstrate that due to the significant impact of large models and their insurmountable limitations, it may bring significant risks to information content governance, and then try to put forward a governance path for generative AI information content under the condition of briefly reviewing information content governance at home and abroad.


1. Generative AI opens a new era of information content production and dissemination


Language has a special meaning for human beings. Heidegger proposed that "language is the home of being", and Wittgenstein put it bluntly, "The boundaries of my language mean the boundaries of my world". In the development of artificial intelligence technology, natural language processing has always been known as the "crown jewel of artificial intelligence". How to enable computers to understand and process human language is an important key point in human-computer interaction. The natural language processing framework used by generative AI large language models has made significant advances in human-machine dialogue and content generation capabilities, which can learn and train on large text datasets to produce complex and intelligent writing, even into images or videos.


1.1 Revolutionary changes in the way information content is produced and disseminated

A history of mankind is a history of information production, communication and dissemination. From the word of mouth in primitive societies, the simple paper and silk in agricultural societies, to radio and television in the industrial age, and then to the development of the Internet, especially mobile communication technology, the production and dissemination of human information content is mainly carried out in two modes: user-generated content (UGC) and professionally generated content (PGC). Before the Internet era, whether it was simple silk, books, newspapers and periodicals, or radio and television, the most easily disseminated and long-term circulation was mainly professionally generated content, and the producers of information content were mainly intellectuals, officials, and professionals in certain fields. In the era of mass media, content producers and gatekeepers such as reporters and editors have also emerged. Overall, professionally generated content is authoritative, reliable, and of good quality. Relatively speaking, the content of word of mouth and street talk is mainly produced by users, and the producers may not be professionals, and there are generally no quality gatekeepers. The so-called "gossip" is mostly "self-produced and self-sold", and its rise and fall are also sudden, and its death is also fast. However, in the Internet era, especially after the widespread application of social media technology, everyone has a microphone and a camera, and the "street talk" in cyberspace can also be widely disseminated and recorded for a long time, and short videos give everyone the opportunity to "be seen". In cyberspace, the volume of user-generated content is naturally overwhelming. On the WeChat platform alone, there are hundreds of millions of audio and video calls and tens of billions of messages sent every day. By the end of 2022, the number of online video (including short video) users in China reached 1.031 billion, and the number of online live broadcast users reached 751 million. The production and dissemination of information content in human society has realized a revolutionary transformation from professional production to user production.

The emergence of generative AI large language models has opened a new era of AI-generated content, which is another revolutionary change in the way human information content is produced and disseminated. The main body of information content production has undergone huge changes, and artificial intelligence can replace human resources in the whole process of information collection, screening, integration, and reasoning, which greatly liberates human resources. Subversive changes have taken place in the production efficiency of information content, and large computing power drives strong algorithms to process big data, in natural language processing such as text classification, sentiment analysis, machine translation, question and answer systems, text generation, etc., computer vision such as image classification, object detection, image segmentation, face recognition, and image generation, autonomous driving such as vehicle control, road recognition, and traffic flow prediction, financial risk control such as fraud identification, risk assessment, and prediction of market changes, medical and health care such as disease diagnosis, pathological analysis, and medical image analysis, as well as smart home, On a variety of tasks in various fields of the Internet of Things, such as intelligent manufacturing and environmental monitoring, high-quality results can be judged and content generation can be carried out efficiently. The dissemination of information content has undergone subversive changes, and the production and dissemination of information have become more convenient, especially the threshold for obtaining professional knowledge has been lowered. The expression form of information content is more abundant, and the use of artificial intelligence creation technology makes the conversion of pictures, texts, and codes more free, and can generate a "digital human" clone with one click, "opening the era of intelligent interconnection".


1.2 The content generation function of the large model

Large models already have multi-modal and cross-modal information content production capabilities. According to the large models released at home and abroad, in terms of information content generation, natural language processing is mainly used as the core architecture, transformers are used as general modules/interfaces, and deep learning models with self-attention mechanisms are used to generate content similar to text or images created by humans. GPT-4 enables the model to acquire the ability to natively support multimodal tasks by pre-training on a multimodal corpus with a variety of data, including text data, arbitrarily interleaved images, and text.

Based on human feedback, large language models such as ChatGPT can learn and improve the output content based on the information input of the user, and can also achieve "alignment" between the expression and intrinsic value of the AI model and the common sense and values of human beings. ChatGPT is also able to use Instruction Tuning technology to better adapt to the user's language habits and communication style, understand the user's questions, and improve the system's adaptability and performance to specific tasks and scenarios.

In terms of the output form of information content, generative AI large models can realize multiple modalities such as text, image, video, audio, digital human and 3D content. Taking SenseTime's "RiRixin" large-scale model series as an example, "Miaohua SenseMirage" is a Wensheng diagram creation platform, which can generate pictures with real light and shadow, rich details, and changeable styles with text, and can support the generation of 6K high-definition images. SenseChat is an efficient chat assistant that solves complex problems in seconds, provides customized suggestions, and assists in the creation of first-class texts, with the ability to continuously learn and evolve. "MingMou" is a data annotation platform with more than 10 built-in general large models and industry-specific large models, which support intelligent annotation of 2D classification, detection and 3D detection in various scenarios such as intelligent driving, smart transportation, and smart city. "Ronin SenseAvatar" is an artificial intelligence digital human video generation platform, which can generate a digital avatar with natural voice and movement, accurate mouth shape, and multilingual proficiency in just a 5-minute live video material. The scene generation platform "Qiongyu" and the object generation platform "Gewu" are 3D content generation platforms, which can generate large-scale 3D scenes and refined objects efficiently and at low cost, opening up a new imagination space for the application of the metaverse and the integration of virtual and reality.

Generative AI models are ushering in the Model as a Service (MaaS) era. Technology giants create general models and provide models to B-end customers in subdivided fields (Tencent Mixed Yuan also targets G-end customers), and customers polish the models, so as to empower all walks of life. At the same time, open the public beta or paid use interface for C-end users to attract players with a deep understanding of the industry, polish and train the model.

Generative AI interacts deeply with users to generate massive amounts of information, providing convenient conditions for users to search for information, consume products, and participate in public life. ChatGPT-3 has 175 billion parameters, trained on about 500 billion pieces of text collected on the web, massive data and powerful computing power to create this very intelligent publicly available artificial intelligence. GPT-4 has a significant improvement in performance compared to other generative large language models, which is an important step towards artificial general intelligence (AGI). Its versatile abilities can be used in a wide range of scenarios such as abstraction, comprehension, vision, coding, mathematics, medicine, law, and understanding of human motivation and emotion, and its performance in task completion in some areas is at or beyond the human level.


1.3 Application scenarios of large AI models

Generative AI can be a human chat companion, and it can generate fluent, contextual, and common sense chats through pre-trained technical support models, and the conversations present a certain "personality" rather than rigid machine speech, so it has the potential to become a virtual companion robot. In a specific field, by learning professional knowledge and using "fine-tuning" techniques, large models can take on the "work" of intelligent customer service. In search services, large models will be better able to grasp human intent and directly generate the "answers" that users want, rather than just providing a series of web links.

The most typical application of large models is writing generation. Depending on the topic and keyword requirements, generative AI can "write" stories, novels, poems, letters, news reports, current affairs reviews, essay outlines, etc., and perform text modifications and polishes, such as grammar correction, text translation, and keyword extraction. Large models can also write code, and according to OpenAI's technical R&D personnel, through the training of large language models, it can generate functionally correct code bodies from natural language document strings. A user once used ChatGPT to write an apology letter for the company involved in the 2023 Shanghai Auto Show "ice cream incident", but the result was faster and more appropriately worded than the company's public relations copy, and the company involved was sharply commented by netizens that its "public relations level is not as good as ChatGPT". GPT-4 can also recognize content based on images, and even understand images with specific connotations behind them, which is the so-called "meme". SenseTime's "Second Painting" in the "RiRixin" large-scale model series, as well as Stable Diffusion and Midjourney, can all use text prompts to generate very creative images.

Generative AI models have begun to show the ability of "experts" in various fields to engage in a certain degree of professional knowledge Q&A and analysis such as basic medical consultation, legal services, and educational Q&A in several fields. For example, SenseTime can "teach" patent law, and GPT-4 can answer exam questions, analyze charts, and can already guide and encourage users to think and get answers step by step like a real human teacher. ChatGPT can help legal workers brainstorm ideas, improve case analysis and document writing, organize citations, and more.

Generative AI models can be clever personal assistants. In life, ChatGPT can help order restaurants, book movie tickets, make travel plans, get weather forecasts, etc., and can also recommend relevant news, music, movies, books, etc. according to the user's interests, and can also customize travel routes, schedules, reminders and other services for users according to their hobbies, working hours and places and other information. For example, after connecting to Alibaba's large model Tongyi Qianwen, the application "DingTalk" can fully assist the office, and can create poetry novels, write emails, generate marketing plans, etc., and generate meeting minutes and to-do lists at any time in DingTalk meetings.

Generative AI also has a lot to offer in the fields of product design, deep synthesis, and manufacturing. In many scenarios such as logo design, clothing design, Internet content illustration, e-commerce illustration, etc., the creative content generation functions such as Wensheng pictures and Tusheng texts can be used. The large model can also generate marketing planning plans based on descriptive text, generate application applets according to functional sketches, and open up application closed loops in multiple fields and industries. In addition, it is also more powerful to generate digital human clones with one click by deep synthesis functions such as connecting smart life and AI face swapping according to text descriptions. Using it for 3D printing, industrial products can be manufactured directly.


1.4 Characteristics of AI information content generation

In addition to the Internet information content production modes such as professionally generated content (PGC), user-generated content (UGC) and hybrid generation, the influence of artificial intelligence generated content (AIGC) mode is becoming more and more significant, which brings about the evolution of content production subjects and production methods, the improvement of content interaction and distribution methods, and the improvement of content production quality and generation effect. AI-generated content has some extremely significant revolutionary features.

The acquisition of information content enables the transformation from presentation to generation. The artificial intelligence model can well summarize the existing knowledge of human beings, streamline and efficiently output according to massive data, and greatly improve the ability of human beings to produce and obtain information. It can write or draft texts, replacing part of the human effort. It has changed the way knowledge is generated and transmitted, greatly lowering the threshold for acquiring professional knowledge, so that the generation of professional knowledge no longer requires decades of professional training from human beings. Compared with the AI tools used in media organizations in the past, the application of this generation of generative AI is open to all users, bringing the possibility of self-release and self-creation, bringing a kind of information inclusiveness, thereby narrowing the knowledge gap in society.

There has been a shift from decentralized to integrated information content provision. Before the artificial intelligence model, the information that people obtained on the Internet mainly came from various scattered web pages, knowledge communities, online encyclopedias, etc. However, generative artificial intelligence completes the integration of massive public knowledge by integrating information and analyzing data, and can interact with people, so as to integrate various functions such as search engines, encyclopedia knowledge platforms, knowledge communities, software open source communities and some social media, and carry out streamlined and efficient inductive output according to the massive knowledge it inherits, which greatly improves the ability of human beings to obtain information. To a certain extent, the large model integrates the search, search, integration and preliminary output of information, which is conducive to promoting the transmission, dissemination and inheritance of knowledge.

The service scenario has realized the transformation from a single domain to a universal one. Generative AI large language models have better versatility, precision, and efficiency, and can be pre-trained or otherwise learned on large datasets, and then fine-tuned to efficiently handle complex tasks such as computer vision and natural language processing. Large language models are trained using a large corpus covering a variety of subject areas, which can mimic human intelligence in a wider range of applications. In the implementation of "model-as-a-service", as a "foundational model of the code layer", generative AI large language models have the ability to become a new generation of infrastructure, which can be applied to various downstream scenarios from search engines, content platforms to application software, including daily work, scientific research and education, and even public services, affecting all walks of life. As a result, the developers of the cornerstone model have become the "gatekeepers" of the digital technology market, with strong market power. This is a truly epoch-making product in the history of AI development: if AlphaGo marks the beginning of narrow AI reaching and surpassing human capabilities in specialized fields, ChatGPT ushers in the era of general AI – that is, an era in which AI has a wide range of learning capabilities and meets or exceeds the capabilities of ordinary humans in most fields.

The dialogue mode has realized the transformation from one-way retrieval to intelligent interaction. How to make computers no longer cold machines, how to enhance the understanding of computers to human beings, and how to make it more convenient for human beings to obtain information are all important driving forces for the development of information technology. Before generative AI large language models, humans acquired knowledge and information either through face-to-face communication, or by consulting library materials, or by internet search engines. It's one-way and boring in the way of getting information. In addition to the communication between people, there is a cold "subject-object" relationship between people, books, materials, and computer networks. But generative AI large language models have dramatically changed the way humans have conversations when they access knowledge and information. Taking ChatGPT as an example, through a generative pre-trained model with massive data, it is trained based on a large number of Internet texts, which can understand and answer questions on various topics, and can carry out natural language expression in a human-like rather than machine-like discourse system. ChatGPT-3 already has the remarkable ability of contextual learning, which can predict contextual vocabulary, learn or imitate patterns in data, and output responses in the corresponding context through corresponding key information matching and pattern imitation. With the increase in the number of model parameters and the continuous enhancement of contextual learning capabilities, the continuity of human-machine dialogue can be ensured, and users can be actively asked questions when they cannot understand instructions. This provides a layer of "personified" communication for human beings to obtain information through large models, so that computer information retrieval is no longer a cold machine operation, but may be an intelligent interaction with a "human touch".


2. Generative AI brings new challenges to information content governance


In a sense, generative AI models are becoming an aggregate of human information content production and dissemination. Information content carriers such as books, newspapers, radio, and television, information providing tools such as news media, search engines, knowledge communities, online encyclopedias, and open source communities, and specific professional identities such as customer service, writers, doctors, teachers, and experts are all integrated into the generative AI model. The large model has become a textbook, a source of knowledge, a "famous teacher" and an "authoritative person", which can "monopolize knowledge", "influence judgment" and "shape cognition" from the source. Large language models have the potential to penetrate into all fields of human production and life, but the limitations and abuse of their technology will bring severe challenges to information content governance.


2.1 Technical limitations

Training data has flaws and limitations. It is impossible to verify the accuracy of all the astronomical-level data required for the pre-training of large models, and if the data is inaccurate or missing, it will inevitably affect the reliability of the results, resulting in "garbage in, garbage out". If the data is biased and contains sensitive information, it may also lead to discrimination and misperception of the generated results. In 2017, a study demonstrated biases and stereotypes in natural language processing data by analyzing the Stanford Natural Language Reasoning (SNLI) corpus. Without access to the internet or the use of plug-ins, the knowledge of large models is often time-bound, for example, GPT3.5 has knowledge limited to events that occurred before 2021, and Google's Bard says that it can search for information online, but there is still a certain time lag. They suffer from limited computing power, insufficient training, and high R&D and operating costs. Large model training can be called the aesthetics of violence, which requires large computing power, big data, and large models, and each training task costs a lot of money. According to the minutes released by SenseTime, on the cloud computing power side, at least 10,000 A100 chips are needed to run ChatGPT, while at present, only SenseTime, Baidu, Tencent, Byte, Ali and High-Flyer have more than 10,000 reserves in China, with a huge computing power gap and extremely high cost.

There is a cap on content generation. High-probability combinations may not be real, and it is difficult to be creative. AI models like ChatGPT can only react based on trained information and can't really access real-time facts or understand context like humans. First, AI content generation is still actually a reorganization of knowledge, rather than knowledge production or reproduction. On the one hand, there is still a gap with human intelligence, the ability to understand the context is still limited, and there is a lack of "human touch", so we can only pursue short-term and large-scale content, but cannot produce meaningful and innovative content. The answers output by the model are generated by its pre-trained neural network, and the parameters in the neural network are randomly initialized, and the training process is optimized for stochastic gradient descent based on the input data, which makes the model likely to give different or even opposite answers to the same question. The answers given are sometimes "conclusive", sometimes "serious nonsense", and "improvised" or "dead and unadmitted" when questioned, essentially due to the fact that the output is randomly selected from multiple alternative answers, probabilistic, and unpredictable. On the other hand, the quality of the output content is highly dependent on the user's ability to ask a question. There is a contradiction between generalization and specialization in the process of natural language processing, and it is difficult to ensure that the results are easy to read without reducing their professionalism. Second, there is the common problem of "hallucination", which makes content "look right and be inherently wrong."

Large-scale model AI has the appearance of personification, but it is still impossible to truly have personality. In digital systems, artificial intelligence does not care about human nature, and it is inevitable that there will be "hallucinations" such as "confident reactions". Thirdly, cross-linguistic and cross-cultural problems, multilingual corpus collection may not be able to grasp the connotation behind the corpus. In the GPT-3 training dataset released by OpenAI, the English corpus is as high as 92.65%, while the second place is French, which accounts for only 1.92%. The corpus input largely determines the resulting output. In the training of large models, too little use of Chinese corpus will not only greatly affect the quality of the content generated by large models, but also greatly affect the Chinese civilization with Chinese language as the main ideographic tool.

Content moderation has complexities that are difficult to control. Due to the inherent algorithmic black box and interpretability flaws, it is difficult to understand the reasoning behind the model's predictions. ChatGPT also writes on its website that the sheer volume of content generated by these models makes manual review and moderation of generated content very difficult. According to OpenAI's paper, GPT-4, while also having these technical limitations, is ostensibly "more convincing and credible than earlier GPT models." This will create a bigger problem. When users rely too much on it, they are likely to be unaware of or ignore mistakes in use.


2.2 Risks of the application of generative large language models

Due to the massive amount of training data required by generative AI large models, as well as their generational, priority, integration, and versatility characteristics, while empowering thousands of industries, they will also generate various huge risks.

2.2.1 Risk of personal information leakage

The process by which a user converses with a generative AI large language model is the process by which personal information is widely collected. When a user asks a question, it may expose personal information that they don't want to make public. However, according to OpenAI's instructions, users can only delete personal accounts, not sensitive personal information. On March 20, a vulnerability occurred in ChatGPT's open-source library, allowing some users to see other users' conversations, names, email addresses, and even payment information. OpenAI had to remind on its official website: "Please don't share any sensitive information in conversations." "In fact, when generative AI is asked to answer a question or perform a task, the information inadvertently provided by the user may be used in the process of training, learning, and improving the model, and thus placed in the public domain. This may not only violate the user's personal privacy, but also reveal other people's information. For example, when a lawyer uses it to review a draft divorce settlement, it can reveal the personal information of the parties to the case. In particular, the large model demonstrates strong inference capabilities, and it is able to write programs according to user needs, which on the one hand will improve the user's product experience, and on the other hand, it may also bring the risk of personal information leakage.

2.2.2 Risk of trade secret leakage

There have been reports of three cases of Samsung Semiconductor leaking trade secrets due to the use of ChatGPT: an employee asked it to check the source code of a sensitive database for errors, an employee used it to optimize the code, and another employee entered the recorded meeting into ChatGPT and asked it to generate meeting minutes. Whether it is a market entity, an academic institution or a government agency, when using a large model, it is inevitable to share certain information with it, which poses a huge risk of leaking trade secrets or even state secret information.

2.2.3 Data Security Risks

The data used for training may be inaccurate or tendentious, and the quality of the data is not guaranteed, or even legitimacy is not guaranteed, resulting in the generated content may be "toxic". As more and more industries and fields access AI-generated large language models, data leakage and compliance risks are becoming increasingly prominent, and once data is leaked, it will bring huge economic and reputational losses to enterprises and industries. Even if it is fragmentary or fragmentary information, ChatGPT may combine it with other data for mining and analysis, so as to infer intelligence information related to national security, public safety, and the legitimate rights and interests of individuals and organizations. Especially for models with overseas servers such as ChatGPT and Bard, if sensitive data is entered during use, it may cause security problems in cross-border data flow, which will bring data security and even national security threats.

2.2.4 Cyber security risks

With a lower threshold for expertise and models making it difficult to identify the user's purpose, generative AI has the potential to provide a convenient tool for cybercrime. By writing cyberattack code, it is capable of generating code in multiple languages such as Python, JavaScript, and more, can create malware to detect sensitive user data, and can also hack into a target's entire computer system or email account to obtain important information. Some experts have detailed how to use ChatGPT to create polymorphic malware, bypass the content policy filters established by OpenAI, and create malicious code. Criminals only need to write marketing emails, shopping notifications, or software updates in English in their native language requirements model to create online scam scripts, and there are few signs of spelling and grammatical errors that make it difficult to identify scams or phishing emails. In addition, the information used by the large model in the process of training account information may be shared with service providers and related companies, which may lead to the risk of data leakage and leave vulnerabilities for cybersecurity attacks in the process.

2.2.5 Algorithmic Risk

Generative AI is essentially the use of algorithms to process massive amounts of data, and algorithms are the key to it. However, because the algorithm itself cannot verify the training data, it is often possible to generate misleading content that seems accurate but is inherently wrong, creating "hallucinations". The accuracy of the content generated by the model is limited, and the model itself cannot identify the authenticity of the written content, which can easily lead to the generation and dissemination of false information. Moreover, algorithms themselves cannot avoid social biases and value tendencies. Algorithms with their own problems may be guided to generate content that violates laws and regulations. Value judgments in data use and algorithm training may also produce "toxic" content, solidifying social biases and discrimination, not only based on race, but also based on gender, beliefs, political positions, social status, etc.


2.3 New challenges in information content governance

Generative AI large language models have the potential to replace the whole process of human thinking in information collection, knowledge acquisition, content evaluation, and thinking and reasoning. In particular, large models have advantages in the fields of natural language processing and computer vision, and when generating graphic content and conducting human-computer dialogue, it may generate huge information content risks due to its reduced information production cost, lower professional knowledge threshold, more aggregated application functions, and wider use fields.

2.3.1 Generative causes a flood of inferior information

Generative AI can write or draft text, replacing some of the human labor, the production cost will be negligible, and the amount of text will skyrocket. The huge growth of content will not only put pressure on the available physical memory recording space, bringing about an explosion of information, but more importantly, it will cause the rapid expansion and mass dissemination of harmful or undesirable content.

First, disinformation worsens the online ecosystem. Generative AI may fabricate false information, output low-quality information that is mixed with real information and false information, and use fluent sentences to explain the fabricated false facts, which is serious nonsense, which is somewhat confusing for groups with limited sources of information. "Automation bias" inclifies users to trust answers from seemingly neutral model outputs. If the powerful content creation capabilities of generative AI are used to generate false information about individuals and enterprises, it will lead to rumors, slander, insults, and slander, especially the use of deep synthesis technology to generate text, pictures or videos of speeches impersonating political figures or key figures, which may also cause social unrest and produce more harmful consequences.

Second, misleading information interferes with personal decision-making and daily life. Generative AI is increasingly taking on the role of "intellectual authority", producing erroneous or misleading content in various professional consulting services such as business planning, legal services, and medical and health care, which will directly affect users' daily lives. When used for transaction planning, due to the existence of "illusions", limited accuracy, and limited contextual understanding, it is easy to "nonsense" and plan the wrong itinerary and schedule. When applied to professional consultations such as medical and health care and legal services, once the knowledge is answered incorrectly, the information generated may mislead users and interfere with their medical consultation or legal litigation activities.

2.3.2 Initial source pollution

Traditional sources of knowledge, such as textbooks and news media, are increasingly being replaced by online platforms. As a large model that integrates the functions of knowledge platform, search platform and generation platform, it may not only become a monopoly source of knowledge, but also may produce source pollution from the source. Information content is created without human supervision, so the ability to produce disinformation on a large scale becomes easier and faster. In the proliferation of "echo chambers" and "filter bubbles" on the Internet, the generation of a large amount of unsubstantiated one-sided content will create a false sense of majority opinion and exacerbate the polarization of opinions.

First, it misleads the view of history. History is objective, but the perception of history can be subjective. Especially in the international community, distortion of history abounds due to ideological conflicts and biases of values. In recent years, differences have constantly arisen in Western society over the understanding of World War II; over the issue of the War of Resistance Against Japan, China, South Korea, and other Asian countries have often criticized Japan for beautifying its war of aggression and distorting history. As human creations, it is difficult for large models to avoid the biases that humans have. In fact, it is not uncommon for false or misleading information to amplify political bias and manipulate user perception in the answer to political questions, and once combined with bot accounts in cyberspace, it may bring greater security risks. Many tests have found that Western models often reflect Western positions and values, and even distort history and facts when it comes to China-related issues.

The second is ideological and value bias. Large language models can have a variety of social biases and worldviews that may not represent the user's intentions or widely shared values. Different countries, political forces, and interest groups all have quite different ideologies and values, and present a realistic power structure, which is reflected in all kinds of information. The datasets required for large model training often encode the ideologies and values of the real society, which may lead to the consequences of reinforcement. Research has shown that most of the data in the Western large model training set is primarily generated from the perspectives of whites, males, Westerners, and English speakers, so the data may be heavily skewed to reflect these structures. The power structure of the real society is encoded in the large model, and the output of the large model reflects the content of the real power structure, which produces the Matthew effect of power, and the result is often to create an oppressive reproduction system and destroy the information ecosystem. In particular, in areas involving ideology and values such as religion and human rights, in areas where national interests are in conflict, and even on extreme issues such as race and the superiority or inferiority of civilization, monopolizing large models is equivalent to monopolizing textbooks, encyclopedias, and libraries. The large model will become a powerful tool for cognitive domain operations, shaping public perception, and manipulating international public opinion.

The third is the challenge of language hegemony. The scale effect of the digital age makes small languages face great challenges. Language is the home of existence, the carrier of culture, and the presentation of civilization. Although generative AI can provide multilingual and cross-language services, large model training requires a huge corpus, even large models such as Wenxin Yiyan in China, which are also trained by code based on the English environment, which may not only have value bias, but also fierce competition between different languages and the civilizations they represent. If it cannot master the integrated and monopoly platform such as the large model, a nation may not even be able to keep its language in the end, and even move towards bubbles and gradually dissolve.

2.3.3 Generic ethical risks

In an atomistic individualistic society, generative AI is increasingly becoming a chat companion and a close "friend" of people, bringing with it a series of ethical challenges.

One is that human beings may have greater confusion and misconceptions about what a "human" is. Due to the excessive competition and involution of the real society, and the influence of more and more atomic individualistic values, individuals in modern society are becoming more and more lonely and alienated from each other. Generative AI models can support chatbots, companion robot services, and even become the "companions" of many lonely individuals, but they can also exacerbate the alienation of interpersonal relationships and the isolation of personal lives. Technology helps humans, but it can make them even happier.

The second is to limit the ability of individuals to make decisions and weaken the dominant position of people. Generative AI presents a trend of de-bodily, de-real, de-openness and de-privacy, which hides the risk of more complete deprivation of human subjectivity by algorithms, and its essence is a manifestation of the alienation of human-machine domestication. Human-machine communication will crowd out the space for interpersonal communication, thereby weakening the social and psychological relevance of the embodied subject: social relations no longer require the physical "presence", and "public life" disappears. In other words, humans have created algorithms, but algorithms have the potential to discipline and reformat humans, subtly changing human behavior and values, and thus eroding human subjectivity. People may entrust the final decision to some automatic text generator, just as they would ask Google existentialism questions today.

The third is to hinder content innovation and knowledge progress. When large language models are applied to writing generation, problems such as manuscript washing, plagiarism, and academic misconduct may occur. Some universities abroad have begun to ban the use of ChatGPT on campus to avoid students cheating on exams or essay writing. Some well-known international journals have also explicitly not accepted AI as a collaborator. Large models can be good tutors, but they can also be used as cheating artifacts. In particular, for minors, over-reliance on generative AI can limit the growth of their minds, thereby jeopardizing sound personality, schooling, and academic training. Since large models simplify access to answers or information, information that is effortlessly generated can negatively impact students' critical thinking and problem-solving skills, amplify laziness and neutralize learners' interest in conducting their own investigations and coming to their own conclusions or solutions.

Fourth, it encourages false propaganda and manipulation of public opinion. In the era of self-media development, the manipulation of public opinion has become a more serious problem. In the 2008 presidential election dispute in Iran, the American social media platform Twitter became an important tool for the opposition. By using social media, the opposition has dramatically reduced the cost of mobilization, which in turn has increased its mobilization capacity. The U.S. government made it clear in its report on funding Iranian dissidents that year that it funded "new media", and even directly asked Twitter officials to postpone system maintenance so that the opposition would not lose its contact channels. The false information originating from Twitter has also been amplified by traditional media such as CNN and BBC. But being smart is mistaken by being smart, and public opinion manipulators often suffer the consequences. In the wake of the Cambridge analysis, some American scholars have predicted that large-scale generative AI models represented by ChatGPT will become a powerful tool for targeting candidates and influencing public opinion in the next round of elections.


3. The status quo of information content governance of generative artificial intelligence large language models


Artificial intelligence brings great possibilities, but it also raises great concerns. Humanity must take precautions against possible risks getting out of control, and universal legislation on the safety and ethics of AI research and development is urgent. Due to the high degree of certainty of the regulatory object, the relevant legislation in the field of specialized artificial intelligence is becoming more and more mature. For example, the regulations for different fields such as autonomous driving, smart healthcare, algorithm push, artificial intelligence investment advisory, and facial recognition can be found in different levels of laws in various countries and regions. How to maximize the effectiveness of generative AI technology while reducing the negative impact of emerging technologies on social development has become an important global issue.


3.1 Large-scale model regulation has become an important issue in Europe and the United States

There are already many people in technology and industry who are wary of generative AI. They believe that AI systems may pose a profound risk to human society, and that advanced AI may represent a profound change in the history of life on Earth, and should be planned and managed with corresponding care and resources. Now, AI labs have fallen into a runaway race, and no one can understand, predict, or control large models, so it is necessary to press the development pause button, greatly accelerate AI governance, and regulate AI R&D. At one point, Italian data protection authorities issued a ban on ChatGPT and investigated it for alleged violations of European privacy regulations. However, due to the fact that the generative AI model is still a small lotus, countries around the world have not formed a systematic regulatory policy and regulatory system.

The EU intends to make adjustments in the legislative process by establishing a dedicated working group to promote cooperation and exchange information on possible enforcement actions taken by data protection authorities. Privacy regulators in some EU countries have said they will monitor the risk of ChatGPT's personal data breach under the EU's General Data Protection Regulation (GDPR). The European Consumer Organisation (BEUC) has launched a call for European regulators at EU and national level to open an investigation into ChatGPT. The European Union is adjusting its AI bill to regulate general AI such as generative large language models, considering requiring OpenAI to submit to an external audit of system performance, predictability, and security settings for explainability. Under the regulatory framework envisaged in the EU AI Act, generative large language models will be classified as high-risk and heavily regulated due to their potential to create harmful and misleading content.

The U.S. government has also begun to take action. On March 30, 2023, the U.S. Federal Trade Commission (FTC) received a complaint from the Center for Artificial Intelligence and Digital Policy (CAIDP), a nonprofit research organization, arguing that GPT-4 does not meet any of the FTC's requirements for "transparency, explainability, fairness, and empirically sound while promoting accountability" for the use of AI, and is "biased, deceptive, and a risk to privacy and public safety" , requiring an investigation into OpenAI and its product GPT-4 to determine whether they have complied with guidelines issued by U.S. federal agencies. On May 4, the Biden administration announced that it would further promote responsible innovation in the United States in the field of artificial intelligence by conducting a public evaluation of existing generative AI systems. In accordance with the Principles of Responsible Disclosure of AI, a group of leading AI developers, such as Google and Microsoft, are required to conduct public assessments on specific AI system evaluation platforms to provide researchers and the public with key information that affects the model, and assess compliance with the principles and practices in the AI Bill of Rights Blueprint and AI Risk Management Framework, so as to facilitate AI developers to take timely steps to solve problems. In January 2021, the U.S. Congress passed the National Artificial Intelligence Initiative Act (NAIIA), which aims to promote U.S. competitiveness in the field of artificial intelligence.

Generative AI, as the forefront of technological competition, has in fact become the patent of a few countries. In most countries, it is still difficult to make a difference in technology development, industrial deployment and regulatory governance. Moreover, the current AI regulation in foreign countries is still mainly focused on the field of traditional AI, rather than generative AI large language models. However, due to the objective concerns about generative large language models in society, there is a voice in the EU that requires generative AI large models to comply with high-risk obligations, which could have a significant adverse impact on the competitive environment for local governments, industries, and corporate goals.


3.2 The current status of domestic regulations

China has initially formed a three-dimensional and all-round normative system for the governance of online information content, which is composed of laws, administrative regulations, judicial interpretations, departmental rules, and a series of normative documents. The information content governance of generative large language models already has a basic legal framework, and has the framework institutional constraints to enable it to develop without harming national security, public interests, and individual rights and interests.

In terms of information content regulation, the information content security regulatory framework composed of laws and regulations such as the Criminal Law, the Civil Code, the National Security Law, the Counter-Terrorism Law, the Public Security Administration Punishment Law, the Cybersecurity Law, the Personal Information Protection Law, and the Measures for the Administration of Internet Information Services explicitly prohibits harmful information that endangers national security, social stability, and false information. The Provisions on the Governance of the Online Information Content Ecosystem also include vulgar information and negative information, which have always been in the gray area, into the legislative system, highlighting the diversity of governance subjects and targets. Norms such as the "Provisions on the Administration of Online Audio and Video Information Services" and the "Provisions on the Administration of Internet Post Comment Services" further build a regulatory mechanism for information content covered by the whole platform, and provide a basis for the content regulation of generative large language models.

In terms of risk response to AI algorithms, the Provisions on the Administration of Algorithmic Recommendations regulate algorithmic recommendation services, opening the process of legalization of algorithmic governance. The Provisions on the Administration of Deep Synthesis of Internet Information Services regulates technologies that use generative synthesis algorithms such as deep learning to produce text, images, and other online information, and regulates technologies for generating or editing text content such as text generation, text style conversion, and Q&A dialogues, providing basic rules for the application of generative large language models.

The Measures for the Administration of Generative AI Services (Draft for Comments), which ended on May 10, 2023, puts forward a series of regulatory assumptions for generative AI services from the whole process of data use, personal information collection, content generation, and content prompt labeling. But the balance between security and development is not easy to wrestle. Regulation first, although it reflects the sensitivity of the regulatory authorities, but the impact on industrial development should also be carefully measured. The new generation of information technology represented by generative artificial intelligence is an important commanding heights in the current field of international competition, because China is in the initial stage of this technical field, the industrial foundation is not strong enough, and the experience accumulation of application impact is not sufficient. For example, it is necessary to distinguish in detail whether the service provider should be liable for product infringement or other liability for the damage that may be caused by generative AI. We should adhere to the principle of inclusiveness and prudence, and leave enough space for technological and industrial innovation on the premise of ensuring national and social security.


4. Explore the way of information content governance of generative AI


Cybersecurity is relative, not absolute, and "zero risk" is not a scientific goal. In the process of model development, it is objectively difficult for developers to foresee all potential risks, and they need to explore and practice in a relatively relaxed environment and within reasonable limits. The risks brought about by technological progress can only be constrained, not completely avoided. For example, problems such as poor accuracy caused by the "illusion" of large models and difficulties in accountability caused by algorithmic black boxes can only be controlled as much as possible, but cannot be completely eliminated.


4.1 Incentive compatibility: Optimize the legal environment for the development of large models

A new round of technological revolution and industrial revolution is vigorously unfolding, and each industrial revolution will have a major impact on the rise and fall of a country, the rise and fall of a nation, and the prosperity and decline of civilization. In the increasingly fierce game between China and the United States, and the United States has suppressed and blocked my country to the limit, whether there is a generative artificial intelligence model, and whether our model is advanced and powerful enough, is a more fundamental question. Strict regulation should be based on advanced technology and strong industry.

Under the influence of the traditional planned economic system and due to the severe international environment at a certain stage, although the state has always adhered to the socialist market economic system and advocated the combination of a promising government and an effective market, some localities and departments are still accustomed to being deeply involved in the market in specific industrial supervision. Especially in the field of the Internet, due to the extreme importance of network security to national security, the Internet has become the forefront and main front of the ideological struggle, resulting in the overall strictness of China's supervision of the Internet industry. In the field of information content governance, the law formulated by the Standing Committee of the National People's Congress has never been passed, and there are still only administrative regulations such as the Measures for the Administration of Internet Information Services, which have been implemented for more than 20 years.

However, a large number of experiences at home and abroad have proved that laws and policies in modern society not only have regulatory functions, but also are important aspects of international system competition. As a comprehensive framework for solving social problems, laws and regulations must be balanced wisely, rather than being as strict as possible, nor as loose as possible. If the commanding and repressive top-down one-way regulation is too rigid, it will make it difficult to enforce the law, or selective enforcement. Due to the excessive power of the regulatory authorities, they will also face a greater dilemma of regulatory capture, which will eventually inhibit technological innovation and industrial development, and miss the country's development opportunities.

In this context, in recent times, developed countries have placed more emphasis on incentive regulation. Practice has proved that if the regulatory measures and rules can be compatible with the incentives of the supervised object, it will not only be easier to achieve the regulatory goals, but also greatly reduce the cost of supervision and improve the enthusiasm for compliance and law-abiding. Therefore, adhering to the principle of the rule of law and implementing the regulatory concept and idea of compatible incentives have become an important part of optimizing the business environment under the rule of law. Because of its stable expectations and long-term benefits, the rule of law is also known as the best business environment.

In the face of the rapid development of generative AI models, the legislative and regulatory authorities must show greater humility and respect for the market, innovation and industrial autonomy, leaving a broader space for the development of new technologies and applications. Considering that computing power is the basis for the development of large models, and the computing power architecture is extremely expensive, in terms of legislation and policy choices, China should provide better policy space for the financing of new technologies and new industries. Considering that large amounts of data are needed for large model training, on the premise of protecting personal information and data security, unreasonable obstacles in data training and other aspects should be eliminated as much as possible in supervision, so as to promote the rational circulation and utilization of data elements. The law must conform to the law, and the regulation must conform to reality. It is necessary to face the risks and challenges brought by generative AI, balance innovation and public interests, ensure the beneficial application of generative AI, avoid social risks, and finally establish an enabling regulatory concept and regulatory model that integrates development and security, conforms to objective laws and development stages.


4.2 Pluralistic co-governance: build a governance mechanism for corporate social responsibility and individual active participation

Technological innovation and industrial leap are the source of the country's prosperity, and adhering to the rule of law and scientific supervision is the institutional guarantee for the country's prosperity. Since the second half of the 20th century, the distinction between "management-oriented" legislation and "governance-oriented" legislation has become increasingly clear. "The social governance model that is compatible with the era of high complexity and high uncertainty should be a cooperative action model, and only when multiple social governance subjects jointly carry out social governance activities under the willingness to cooperate, can we solve various social problems that have emerged and achieve excellent results in social governance. ”

The Internet industry is highly specialized due to its technical complexity. The history of the Internet shows that while the supporting role of the government and the state cannot be ignored, the role of the scientific and technical communities is equally important. Adhering to the spirit of open source, the exchanges between scientists and professional professionals and the consensus reached have greatly shaped the Internet protocols, standards and rules, and given a strong impetus to the development of the international Internet. In particular, as a new technology and industry, the complex code world and technological development behind the Internet are often ahead of the world of daily life, and it is impossible for the public to fully understand them immediately, including the regulatory authorities. The potential for development is not always clear. Without enough patience and tolerance, without a gentle, rational mindset, it's easy to stifle vital innovation through fear of risk. In the field of new Internet technologies and applications, the pursuit of absolute security often leads to greater insecurity. In this context, developed Internet countries, including China, often pursue the concept of pluralistic governance and social co-governance, which not only mobilizes enterprises and society to fully participate, but also reserves a broad space for the development of new technologies and applications.

As a new trend in the development of Internet information technology, generative artificial intelligence models have shown explosive and revolutionary potential, and by empowering thousands of industries as a productivity tool, it is likely to bring great benefits to future technological innovation, industrial leap, social governance, and personal well-being, and even become an important factor in the country's comprehensive competitiveness. In this case, it is necessary to first support and support the development and deployment of large models, strengthen corporate social responsibility, standardize data processing and personal information protection, ensure that the development and application of AI models comply with moral and ethical standards, and promote algorithms for good. It is necessary to strengthen risk identification and data traceability, improve technical governance capabilities, clarify data sources and training processes, identify potential deviations and other risks through datasets, and monitor content output and identify risks through manual review or the establishment of monitoring systems. Establish a feedback and complaint mechanism to receive, monitor and evaluate risks that arise in real time, and take remedial measures in a timely manner.

The application and impact of generative AI models is global and requires the joint efforts of R&D institutions in various countries to harmonize technical standards. As the largest Internet country, we must also have the awareness of participating in international Internet governance and providing Internet public goods to the international community, support China's large models and platforms to participate in and organize global technology communities, and make China's contributions in technology, ethics and rules.

Of course, it is also necessary to improve citizens' digital literacy and avoid the digital divide caused by the imbalance in the application of generative large language models. First of all, it is necessary to enhance users' comprehensive understanding of the application of new technologies, and encourage the public to look at and evaluate new technologies with a scientific and rigorous attitude, and not blindly follow or oppose them. Second, to educate the public about neural networks, deep learning, and other technologies to help people understand the operating principles and limitations of generative AI, and avoid technology dependence. Finally, the ability to distinguish between true and false information should be enhanced, and the public should be guided to maintain a certain rational attitude and ability to distinguish the output of generative AI.


4.3 Rule by law: Construct a legal framework for generative large language models

Under the framework of laws and regulations such as the National Security Law, the Cybersecurity Law, the Counter-Terrorism Law, and the Measures for the Administration of Internet Information Services, all online communication platforms that engage in news and information services, have media attributes and public opinion mobilization functions are included in the scope of management, and content that endangers national security, undermines ethnic unity, and disrupts social stability is strictly prohibited. First, China has coordinated the development of network information content with the construction of a network power, which has effectively promoted the rapid development of network information technology and the great enrichment of information content. Second, the construction of network civilization should be used to coordinate the construction of online information content, create an upward and benevolent online fashion, and promote the public to consciously resist the erosion of illegal and negative information. Third, the establishment of the rule of law on the Internet is used to coordinate the governance of the online information content ecosystem, effectively curbing the spread of illegal and negative information in cyberspace and optimizing the online ecology.

As a new platform for the production and dissemination of information content, the generative AI model has not yet shown a full picture, but its characteristics of generation, integration, versatility and intelligent interaction are making it the main monopoly of information production and dissemination. Therefore, in terms of legislation and supervision, it is necessary to identify their risks as accurately as possible, and improve the chain regulation from data to algorithm to content within the original information content governance framework. First, regulate the collection, storage, and use of user data, and prevent user data from being used for harmful purposes and generating false, erroneous, or misleading content. Second, improve the algorithm filing system, and guide enterprises to establish a third-party review or self-discipline mechanism for various content such as text, images, and videos generated by AI. Third, it takes into account the freedom of individuals to access knowledge and create content while identifying and regulating harmful information.

The first is to establish a scientific and clear mechanism for assuming legal responsibility. For generative AI service providers, legislation requires them to ensure the reliability and accuracy of data, fulfill content review obligations to avoid the generation of harmful information, perform special marking obligations to identify deep synthetic content in a conspicuous and conspicuous manner, and establish mechanisms to prevent, promptly identify, and stop the generation and dissemination of harmful and negative information. For users, when the service provider has assumed the responsibility for security management and fulfilled the duty of prudence, the user shall bear criminal liability for using the model as a tool for cybercrime. Other information platforms shall promptly screen false information and other harmful or negative information generated by models, and prohibit or restrict the dissemination of information on the platform. According to the nature and consequences of different acts, different types of responsibility are determined.

The second is to coordinate the rule of law at home and the rule of law related to foreign affairs. At present, the mainstream generative AI large models are mainly distributed in China and the United States, and the United States is in the leading position in large models, which has great advantages. Article 50 of the Cybersecurity Law stipulates that "technical measures and other necessary measures shall be taken to block the dissemination" of foreign countries that use generative large language models to infringe on China's interests, interfere in China's internal affairs with political manipulation and ideological bias, or transmit other information suspected of violating the law or committing crimes. In fact, for the use of generative AI large language models by foreign governments or relevant organizations to transmit information that violates China's laws and regulations, not only should we take technical measures to prevent it, but also explore the establishment of countermeasures to better safeguard national sovereignty, security, and development interests.