Return to the Library

A Reshaped World? New Features of Artificial Intelligence and National Security under the Rise of ChatGPT


Scholars from Huaqiao University explore the implications of generative AI for China’s prosperity and national security, following the launch of ChatGPT. They emphasize the pivotal role leadership in AI research and applications will play in global power distributions going forward, given implications for standards-setting ability, productivity growth, and information control.

FacebookTwitterLinkedInEmailPrintCopy Link
Original text
English text
See an error? Drop us a line at
View the translated and original text side-by-side



Science and technology (S&T) competition has increasingly become an important part of the strategic competition among powers. The technological progress of emerging countries will affect the distribution of power in international relations, and their innovative activities may have externalities on dominant countries. Therefore, when any country makes a breakthrough in S&T competition, this will help the country gain an advantage in the fierce competition among nations. As a representative of cutting-edge technologies in recent years, the impact of the development trend of artificial intelligence (AI) on the international community has gradually become apparent: Both the series of S&T wars launched by the United States against China and the technological alliances established by the United States, Japan, South Korea, and other countries to restrict China’s S&T development exhibit the characteristics of scientific and technological competition. In particular, the emergence of the sensational application ChatGPT at the end of 2022 once again highlighted the urgency of S&T competition. As a generative AI launched by OpenAI, ChatGPT features “strong interaction,” “strong understanding,” and “strong generation.” Through dialogue with ChatGPT, people can quickly retrieve information, write abstracts of papers, write computer applications, and perform other operations. Compared with the AI of the past, these breakthrough capabilities have allowed people to see the dawn of the development of artificial general intelligence. This type of AI is an AI system with universal applicability, autonomy, creativity, and learning capabilities. This is expected to become an important driving force for the progress of human society. Microsoft Chairman and CEO Satya Nadella excitedly stated that ChatGPT will be a killer application that changes productivity and the way we work in the future. From the perspective of national security governance, the application of AI in the cognitive and physical domains can certainly improve the country’s governance level in the security field. However, the rapid development of AI based on big data models as represented by ChatGPT has significantly expanded the breadth and depth of the content and fields involved in the national security field in a short period of time. The widespread use of this type of AI will profoundly change the security situation countries face.

科技竞争愈发成为大国战略竞争中的重要一环。新兴国家的技术进步会影响国际关系中的权力分配,其创新活动可能会对主导国家产生外部性影响,因而任何国家在科技竞争中的突破都将有助于该国在激烈的国家竞争中占得优势。人工智能作为近年来前沿科技的代表之一,其发展态势对国际社会的影响逐渐凸显:无论是美国对中国开展的一系列科技战,还是美日韩等国围绕限制中国科技发展而建立的技术同盟,都展现出科学技术竞赛的特征。特别是2022年年底现象级应用ChatGPT的出现,再次彰显出科技竞争的紧迫性。作为由OpenAI公司推出的生成式人工智能(Generative AI),ChatGPT具备“强交互”“强理解”“强生成”等特点,人们可以通过与ChatGPT对话进行快速检索信息、撰写论文摘要和编写计算机应用程序等多种操作。这些相较过往人工智能出现的突破性能力,让人们看到了研发通用型人工智能(Artificial General Intelligence)的曙光。该类人工智能是具有普适性、自主性、创造性和学习能力的人工智能系统,有望成为推动人类社会进步的重要驱动力。微软董事长兼首席执行官萨蒂亚·纳德拉(Satya Nadella)惊叹,ChatGPT将是未来改变生产力和工作方式的杀手级应用。从国家安全治理来看,人工智能在认知域和物理域的应用的确能提升国家在安全领域的治理水平,但以ChatGPT为代表的基于大数据模型下快速发展的人工智能,短时间内大幅扩展了其在国家安全领域涉及内容和领域的宽度和深度,对此类人工智能的广泛使用将深度改变国家面临的安全态势。

In recent years, discussions on the development of AI and international relations both inside and outside China have increased, and specific issues of concern have gradually turned to international political and national security issues, such as the relationship between AI technology and national power. In current research, scholars’ discussions of political issues around AI as represented by ChatGPT mainly include the following: The first is the relationship between the development of AI technology and changes in power among countries. As AI becomes the core force behind a new round of industrial revolution and is seen as the “technical foundation” of the innovation paradigm of the future, AI research has become a key factor for countries vying for S&T supremacy. Therefore, the current contest among technological powers in AI research will lead to increasingly fierce competition among countries. The changing composition of national power will further disrupt the originally stable structure of the balance of power. The strategic behavior of countries will change, and the gap between developed countries and developing countries will gradually widen. The second is the increasing complexity of AI and the national security situation. Some studies argue that the widespread use of AI will make the security situation that countries face more complex. The application of AI in military equipment has partially changed the format of war, and the United States and Western countries have become more willing to launch regional wars. The third is research on the impact of ChatGPT itself. At present, the focus of academic circles is on the impact of the emergence of ChatGPT on academic ethics and social development, with less attention paid to the impact of ChatGPT on the current national security situation.

近年来,国内外有关人工智能与国际关系发展的讨论逐渐增多,关注的具体议题逐步转向人工智能技术与国家权力之间关系等国际政治和国家安全方面的议题。从当前研究看,学者们围绕以ChatGPT为代表的人工智能开展政治类议题讨论主要包括以下几种:一是人工智能技术发展与国家间权力变化间的关系。随着人工智能成为新一轮产业革命的核心力量,其被认为是未来创新范式的“技术基底”,人工智能研究成为各国争夺科技竞争权的抓手之一。因此,当前技术大国在人工智能研究方面的较量会导致国家间竞争态势愈发激烈,国家权力的组成发生变化并进一步打破原本稳定的权力均势结构,国家的战略行为发生改变,发达国家与发展中国家之间的差距将会逐步放大。二是人工智能与国家安全态势复杂化。部分研究认为,对人工智能的广泛运用会使国家面临的安全态势更加复杂, 人工智能在军事装备上的应用部分改变了战争形态,美西方国家发动地区性战争的意愿更加强烈。三是研究ChatGPT本身带来的影响。目前,学界关注重点在ChatGPT的出现对学术伦理带来的冲击与对社会发展的影响上,但是较少涉及ChatGPT对当前国家安全形势的影响。

The research discussed above makes a relatively comprehensive study of the relationship between AI and the development of international relations from the perspectives of changes in power, the transformation of the international power structure, and competition between major powers. However, current research focuses on the relationship between AI and the relative dynamics of power among countries. There is relatively little research on the impact of the development of generative AI as represented by ChatGPT on national security. With the progress of the times and the advancement of technology, the intrinsic meaning of security and power are changing. In the field of non-traditional security, power is present at the traditional technical level and in the game of international systems and standards, and is also reflected at the level of discourse in the struggle for dominance. Power manifests itself in technical power, institutional power, and discursive power. Although discursive power is less obvious than the other two types of power, in a sense, it is only when a country’s comprehensive power in multiple dimensions is enhanced that this country can safeguard its rights and interests in this field. Different from any previous technological progress, ChatGPT’s discursive power is shaped by technology and manifested in human-computer interaction. This means that whoever can influence the input content and output preferences of the machine by setting its algorithms will hold discursive power. Compared with all previous strategic power games, the ideological penetration and security risks caused by this game are more subtle. The setting of algorithms and content preferences will be decisive factors in the production of discursive power. The emergence and development of generative AI as represented by ChatGPT will further increase the intensity of the discursive power game among countries and will also further enrich and rewrite the meaning of security.


Every technological revolution in history has brought immense changes to the economy and society, but they have also brought new challenges and imposed new requirements for national governance systems and governance capabilities. From the perspective of maintaining national power and national security, we must squarely face the challenges and opportunities brought by this technology. Therefore, this article aims to answer the following questions: In the context of increasingly fierce competition among countries for dominance in the AI field, how are ChatGPT and generative artificial intelligence different from the AI of the past? How will this type of AI change the features of the current state of national security? What methods should we adopt to respond to changes in real security challenges? The answers to the above questions will help us squarely face the challenges that the new technologies are bringing to national security and provide a reference for maintaining the national security order and taking hold of the development of the security situation.


I. The development of AI technology and the rise of ChatGPT


ChatGPT and related technologies are the latest data points plotting out the rapid development of AI technology. To understand the impact of ChatGPT on current societal development and international security, we must first understand the operating mechanisms of ChatGPT and how it changes and impacts the combined application of AI technology.


(i) The development of AI technology and the features of ChatGPT’s capabilities


Research on AI began in 1956 when scientists such as John McCarthy and Marvin Lee Minsky discussed how to use machines to simulate human intelligence. The concept of “artificial intelligence” proposed in these discussions opened the door to research in the academic discipline of AI. After several technological upgrades, breakthroughs were continuously achieved in the technological development and application of AI. The nature of the technology AI is based on has gradually shifted from the “logical reasoning” of the 1.0 era to the “knowledge engineering” of the 2.0 era and then to the “machine learning” of the 3.0 era. Along the way, AI capabilities improved rapidly, and the scope of applications gradually expanded. After the third wave of AI technology development, AI technology has also crossed the threshold of industrialization, gradually taking on the characteristics of a systemic variable in the international system. Countries continue to expand the scope and depth of their applications of AI technology in military, economic, and other fields: Smart weapons and information processing platforms such as unmanned aerial vehicles (UAVs) have gradually changed the face of traditional warfare, and AI’s empowerment of industries and innovation have greatly increased the rate of production and economic development. AI is seen as an important part of the new round of S&T revolution. As early as 2016, analysts realized that the development of AI technology has crossed the threshold of affecting human political life and would become an important influencing factor in the international system in the future. The unprecedented importance of AI technology and equipment may make policies for this field into the most critical element of national policy. The United States is well aware of the significant momentum that AI contains in the fields of national defense, military affairs, and industrial innovation. Therefore, it has always closely linked leadership in AI with the maintenance of U.S. hegemony, continuously introducing many AI development strategies. The competition for the initiative in AI technology R&D and application has become an important component part of the S&T competition between major powers.

有关人工智能的研究起源于1956年约翰·麦卡锡(John McCarthy)和马文·明斯基(Marvin Lee Minsky)等科学家对有关如何用机器模拟人的智能的探讨。探讨中提出的“人工智能”(Artificial Intelligence)概念,打开了人工智能学科研究的大门。在历经数次技术升级后,人工智能的技术发展与应用不断取得突破,其所依据的技术本质从1.0时代的“逻辑推理”,逐步转向2.0时代的“知识工程”与3.0时代的“机器学习”,人工智能能力快速提高,适用范围逐步扩大。在人工智能技术发展的第三次浪潮后,人工智能技术也跨越了产业化的门槛,逐步呈现出国际体系中系统性变量的特征。国家不断扩大人工智能技术在军事、经济等领域的应用范围与深度:无人机等智能武器与信息处理平台逐步改变了传统战争的面貌,人工智能对产业的赋能与革新大幅提高了生产与经济发展速率,被认为是新一轮科技革命的重要组成部分。早在2016年,已有分析者意识到人工智能技术的发展已经跨越了影响人类政治生活的门槛,未来将成为国际体系中的重要影响因素,人工智能技术和设备史无前例的重要性可能使该领域的政策成为国家政策中最关键的要素。美国深知人工智能在国防军事、产业革新等方面蕴含的重大动能,因而始终将领跑人工智能与护持美国霸权紧密联结,不断出台多项人工智能发展战略。人工智能技术的研发和应用主动权竞争成为大国科技竞争的重要组成部分。

S&T companies also continue to accelerate the pace of their exploration of AI technology development. In 2017, Google proposed a new AI deep learning model, the Transformer model. The researchers introduced a self-attention mechanism into the model and proposed a multi-head attention mechanism that can improve the parallel efficiency and operational performance of the model, greatly improving the computational speed of AI when faced with large-scale data. This gives AI strong abilities to understand contextual relationships. After that, OpenAI developed the Generative Pre-Training (GPT) model based on the Transformer model, introducing an unsupervised pre-training mechanism and supervised fine-tuning method into model training. This, combined with the prompt learning method, gives AI powerful natural language understanding capabilities. OpenAI successively developed several versions based on the GPT model. By the time it reached the GPT-3 model, the number of parameters used to train the model had reached 175 billion, achieving in-context learning and allowing the model to produce high-quality answers from a small number of samples. ChatGPT is a conversational AI launched by OpenAI based on the GPT-3.5 model. Researchers at OpenAI introduced the reinforcement learning from human feedback method into the training of ChatGPT. Engineers provided dialog data during AI training, created and used the reinforcement learning reward model, and used proximal policy optimization to fine-tune the model. AI sorts the generated answers by ranking method during the training and learning process, and feeds them back to the system for a new round of training. This way, it can learn the expressions and grammatical rules of human language through large-volume training, thereby simulating the generation process of human language. Therefore, ChatGPT can generate text content in a more natural manner and is viewed as a new-generation knowledge invocation and processing tool. The combination of the above technologies gives ChatGPT more unique capabilities compared to traditional AI.

科技公司也在继续加快探索人工智能技术发展路径的脚步。2017年,谷歌公司提出新的人工智能深度学习模型Transformer模型。研究人员把自注意力(Self-Attention)机制引入模型内部,提出能够提高模型并行效率和运行效果的多头注意力机制(Multi-head Attention),大幅度提高人工智能在面对大规模数据时的运算速率,使其具有良好的理解上下文关系的能力。此后,OpenAI公司基于Transformer模型开发出GPT(Generative Pre-Training)模型,将无监督预训练机制和有监督微调方式引入对模型的训练中,结合提示学习(Promote Learning)方式,使得人工智能具备了强大的自然语言理解能力。OpenAI公司基于GPT模型陆续开发数个版本。到GPT-3模型时,用以训练模型的参数规模已达1 750亿,实现了情景学习(In-context Learning),使得该模型可以通过少量的样本便产生高质量答案。ChatGPT便是OpenAI公司基于GPT-3.5模型推出的对话型人工智能。OpenAI公司的研究人员将基于人类反馈的强化学习方法(Reinforcement learning from human feedback)引入对ChatGPT的训练,由工程师在训练人工智能时提供对话数据,创建和使用强化学习的奖励模型(Reward Model),使用最近策略优化(Proximal Policy Optimization)来微调模型。人工智能则在训练学习过程中将生成的答案通过排名方式排序,并反馈到系统中开展新一轮训练,使其能够通过大量训练习得人类语言的表达方式和语法规则,从而能够模拟人类语言的生成过程。因此,ChatGPT可以较为自然地生成文本内容,被视为是新一代知识调用和处理工具。以上这些技术的组合,共同赋予了ChatGPT相较于传统人工智能更独特的能力。

Figure 1 Features of ChatGPT

图1  ChatGPT特点

Source: Compiled by the author

First, in terms of linguistic expression, ChatGPT can efficiently understand and interpret human language and produce corresponding answers that conform to the laws of natural language, demonstrating strong anthropomorphic expression capabilities. Based on reports, ChatGPT is capable of generating impressive, detailed, human-like written text. Compared with the mechanized and programmed expressions found in interactions with previous AI systems, the ChatGPT model has a built-in scientific and humanized AI language system. Its surprising language “understanding” and expression capabilities exceed those of 90% of people. This enables ChatGPT to rid itself of the lack of emotional expressions found in the output process of previous AI systems during its operation. ChatGPT can conduct basic logical deductions and emotional judgments based on the conversation content, and output text that is as consistent with the context and language characteristics as possible. This gives it broad prospects for use in human-computer interaction fields such as text translation.

第一,在语言表达方面,ChatGPT能够高效理解和解读人类语言,并相应地产出符合自然语言规律的回答,展现出极强的拟人化表达能力。据报道,ChatGPT能够生成令人印象深刻的、详细的、类似人类的书面文本。相较于以往人工智能在交互过程中的机械化、程序化表达,ChatGPT模型内置了科学化和人性化的人工智能语言系统,其出人意料的语言“理解”和表达能力,已经超过90%的人, 使得ChatGPT在运行过程中能够摆脱以往人工智能输出过程中缺少感情色彩的表达方式,并能根据对话内容进行基本的逻辑推演和情感色彩的判断,输出尽可能符合语境和语言特点的文本。这使得其在文本翻译等人机交互领域表现出广阔的使用前景。

Second, ChatGPT shows far superior capabilities compared to traditional AI. ChatGPT shows people the capabilities and potential of large language models based on deep learning, that is, large-scale models with parameters in the hundreds of billions, which can break the traditional scaling law to achieve a qualitative leap in model capabilities. ChatGPT has already demonstrated document translation, data code writing, and other capabilities. The latest GPT-4 model has further demonstrated advanced reasoning capabilities that far exceed ChatGPT, and can even produce functional websites based on handwritten requirements in a few minutes. These capabilities make ChatGPT an auxiliary tool that can greatly improve the efficiency of learning and research. At the same time, ChatGPT also possesses strong learning capabilities. The AI of the past rarely had the ability to make interactive corrections, and its answers were often based on certain “patterns” that were “summarized” from the data. Once the information received did not conform to these patterns, the answers it gave would deviate significantly. Based on self-reinforcement learning through unsupervised pre-training and human feedback, ChatGPT can continuously correct errors in interaction with human discourse and autonomously strengthen its language processing capabilities. Limited by the size of the training database and the failure to update the training data in a timely manner, ChatGPT sometimes makes minor or even common-sense errors. However, it quickly makes corrections and produces records after receiving user feedback. After multiple rounds of training, the accuracy of its output results significantly improves, and it achieves a high degree of intelligentization through continuous technological iteration. Of course, limited by its training data and algorithms, ChatGPT may sometimes “talk nonsense in a very serious tone.” This requires a further strengthening of algorithm capabilities and expansion of the scale of data used in AI training to improve the accuracy and precision of its answers.


Third, the emergence of ChatGPT has made people realize the broad application prospects of generative AI. As a generative AI, ChatGPT can improve natural language processing and natural language understanding in many applications, and demonstrates the ability to generate multi-modal information. In addition to language-based question and answer, ChatGPT has shown us a variety of functions including language translation and text summarization and generation. This reflects the outstanding advantages of generative AI, namely, the ability to generate new content based on training data rather than simply logically arrange knowledge. The AI developed in the past was mainly decision-making AI, which primarily learns the conditional probability distribution in the data and then performs analysis and makes judgments based on the data. Starting from this basis, generative AI has achieved a breakthrough. It can not only perform prediction functions similar to decision-making AI, but can also perform in-depth learning, induction, and creation based on large amounts of data. After receiving instructions, it can independently make judgments and generate corresponding content. Currently, generative AI is being used in many fields, including ultra-high-definition video generation, remote medical diagnosis, and computer code generation. People with no basic skills in painting can even use AI to create award-winning paintings. Generative AI is likely to be the main direction of future AI development. It is foreseeable that, as the training cost of generative artificial intelligence decreases and computing efficiency improves, more auxiliary tools will be used in fields such as drug and chip design. Researchers can optimize product design layouts with the assistance of such AI, reducing the cost and time required for drug and material discovery, and thus greatly improving industry productivity.

第三,ChatGPT的出现使得人们认识到生成式人工智能所具有的广阔应用前景。ChatGPT作为生成式人工智能,能够在许多应用中改善自然语言处理(Natural Language Processing)和自然语言理解能力(Natural Language Understanding),展现出多模态信息的生成能力。除语言问答外,ChatGPT已经为我们展示了包括语言翻译、文本摘要与生成在内的多种功能,这反映出生成式人工智能的突出优势,即能够基于训练数据生成新的内容而非对知识进行基本的逻辑排列。以往开发的人工智能主要是决策式人工智能,主要学习数据中的条件概率分布,根据数据进行分析判断。生成式人工智能则在此基础上实现突破,不仅能够与决策式人工智能发挥类似的预测功能,还能够基于大量数据进行深度学习、归纳及创造,在接收指令后能独立作出判断与生成对应内容。目前,生成式人工智能已被运用在包括超高清视频生成、远程医疗诊断、计算机代码生成等在内的多个领域,没有绘画基础的人甚至能借助人工智能创作绘画作品并获奖。生成式人工智能很可能将是未来人工智能的主要发展方向。可预见的是,随着生成式人工智能的训练成本下降和计算效率的提高,会有更多辅助性工具被用在药物和芯片设计等领域,研究人员可以通过此类人工智能的辅助优化产品设计布局,降低药物和材料发现的成本与时间消耗,以此大幅提升行业生产效率。

(ii) ChatGPT intensifies competition among S&T companies and pushes society to consider it from all sides

(二)ChatGPT激化科技企业竞争态势 引发社会多方位思考

The emergence and popularity of ChatGPT have attracted S&T giants to continue to increase their investment in AI research, launching a new round of competition in this field. In terms of the development of AI technology, after OpenAI launched ChatGPT based on the GPT-3.5 model, it continued on to launch the GPT-4 model that provides stronger comprehensive capabilities. Google has accelerated the training of ChatGPT-like AI and launched the latest version of the large language model PaLM and a new tool MakerSuite. Developers can use this tool to rapidly prototype their own ideas, achieve real-time engineering, generate synthetic data, adjust custom models, and perform other functions. Chinese S&T companies and universities are also stepping up their efforts to develop ChatGPT-like AI in an attempt to catch up with this wave. For example, Baidu quickly launched a large model “Wenxin Yiyan,” benchmarked against ChatGPT.


At the same time, ChatGPT’s powerful capabilities have impacted many areas of society. ChatGPT is essentially a conversational AI, and its output results are mainly text. Therefore, in some information exchange fields based on text exchange, ChatGPT will have a transformative impact on existing interaction and production methods. In the field of education, for example, ChatGPT’s multi-modal applications, information retrieval, and result production capabilities will gradually revolutionize the existing education system. It will not only serve as a “prosthetic limb” to gradually narrow the gap between the knowledge-disadvantaged and others, but also improve the ease of use and accuracy of the adaptive learning systems and enhance the completeness and creativity of teaching results. In the field of public administration, ChatGPT’s rapid processing capabilities for text and other data enable it to help the national government quickly handle certain public matters and realize the integrated development of intelligent government affairs including intelligent decision-making, intelligent management, intelligent services, and intelligent supervision. However, social changes are often accompanied by corresponding risks. As a new knowledge invocation tool, ChatGPT brings ethical risks and knowledge plagiarism to the academic world, which also makes whether to allow its use in academic research a controversial issue. Some people believe that “ChatGPT may be the end of civilization.” They think that widespread reliance on ChatGPT will make people lose their critical thinking. Therefore, how to define the role of similar AI in papers and experiments has become a major problem that the academic community will face.

同时,ChatGPT拥有的超强能力给社会多个领域带来冲击。ChatGPT本质上是一个对话式人工智能,其产出结果方式主要是文本输出,因而对于某些基于文本交换进行信息交流的领域,ChatGPT将会对既有的互动与生产方式产生变革性冲击。例如,在教育领域,ChatGPT拥有的多模态应用、信息检索与结果产出能力将逐步变革现有教育体系,不仅作为“义肢”逐步缩小知识弱势者与他人的差距,还能够提高自适应学习系统的易用度与准确度,提升教学成果的完成度与创意感。而在公共管理领域,ChatGPT拥有的对文本等数据的快速处理能力使其能够帮助国家政府快速处理部分公共事项,并实现包含智能决策、智能管理、智能服务和智能监管在内的智能政务一体化发展。 但是,社会变革往往也伴随着相应风险。ChatGPT作为新型知识调用工具,其给学术界带来的伦理风险、知识剽窃行为也让人们在允许其介入学术研究这一问题上存在争议,有观点认为,“ChatGPT可能是文明的终结”,认为对ChatGPT的广泛依赖将会让人的思维失去批判性。因而如何界定类似人工智能在论文和实验中的角色,成为学术界将要面临的一大难题。

The potential security risks brought about by ChatGPT have caused widespread concern in society. Elon Musk believes that AI (as represented by ChatGPT) is “both positive or negative and has great, great promise, great capability,” but that “with that comes great danger.” Viewed from the level of national security, the impact of the development of ChatGPT and AI on national security has received widespread attention. The U.S. National Institute of Standards and Technology released an AI risk management framework to help researchers think about the development of AI technology, and measure and monitor AI risks and their potential negative impacts. Italy has banned the use of ChatGPT on the grounds of leakages affecting “personal privacy.” At the societal level, the shock caused by ChatGPT also made people worry about the breakthrough development of AI. In Japan, nearly 70% of respondents believe that the country should impose stricter supervision on the development of AI. From this, we can see that ChatGPT will not only have impactful results on traditional AI application fields, but its unpredictable functions and application spillovers will also have a multi-faceted impact on society. Therefore, discussing the impact of the emerging applications of ChatGPT on national security will not only help to clarify doubts about its application and the technical mechanism behind this, but also help to comprehensively understand and correctly view the disruptive influence of generative AI on society in the future.

ChatGPT可能带来的安全风险引发社会的广泛担忧。埃隆·马斯克(Elon Musk)认为,“(以ChatGPT为代表的)人工智能具有巨大的前景与强大的能力,但随之而来的也是巨大的危险”。从国家安全层面看,ChatGPT和人工智能发展对国家安全带来的影响已被广泛关注。美国国家标准与技术研究院发布人工智能风险管理框架,旨在帮助研究者思考人工智能技术发展、衡量和监控人工智能风险及其潜在消极影响。意大利则曾以“个人隐私”泄露为由禁止ChatGPT在意大利境内的使用。 在社会层面,ChatGPT带来的震撼也使得人们对人工智能的突破性发展感到担忧。在日本,接近70%的受访者认为,国家应该对人工智能的发展进行更加严格的监管。由此可见,ChatGPT并不只对传统人工智能应用领域带来冲击性结果,其难以预料的功能和应用外溢也将给社会带来多方位影响。因此,围绕ChatGPT这项新兴应用对国家安全带来的影响展开讨论,不仅有助于对该项应用及背后技术机理的释疑,更有助于全面理解和正确看待生成式人工智能对未来社会的颠覆性影响。

II. New features of national security under the influence of ChatGPT

二 ChatGPT影响下的国家安全新特征

High-tech innovations and breakthroughs have had multiple overlapping influences on national development. From the perspective of international power comparisons, a new round of scientific and technological revolution carried by AI technology is accelerating the reshaping of the global power order. From the perspective of social and economic development, AI can help countries establish a more sustainable growth model that is free from economic crises. Looking back at the development and evolution of several technological revolutions from the past, we can see that they often have a significant or even disruptive impact on national security. The “cross-field extensibility” of national security studies naturally coincides with the broad adaptability of AI, and the development of AI technology is naturally deeply embedded in various national security fields. In the fields of political security, economic security, and military security, AI, as a systemic element itself, will create a wide range of new issues and become the central node of an issue network. In the fields of network security, nuclear security, and homeland security, where security has a relatively specific meaning, AI will become an important enabler, shaping the semi-central area of the network. The remaining national security fields are more loosely linked to AI, forming the peripheral area of the issue network. Therefore, it is very important to explore the multiple overlapping influences of ChatGPT on the national security situation.


(i) The information field will become more complex, increasing the ability of some countries to manipulate international public opinion

(一)信息场域进一步复杂化 助长部分国家操纵国际舆论的能力

Developed Western countries are relatively mature in the AI field, so they have been able to strengthen their control over international public opinion by seizing opportunities and technological advantages. This is more conducive to their values-based targeted manipulation of public opinion and the establishment of information cocoons. Rapidly developing AI will facilitate the generation of disinformation, allowing the creation of a false appearance of consensus and hiding underlying societal fractures. For a long time, the United States has relied on its S&T hegemony and international voice established after World War II. Through its control of international media, the United States has continued to project anti-China public opinion and a negative image of China with views such as the “debt threat theory” and “economic aggression theory.” In terms of propaganda strategies, the Western media often only take up political issues and human rights issues that are in line with Western interests in order to set the agenda, while focusing on negative issues in their reporting. Once this sort of public opinion gains dominance, it can hide its value orientation under the cloak of “collective will” through discourse suppression and creation functions. Generative artificial intelligence, as represented by ChatGPT, has great advantages in conducting public opinion wars: After purposeful training, this type of AI can continuously output speech that is politically biased and contains false information, and convert such speech into the formats of images, videos, and online speech. When exported to international data networks through various public opinion communication channels, this speech will directly or indirectly interfere with international public opinion and affect the public perception of other countries. On this basis, countries with technological and data advantages can combine AI developed based on big data models with various Internet bots to create a new type of “cyber army.” By mixing in false information and distorting facts, this army can concoct large volumes of speech that smear the images of other countries. This not only directly affects public opinion and tarnishes the images of other countries, but also helps the perpetrating countries strengthen their dominance and initiative in the field of international public opinion.


After the large-scale application of generative AI as represented by ChatGPT, the public opinion ecosystem will be further complicated and muddied, and the validity and credibility of Internet information will be further reduced. New media gives every individual the right to speak out. The emergence and application of generative AI such as ChatGPT has reduced the information asymmetry among different individuals in the social information field and greatly increased the frequency and weight of individual voices in public opinion. Information entropy in the public opinion field will greatly increase. The social information field and public opinion field further present the features of “multi-centering,” and the government’s authoritative position in information release and public opinion guidance is severely challenged. With the rapid development of generative AI capabilities, the threshold for creating false information such as fake videos will be significantly lowered. AI can use deep fakes to manipulate elections, exacerbate social divisions, and reduce society’s trust in the government. The credibility of such deep fake content is rapidly increasing with the development of AI. Even originally accurate content can be dismissed as deep fake content, directly affecting the credibility and authority of the government. In recent years, fake news circulating on social media has caused much controversy. For example, fake videos of the U.S. president’s speeches are constantly circulating online, causing people to worry about the president’s health. Some videos have even had a direct effect on the battlefield of public opinion. After the widespread use of generative AI developed based on general large models as represented by ChatGPT, more social groups will use this technology to produce videos, charts, and other information. The power to produce highly credible news and information will no longer be concentrated in the hands of the state and mainstream media, and the entities involved in information dissemination will become more diversified. This will further increase the complexity of Internet information and intensify the existing trend towards the fracturing of Internet society. Social groups that have mastered the technology will compete with the state over the right to interpret the truth, and false information that is difficult to identify as such will further reduce the authority and credibility of the news media and even the government.


In addition, ChatGPT’s “bias amplification” ability will amplify biases and stereotypes in its training data, further marginalizing groups that are already at a disadvantage in society. When using ChatGPT, some users have found it responds to certain questions about Black women by downplaying their contributions, and it sometimes deliberately belittles certain political figures. This is because the data from the current international Internet contains a large amount of biased information. Stereotypes, prejudices, and discrimination have been widely recorded in machine learning methods, and R&D institutions, limited by algorithm capabilities and costs, are unable to completely scrub out such information. Countries with monopoly advantages in international network data management can artificially create “information cocoons” for AI by methods such as inserting malicious information to pollute databases. This will make it so AI develops a thinking framework with ideological and cognitive biases during the training process, affecting its output results and thereby potentially affecting the viewpoints and cognition of users.


(ii) Data security will be difficult to safeguard, and ideological infiltration methods will grow more diverse

(二)数据安全难以保障 意识形态渗透方式更加多元

With the widespread use of AI based on large models and big data, the risk of leakage of important data of citizens and countries will suddenly increase, causing a sharp increase in security risks. At present, the new generation of AI is made possible by deep learning and operations based on big data. The interactive use of large amounts of sensitive data by AI during deep learning not only exposes the private information of humans to AI, but also greatly weakens the government’s ability to supervise data and information. The political security issues caused by personal information leakage already largely extend beyond the traditional security field. From a government perspective, a national government can quickly improve the efficiency of its information collection and decision-making by using ChatGPT, but this requires the government to “feed” relevant documents and data to ChatGPT so that it can make decisions more in line with actual situations. This undoubtedly increases the risk of national data being stolen: On the one hand, the working mechanism of ChatGPT is still a “black box”, and the possibility that it can store and transmit relevant data during its operations cannot be ruled out. On the other hand, the functionality of ChatGPT enables it to fully understand the relationship between data in a short time through self-supervised learning capabilities. In a feature update in May 2023, OpenAI lifted the restrictions on ChatGPT’s Internet connection capabilities, allowing it to “know” the latest information, using specific third-party software as a springboard. Although it still cannot access certain “security-protected” websites, the unlocking of networking capabilities has strengthened ChatGPT’s deep data retrieval and serialization capabilities. ChatGPT and similar generative AI systems collect and associate targeted large-scale data through keyword reminders and other methods, creating “portraits” of certain events or plans. For example, a country’s government could use this technology to collect targeted information on port shipments and important resource transfers in some other country, using this small window to dynamically perceive the domestic conditions of other countries in order to get an early warning of possible military activities. This would significantly reduce the confidentiality and security of government behavior. Without specific countermeasures, countries weak in S&T will be unilaterally transparent to technologically powerful countries, and technologically powerful countries will always hold the strategic and tactical initiative. The power structure among countries will become more unbalanced, forming a “Matthew effect” in the field of international power.


From the perspective of user information exchange, the dialogue between users and ChatGPT is not only a process of AI self-reinforcement learning, but also a process of the disclosure of users’ personal information. When a user uses ChatGPT for information query, ChatGPT can learn the user’s speech characteristics, current areas of interest, and even identity through information exchange. In a networkized information society, each citizen’s state of mind, behavioral habits, and characteristics can be obtained through the collection and processing of monitoring data such as their real-time public speech and online behavior. The constant communication between the public and ChatGPT actually gives AI the opportunity to conduct in-depth analysis of user portraits for the group. Language models not only offer the potential to produce lower-cost propaganda, but can also improve the effectiveness of propaganda by tailoring the quality of propaganda to specific groups. Thus, such models can serve as content generators and disseminators for cognitive warfare between states. Through in-depth group portraits, a state is able to “prescribe the right medicine” and carry out ideological subversion using propaganda models that are more familiar and favorable to the target group. In this way, subversive activities transition from a method of “indiscriminate delivery” to one that is “precise and differentiated,” significantly increasing the effectiveness and efficiency of ideological penetration activities. In addition to drawing group portraits, ChatGPT also gives humans the ability to accurately construct portraits of elites. Currently, AI systems are already able to accurately portray the personal images of other countries’ political elites, who are seen as opponents in a strategic game. AI can analyze and grasp the personality traits of political leaders by using their public speeches on different occasions as data. It can even predict the language characteristics of political leaders based on language data. This makes targeted attacks on perceptions of leaders more difficult to defend against.


(iii) The imbalance in power development among countries will further increase, intensifying changes in domestic and international power structures

(三)进一步加大国家间实力发展的不平衡 国内国际权力格局变动加剧

With the in-depth development of AI technology, the existing competition among technological powers will further intensify. Changes in AI technology will further consolidate the power of the pioneer countries to develop technology in the knowledge field, and the control of this power by major countries will become more invisible and natural. When a certain party’s AI research achieves a breakthrough in algorithms or computing power, this advantage can spill over to other fields, empowering rapid breakthroughs and development in other technologies. Therefore, the cross-field application of such technology will lay down rules for the research ideas of others, and latecomers will rely on this path of research. The dominant party can therefore dominate and formulate the rules and discourse in the general large model research and application fields. At present, the exponential development trend in AI performance is rapid and sustained once it gets going. It is easy for those with advantages in research to transform their algorithm and computing power advantages into an insurmountable superiority over disadvantaged parties, and gain most of the fruits of development. This creates a “winner takes all” situation. Such technological progress will ultimately create winners and losers at the national level. Especially at this critical moment when the international power structure is changing, breakthroughs brought about by technological changes will be regarded by all parties engaged in competition as the “key” to winning the contest. This will inevitably result in more intense, zero-sum competition between different countries or political alliances in this field. Today, the Group of Seven has reached a consensus through the establishment of the “Hiroshima AI Process” in an attempt to “preemptively” seize the leading voice in the international discourse around issues such as suitable values for AI. This means that the establishment of organizations such as transnational data alliances and algorithm alliances may become an important strategy for S&T competition among countries. Technology and data protectionism centered on big models, big algorithms, and big data will become the core content of S&T competition among major powers. International confrontation situations such as S&T blockades and even S&T cold wars may become new trends in international competition.

随着人工智能技术的深度发展,技术大国之间的原有竞争态势将会进一步激化。人工智能技术的变革会进一步巩固知识领域中技术先发国的权力,大国对权力的控制变得更加隐形且自然。当某一方的人工智能研究实现算法或算力的突破,其能将这种优势外溢到其他领域,赋能其他技术快速突破发展。而这种技术的跨领域应用会规制他者的研究思路,后来者会依赖这种研究路径,优势方因而能够掌握对通用大模型研究和应用领域的规则以及话语主导权与制定权。当前,人工智能效能的指数级发展态势,呈现出“一步快、步步快”特点,研究优势方很容易将算法和算力优势转化为对其他劣势方的碾压态势,获得大部分的发展成果,形成“赢者通吃”的局面, 此类技术进步最终在国家层面会制造赢家和输家。特别是当下正处于国际权力格局变动的关键时刻,技术变革带来的突破将会被竞争中的各方视为赢得竞赛的“钥匙”,这就造成不同国家或政治联盟必将在此领域展开更激烈、零和性的竞争。如今,七国集团已就建立“广岛人工智能进程”达成共识,试图“先发制人”抢夺人工智能适用价值观等方面的国际话语权。这意味着组建跨国数据联盟、算法联盟等组织可能会成为国家间展开科技竞争的重要方略,以大模型、大算法和大数据为核心的技术与数据保护主义将成为大国科技竞争的核心内容,科技封锁乃至科技冷战等国际对抗态势可能将成为国际竞争的新态势。

The first feature is that unbalanced S&T development will worsen the “Matthew Effect” in the development among countries, making it more difficult for developing countries to change their disadvantaged position in the global economic system. The capitalist global economy is based on a worldwide division of labor, with different actors assuming different economic roles. The advantages of developing countries participating in the global economic system are mainly their relatively low labor costs and ample supply of natural resources. However, in the AI era, there is a possibility and tendency for peripheral areas to be permanently marginalized. Although generative AI can significantly improve the efficiency of industrial production and experimental R&D, many developing countries lack the digital infrastructure needed to power AI, the innovation environment needed to build new models that utilize AI, and the skills to fully utilize its power. This means that these countries can only continue to be forced to accept industries transferred to them from developed countries. Generative AI has extremely high requirements for cumulative training and infrastructure, which determines that the R&D and application of this technology have cumulative features. Long-term investment is needed to achieve in-depth research and development of the technology. This makes it difficult for other countries, especially countries under chip import and export restrictions, to achieve “overtaking on the curve” (弯道超车, using new opportunities to overtake current leaders), widening the “rich-poor gap” among countries in the development of AI industries. The replacement of labor by AI technology has further diluted the demand for labor in the capitalist economic system, while the development of the AI industry has not yet created new labor positions. This has further weakened the relative advantages of developing countries, and their relative position in the global economic system is further shifting to the periphery. Countries with advantages in technology and capital can build a “center–periphery” economic and S&T development structure for the new era, relying on their advantages in the development of AI technology to continuously strengthen their core position in the economic and S&T research fields. Meanwhile, countries at the periphery of the structure will be stuck with the choice of obtaining some core technologies through exchanges with and learning from core S&T countries in order to promote the upgrade of their national industrial structure, or else continuing to receive industries transferred from developed countries. The former choice means that peripheral countries will continue to strengthen their dependence on central countries, making it more difficult to break out of the system structure that exploits them. The latter choice means that relatively underdeveloped countries will miss out on the wave of cutting-edge technology R&D and applications, and their development potential will be further reduced, ultimately leading to the further expansion of the gap created between countries by the “Matthew Effect.”

第一个特征是,科技发展不平衡将恶化国家间发展的“马太效应”,发展中国家在全球经济体系中的不利地位更加难以改变。资本主义世界经济以世界范围内的劳动分工为基础,不同行为体承担不同的经济角色。发展中国家参与世界经济体系的优势主要是相对低廉的劳动力成本和充足的天然资源供给,但在人工智能时代出现了边缘地区被永久边缘化的可能和倾向。尽管生成式人工智能能够大幅度提高工业生产与实验研发效率,但许多发展中国家缺乏为人工智能提供动力所需的数字基础设施、建立利用人工智能的新模型所需的创新环境以及充分利用其力量的技能,意味着这些国家只能继续被迫接受来自发达国家的产业转移。生成式人工智能对训练积累和基础设施的极高要求决定了该技术的研发和应用具有累积性特征,必须依赖长期投入才能够实现技术的深度研发,使得其他国家特别是芯片进出口受限国家难以实现“弯道超车”,加大了国家间人工智能产业发展的“贫富差距”。而人工智能技术对劳动力的替代进一步稀释了资本经济体系对劳动力的需求,人工智能产业的发展暂时也没有实现创造新的劳动岗位, 这使得发展中国家拥有的相对优势进一步被削弱,在世界经济体系中的相对位置进一步向边缘推移,技术和资本占据优势的国家能够据此构建新时代的“中心—外围”经济与科技发展结构,凭借其在人工智能技术发展上的优势不断强化在经济和技术研究领域的核心地位;而处于边缘结构的国家将陷入选择在与核心科技国家的交流学习中获得部分核心技术来推动国家产业结构的升级,或者继续维持承接发达国家转移产业的抉择中。前者意味着边缘国家将继续强化对中心国家的依赖,更加难以突破体系结构对其的剥削;后者则代表发展相对落后国家将错失对前沿技术研发和应用的浪潮,发展潜力将会被进一步缩减,最终导致国家之间发展“马太效应”的进一步扩大。

The second feature is a further enhancement of the power of non-state actors and a trend toward the “multi-centering” of political power. Big data, algorithms, and even AI, while bringing profound changes to society, are also becoming effective means for platform companies to capture power. Core technologies such as the learning iteration of generative AI are mainly in the hands of S&T giants such as Google and Microsoft. Such ownership will give rise to a group similar to a “transnational S&T community.” The information monopoly of S&T companies gives them power over the acquisition and interpretation of information. The originally relatively stable boundaries of power between sovereign states, especially in terms of power to produce public goods, will become blurred or even disappear due to the arrival of new entities. The data training and algorithms used for the update and iteration of ChatGPT and other similar generative AI are mainly in the hands of large technology companies, especially Internet companies, and groups of scientists, so it is difficult for the government to get involved. Moreover, the broad application prospects of generative AI in many fields allow S&T companies to participate in national affairs and share some power, which in turn creates the possibility for these companies to intervene in or even dominate affairs over which the state originally had a monopoly of power. Through long-term investment, technology R&D, and the practical application of research results, S&T giants will assume the role of producing and providing public goods in many areas of society. Therefore, in some fields such as economic and social management, the right to supply public goods will to a certain extent be transferred from the traditional political authority held by nation-states to the power of capital. Government power will be further differentiated, and the disconnect between technical authority and bureaucratic authority may gradually erode the effectiveness of the national government’s actions and allow national power to be “hijacked” by S&T companies.


(iv) Generative AI will enable battlefield weapons, turning the cognitive battlefield into an important area of contest

(四)生成式人工智能赋能战场武器 认知域战场成为重要博弈场域

Technical equipment equipped with new types of AI will significantly affect and change the current battlefield situation. AI has been called the third revolution in warfare, following the invention of gunpowder and the atomic bomb. ChatGPT’s high adaptability and powerful information processing and output capabilities give it broad prospects for battlefield applications. The CIA is currently discussing how to use ChatGPT and similar programs to assist in intelligence collection and espionage operations. The large-scale application of ChatGPT and similar AI in military operations will greatly enhance military capabilities. In terms of battlefield target recognition and combat information processing, ChatGPT can be loaded onto drones and military equipment to perceive battlefield information in real time. Recognition networks equipped with the GPT-4 model can effectively improve the accuracy of target recognition in complex environments, greatly improving the combat effectiveness of drones. ChatGPT can also efficiently connect different battlefield information perception systems in series to create intelligence-sharing terminals through perception, aggregation, processing, and output methods. AI based on the GPT model architecture can continuously filter large volumes of battlefield information and data and flag its important content, improving the efficiency of manual information processing and analysis and providing command departments with accurate information and data and decision-making bases. In terms of battlefield command, combat command systems controlled by generative AI have faster response rates, higher-level decision-making, and more outstanding efficiency compared to traditional command systems. In the face of increasingly complex battlefield formats and exponential growth of battlefield information, generative AI based on big data technology and trained on military decision-making styles can efficiently analyze the combat needs of commanders and provide highly accurate risk assessment and decision-making suggestions through large-scale model processing, combat simulation, and deduction based on war game deductions, helping commanders discover and defeat threats before they arise and greatly improving the level of ad hoc decisionmaking on the battlefield. In terms of logistics, ChatGPT can be combined with technologies such as the Internet of Things (IoT) and cloud computing to dynamically monitor the use and deployment of strategic military materials. It can improve warehouse management efficiency by analyzing data such as the quantity of stored materials and maintenance status, achieving a dynamic balance between demand and resource transportation to provide optimal logistics supply solutions. Therefore, when ChatGPT is applied in the combat field, it will significantly raise the military’s intelligentized information acquisition and decision-making level. The military’s combat capabilities are closely linked to the intelligentization level of its equipment. Battlefield information collection and decision-making methods will transform by moving in the direction of “cloud collection” and “AI decision-making.”


The emergence of ChatGPT will also highlight the importance of cognitive confrontation capabilities between states. New cognitive warfare and information warfare weapons developed based on generative AI, big data, and big algorithms may reduce trust between countries. NATO’s cognitive warfare doctrine holds that victory is defined more by capturing the psychological and cultural high ground than the geographical high ground. Cognitive warfare is highly dependent on support from network information technology and social media platforms. Social media platforms and networks not only give each individual the ability to obtain information and communicate instantly, but they have also become important platforms from which hostile countries launch targeted attacks. Such attacks can be divided into cognitive attacks, such as transmitting false battlefield information and fake speeches by leaders, and information attacks such as destroying other countries’ information transmission systems and intercepting and deciphering information transmitted by other countries. In these areas, generative AI will become a powerful tool for cognitive warfare: After attack information is processed by generative AI, its capabilities for concealment and deception will be greatly improved, gradually giving cognitive warfare the features of “highly interactive, occurring throughout all time and space from multiple perspectives.” The military can train generative AI to continuously strengthen its capabilities in the production and delivery of false information. It can intensively carry out multi-angle information bombing and emotional incitement on the opponent’s people through AI generated images of battlefield witnesses and victims. Through continuous interaction and communication with the public, it can continuously indoctrinate people with falsified facts and create targeted information fog, which in turn can generate war-weariness and anti-war sentiment in other countries and threaten the stability of public opinion in the domestic societies of other countries, ultimately overcoming the opponent’s will to resist. At the same time, generative AI can launch non-stop information bombing against the enemy’s information receiving systems, affecting the normal operation of the enemy’s network technology facilities, cutting off the enemy’s information reception, and reducing the efficiency of internal communication within the enemy’s military. Therefore, the outstanding performance of generative AI in cognitive warfare may aggravate public opinion relations between countries. Every country will always be on guard against possible disinformation attacks and tend to blame any disruptions arising from public opinion within society on the actions of other countries, especially enemy countries. This will further weaken the foundation of strategic mutual trust between countries.


(v) ChatGPT training and operations will impact existing legal requirements and significantly lower the threshold to crime

(五)ChatGPT训练与运转冲击现有法律规定 犯罪门槛大幅降低

At present, the party who is responsible for materials related to ChatGPT is not clear, and ChatGPT can easily infringe on the rights of others during training and use. As a very large model with more than 100 billion parameters, ChatGPT cannot provide a clear logical reasoning process and clear data source for each decision or output result produced during its operation. Even the “real” sources it provides may be accompanied by real details and false citations. When an accident results from the use of the output results of generative AI, there is no clear subject who can bear this part of the responsibility. Considering this issue, the U.S. Supreme Court once discussed the entities liable for AI chatbots, and one justice suggested that “the legal protections that shield social networks from user content lawsuits may not apply to AI-generated work.” At the same time, the large amount of data used to train ChatGPT means that it is difficult to prevent the theft and misuse of personal information, significantly increasing the risk of data leakage. It is difficult for the existing legal system to restrict related behaviors. The capabilities of generative AI such as ChatGPT come from the capture and learning of massive data, but this involves issues such as whether the behavior of capturing data is compliant with laws and whether the captured data will be used for other purposes. It is also difficult to verify the legality and validity of data sources. This is deeply related to the legality of the source of generative AI capabilities. In addition, the training and operation methods of ChatGPT undermine the public’s informed consent mechanism, in that data capture is carried out by S&T companies without individual consent, making it difficult for the public to protect the privacy of personal data. It is even difficult for citizens to detect the leakage of their private personal data. Payment information, addresses, taxi records, and even chat records may be easily obtained by others due to program design flaws and widespread user use, making it increasingly difficult to protect the privacy of information. In addition, the relevant companies are not transparent about the procedures used to process training data. Personal information and confidential government information may not be fully processed before being used directly for training. After the training, the relevant data may also be deliberately retained by the company, increasing the possibility of data leakage.


At the same time, the widespread use of generative AI will lower the threshold for cybercrime and make it more difficult to track criminal behavior. Europol believes that ChatGPT is a valuable resource for potential criminals who lack the technical knowledge to generate malicious code. As a conversational AI, ChatGPT can quickly produce the required type of code and generate programs based on user requirements, making it a natural criminal tool when in the hands of criminals. Although each company currently claims to have set up a program for rejecting inappropriate requests for its artificial intelligence products, the current programs do not yet have the ability to identify the deeper intentions behind requests, and users have still found that they can “trick” the program during use to achieve their purposes. At the same time, generative AI has the ability to edit batches of text and forge pictures and videos on a large scale. Criminals can combine these capabilities to carry out telecommunications fraud and other activities in a way that better aligns with cultural and social characteristics, so that their tactics are more “authentic” and “effective.” Criminals can take advantage of information gaps to create scams that are more in line with logic and strike at the weakness of people who lack the ability to discern information. The frequency and success rate of these crimes will increase significantly.


(vi) ChatGPT will bring about structural unemployment, affecting the labor market and social stability

(六)ChatGPT带来结构性失业现象 影响劳动力市场及社会稳定

As an important component of national security governance efforts, employment risk governance is embedded in the system of national development practices. This technological revolution, represented by AI, may bring about many types of unemployment. ChatGPT and other generative AI will permanently replace some jobs in certain industries, posing the risks of collective and long-term unemployment risks to social development. As a new productivity tool, ChatGPT will inevitably have a severe impact on the current ways in which people work and the balance of supply and demand in the labor market. Companies can quickly obtain and extract information through generative AI such as ChatGPT. For example, OpenAI Codex, which is based on the GPT-3 model, is proficient in more than a dozen programming languages and improves programmers’ efficiency through “memorizing code”. Generative AI has also greatly improved the productivity and creativity of the workforce. Low-skilled workers can use tools such as ChatGPT to narrow the gap in work quality between themselves and high-skilled workers.

就业风险治理作为国家安全治理事业的重要组成部分,内嵌于国家发展的实践体系中。此次以人工智能为代表的技术革命可能将带来多类型的失业现象,ChatGPT及其他生成式人工智能将永久性地取代某些行业的部分岗位,对社会发展带来群体性和长期性失业风险。ChatGPT作为全新生产力工具,必然对当前人们的工作方式和劳动力市场的供需平衡产生剧烈冲击。公司可通过ChatGPT等生成式人工智能对信息进行快速获得和提取,例如,基于GPT-3模型产生的OpenAI Codex精通十几种编程语言,通过“背代码”方式以提高程序员工作效率。生成式人工智能还大幅提高了劳动力的生产效率与创造力,低技能工人则能通过使用ChatGPT等工具缩小他们与高技能工人之间工作质量的差距。

However, the substantial improvement in productivity also means that the labor force required for existing jobs will be reduced, as well as the demand for manual workers and workers doing basic intellectual work, and it will be possible to replace some repetitive labor and creative work with generative AI. From the perspective of economic development, the development and application of disruptive and breakthrough technologies will affect the existing economic format and economic distribution relations, reshape the current employment and social and economic pattern, and produce destructive substitution effects on existing technical groups. A report released by OpenAI in March 2023 shows that approximately 80% of the U.S. workforce will see some of their work tasks affected by GPT models and related technologies. A report released by Goldman Sachs shows that, at present, about 2/3 of the jobs in the United States and Europe are “affected to some extent by AI automation,” and as many as 1/4 of the jobs can be completely performed by AI. Different from the replacement of low-end manufacturing jobs by AI of the past, the job positions that generative AI can handle are more concentrated among positions where similar operating rules and procedures are applied and that require frequent processing of large amounts of similar data. These behaviors can be easily completed through generative artificial intelligence. For example, you only need to input basic news copywriting rules and information capture procedures into generative AI, and it can automatically capture news data and “fill” it into a manuscript, greatly reducing the time required for news writing. Therefore, the widespread use of generative AI such as ChatGPT in these fields will reduce corporate operating costs while improving operational efficiency. The main production entities in this field will shift from humans to AI, and some workers doing intellectual tasks will lose employment opportunities. At the same time, we should also note that the current ChatGPT is a milestone in the development of generative AI, but it is by no means the end of that development. In the future, generative AI may serve as the “control console” and “core brain” of other AI technologies or industrial technologies. Besides some positions that require innovative thinking or subjective adaptability, most fields will see the introduction and dominance of production entities based on generative AI, unleashing an unpredictable surging tide of labor unemployment.

但是,工作效率的大幅提高也意味着既有工作所需劳动力将减少,体力劳动者及部分简单脑力劳动者需求量将大幅度下降,部分重复性劳动和创造性工作存在被生成式人工智能替代的可能。从经济发展的角度看,颠覆性和突破性的技术发展与应用都会影响既有的经济形态和经济分配关系,重塑当前就业和社会经济的格局,对原有的技术人群产生破坏性的替代效应。OpenAI公司2023年3月发布的报告显示,大约80%的美国劳动力会有部分工作任务受到GPT模型和相关技术带来的冲击。而高盛公司发布的报告则表明,美国和欧洲目前约有2/3的工作“在某种程度上受到人工智能自动化的影响”,并且多达1/4的工作可以完全由人工智能完成。 不同于以往人工智能对低端制造业岗位的替代,生成式人工智能可胜任的岗位更多集中在拥有类似运作规则和运转步骤,需要频繁处理大量相似数据的岗位,这些行为都能轻易通过生成式人工智能完成。如只需将基本新闻文案规则与信息抓取程序输入到生成式人工智能中,它就能自动地抓取新闻数据“填写”到文稿中,大幅节省新闻撰写所需的时间。因此,ChatGPT等生成式人工智能在这些领域的广泛使用将在降低企业运营成本的同时提高运营效率,该领域的生产主体从人类转向人工智能,部分脑力劳动者将会因此失去就业机会。同时我们还应注意到,当前的ChatGPT是生成式人工智能发展的一个里程碑,而绝非发展的结束。未来生成式人工智能可能将作为其他人工智能技术或者工业技术的“操纵台”和“核心大脑”,除了部分需要创新性思维或者主观应变能力的岗位外,将有更多领域被以生成式人工智能为核心的生产主体介入和主导,不可预期的劳动力失业潮将汹涌而至。

An organized schematic of the impact of ChatGPT on national security is shown in Figure 2.


Figure 2 Schematic diagram of ChatGPT’s impact on national security

图2 ChatGPT影响国家安全的示意图

Source: Compiled by the author

III. Strategies to cope with the new security situation

三  新安全形势下的应对策略

The emergence of ChatGPT poses unprecedented challenges to national security. At the domestic level, ChatGPT impacts existing legal norms and ethics, affects the labor market, and influences ideological stability. At the international level, this technology will affect the dynamics of S&T competition in the international community and impact the international power structure. Needless to say, we can say these challenges are massive, or even disruptive. Therefore, we must prepare for a rainy day and make corresponding preparations in advance, strengthen international cooperation in the construction of innovation systems, and work hard to improve laws and regulations in order to lay a solid foundation for safeguarding China’s national security in the ChatGPT era.


(i) Internally, we must accelerate the construction of an open, collaborative, and diversified AI innovation system.


First, we can establish decision-making institutions including an AI development committee to solve the issue of an overall governing force in the AI field. Official departments such as the Ministry of Science and Technology and big data administration bureaus can coordinate the development pace of technology-based enterprises and promote the improvement of China’s general large model R&D and innovation capabilities. The development of existing S&T enterprises currently focuses more on technology R&D and application centered on market competition, with investment in applications far exceeding investment in basic theoretical research. The situation of internal competition hinders improvements in technological innovation and application capabilities. Therefore, taking accelerating basic theoretical development and breakthroughs in major technological applications as our point of departure, we should innovate the technology R&D and application paths of existing enterprises and universities. Based on the diverse and interconnected characteristics of general large model R&D, we should fully realize the interaction and collaboration between various entities and levels in the R&D and application chain to improve our original innovation capability in the general large model R&D system. We should focus on solving the pain points and difficulties in the development of general large models and accelerate the pace at which we are catching up with international leaders.


From the perspective of lateral R&D, the government should strengthen support for basic research, especially model data training, and use tax subsidies and other preferential policies to encourage enterprises to make incubation investments around basic theoretical research. Entities such as technology enterprises, universities, and governments play complementary roles: Different enterprises can integrate their data resources, build shared open-source databases and computing engines, reduce barriers to entry by evenly sharing scientific research costs, facilitate large model training for innovative S&T companies, and build a technology R&D model in which “the government provides basic support, multiple parties share responsibilities, and enterprises verify.” In view of the difficulties in the development of existing large models, enterprises and universities should set up key joint research teams to explore the feasibility of different solutions and paths through long-term research. We should leverage the capabilities of S&T companies to convert research results into applications in order to quickly verify technical paths, thereby breaking through the obstacles to the development of large models. From the perspective of vertical applications, entities up and down the production chain should orient themselves to the actual application needs of society, give full play to the multi-modal application capabilities of general large models, create innovative application solutions based on large model capabilities, and quickly realize good iteration and upgrade of application capabilities through methods such as research and interactive feedback. The government should collect opinions from experts and scholars, set up “test fields” for the conversion of technology into applications around key application scenarios, continue to empower enterprises to conduct large-scale experiments on applications, innovate practices for generative AI operations, achieve the integration of the needs of generative AI technology and the current needs of social development, and accelerate the innovative application of generative AI in multiple fields.


In addition, we should gradually reform the talent training model and delivery system, build a high-tech talent cultivation alliance linking universities, research institutes, and enterprises, and provide more intellectual support for generative AI R&D. Talent is the first resource, and the technological innovation system cannot operate without the support of talents. With the explosive rise of ChatGPT, the demand for talents studying the training and application of general large models has surged. However, China still has a large gap in research-type talent in the field of the development of general large models, with a shortfall of nearly 30 million relevant talents. Therefore, in order to promote the rapid development of generative AI, it is necessary to speed up the cultivation of domestic talent.

其次,逐步改革人才培养模式与输送体系,建设高校、研究所、企业三体联动的高新技术人才培育联盟,为生成式人工智能的研发提供更多智力支撑。人才是第一资源,技术创新体系的运转不能缺乏人才支撑。随着ChatGPT的爆火,对研究通用大模型训练和应用的人才需求激增,但当前中国在研发通用大模型方面仍然存在较大的研究型人才缺口,相关人才的缺口接近3 000万。因此,为推动生成式人工智能的快速发展,就必须加快国内人才的培育速率。

The education system, led by universities, should gradually expand the scale of AI research and development majors, subdivide the research directions of these majors, increase the cultivation and application conversion courses on general large model training, and improve the basic AI literacy of school students. At the level of high-quality talent delivery, universities can directly connect with enterprises and research institutes and regularly push high-quality talents to existing AI R&D and application systems through the establishment of joint internship mechanisms.


The government and enterprises should increase their support for relevant talents, attract top talents in relevant fields, selectively introduce some teams of outstanding scientists, and directly enhance the country’s innovation and research capabilities in this area. For example, the government can include talents in the purification of general large model training data and model applications in the existing overseas high-level talent recruitment plan requirements, and by setting up admission criteria and graded benefits, continuously find and attract high-quality overseas talents to return to China to advance their careers. Objectively, there is still a large gap between China and foreign countries in terms of large models and other technologies. Attracting overseas talents should become an important means of quickly improving R&D capabilities. To this end, we should break down ideological constraints, recruit talents with a more open and inclusive attitude, and build a solid talent foundation for the development of AI in China.


(ii) Externally, we must share in the AI technology development achievements to break down the situation of technological hegemony

(二)对外共享人工智能技术发展成果 破除技术霸权状况

As a new technological singularity, generative AI can effectively empower different industries, produce industrial innovation, and greatly improve productivity. It should become a powerful helper for realizing a community with a shared future for mankind. However, most of the value created by AI has not been truly enjoyed by most countries. Instead, it has become a new tool for countries with technological hegemony to expand their monopolistic advantages. At present, the main force in the development of large models is still the leading S&T enterprises, and various R&D entities have not yet reached a consensus on technical standards and usage rules for the development of general large models at the level of international society. At present, China has a certain degree of influence in the development of AI and has the ability to unite other countries to change the “haves” and “have-nots” situation in the field of AI R&D. Therefore, under the new situation, China should continue to adhere to the concept of a community with a shared future for mankind and realize technology and research achievement sharing within the international community so as to slow down the trend of technological polarization on a global scale, win more friends from among developing countries through technology sharing, and breaking the United States’ “Technological Democratic Alliance” (技术民主联盟).


China should actively explore and discuss international rules and standards related to the development of general large models internationally and actively pursue the right to make rules and technological leadership in new technology fields. This requires China to launch initiatives in multiple areas. In terms of data connectivity, China can take the lead in establishing regional organizations such as a General Large Model R&D Cooperation Organization. On the basis of establishing deeper strategic mutual trust with other countries, China can engage in cooperation based on data and technology exchanges, sign data sharing agreements to build larger databases covering social data from multiple countries, and jointly establish a team of experts from various countries to manage the database. While ensuring data security, this would provide solid support for the training and strengthening of general large model capabilities, effectively broadening the scope of applications and depth of use of the model. In terms of application sharing, China can work with other countries to develop common large-model applications according to local conditions. According to the needs of different countries, such as disaster early warning and social service management, and through methods such as resource exchange and establishment of joint ventures, with guarantees provided by government and technical support and maintenance from leading enterprises, we should cooperate with different countries to develop general large model applications and build digital information hardware facilities, realizing the direct sharing of generative AI achievements. In terms of sustainable development, China can help other countries train the information talents they need and build a basic information talent cultivation system by dispatching scientific research teams, remote guidance, and mutual visits for learning. These practices can not only effectively share AI development achievements and “realize the dreams” of other countries’ informatization and intelligentization development goals, but also effectively enhance China’s influence in international digital information standards and application development paths. In addition, by setting agendas and discussing data processing standards at AI conferences and in international organizations, China can regulate the application scenarios and capabilities of general large models, coordinate the steps taken by various countries in global AI governance, and make positive contributions to guide the development of general large models so that they work for the good.


(iii) We must innovate existing domestic governance tools and improve laws, regulations, and policies in the AI field

(三)创新现有国内治理工具 完善人工智能领域法律法规政策

Network technology is neutral in itself. We must not only observe the impact of emerging technologies, represented by ChatGPT, on the current international security order, but also realize that this is an important opportunity to innovate security governance tools. After all, the new security issues brought about by new network technologies should also be addressed through the development of corresponding governance tools. With the rapid improvement of generative AI capabilities, existing security detection and processing tools are unable to deal with the threats it brings. For example, in the face of forged video and simulated data attacks, traditional information detection tools have difficulty determining the authenticity of information. Only supervision tools built on the same technical foundation can detect and manage such attacks. This also requires enterprises to assume the responsibility of using AI information review technology to regulate the proliferation of false information and frequent criminal frauds that may be induced by AI.


At the same time, generative AI can revolutionize existing governance tools. Based on this technology, enterprises and governments can jointly develop highly automated tools that require data collection and processing, such as natural disaster early warning tools. This can improve society’s early warning and response capabilities for traditional governance issues. Generative AI can also play a supporting role in social services and other fields. Therefore, the comprehensive utilization and development of “generative AI + X” services will help promote the establishment of smart governments and friendly societies. This is a key step in creating an interconnected and intelligent society.


In addition, legal and regulatory systems should also be improved to regulate the development and application paths of generative AI. First, the law should establish a reasonable system of legal regulations around the copyright of the content generated by generative AI and the legitimacy of the source of the content, provide basic legal support for the content created by generative AI, avoid hindering the training and development of generative AI due to violations of intellectual property protection and other regulations so as to encourage and regulate the large-scale production of S&T companies and the deployment of their applications in multiple fields, achieving the stable implementation of the achievements of generative AI research. The abuse of big data by S&T companies itself touches on the right to keep personal information private and the right to know, but excessive restrictions will hinder the development and training of big data models by S&T companies. Therefore, laws and regulations must be introduced to clearly define the boundaries of data that can be used for general large model training and find a balance between ensuring model training and protecting information privacy. For example, using document-based generative AI as an example, the government should negotiate with enterprises on the data types and scale of training texts and “feed” the corresponding content to the AI to avoid intrusion into some unauthorized and private data.


Second, the government and enterprises should implement effective supervision of the AI industry with clearly defined responsibilities, and implement AI technology safety standards. Enterprises must establish a standardized production system when training and using generative AI, consult with the government and submit documentation in a timely manner on the training data and related feedback of generative AI, focus on cooperation and development of high-quality training data, and ensure the reliability of generative AI capabilities from the source. The government should set up a review team for AI-generated content to regularly review and inspect the operation and use of generative AI. This can then be used as a basis for the assignment of scores linked to the credit rating and preferential tax treatment of development companies, forcing companies to pay attention to safety standards in the process of developing and training AI and ensuring that generative AI does not incorporate discriminatory and biased values into its output framework during the training process. Enterprises should also create compliance design plans to promote AI innovation and ensure system transparency and explainability.


Finally, the law should clarify the entities responsible for generative AI and improve the legal liability system accordingly. Transparency and accountability are critical issues in the development and deployment of generative AI, as they are key to ensuring that the technology is trustworthy and remains fair and secure. If a technology or application lacks an entity that can bear responsibility, criminals who use the technology as a criminal tool can easily escape relevant responsibilities. This actually encourages entities in society to abuse it. Therefore, when formulating future laws, we should pay attention to the sharing of responsibility among the designers, users, and maintainers of generative AI, clarify the rights and responsibilities of different entities in the form of laws and regulations, establish corresponding responsibility tracking mechanisms, unify standards for the division of responsibilities, establish reasonable and reality-based principles of attribution, reasonably allocate responsibility for accidents to people or “actual actors,” and cooperate with network information security departments to jointly combat new types of cybercrimes. At the same time, the law must also put forward higher requirements for the producers and maintainers of generative AI and fundamentally reduce the possibility of generative AI being used to commit crimes.



We can expect generative AI, as represented by ChatGPT, to change the security situation facing the country, but it will also bring new opportunities for national development. The report to the 20th Party Congress pointed out that we must “promote the development of integrated clusters of strategic emerging industries and build a number of new growth engines such as new-generation information technology, AI, biotechnology, new energy, new materials, high-end equipment, and green environmental protection.” This requires us to look at the impact of this technological breakthrough on the country’s development situation from multiple perspectives and realize that generative AI as represented by ChatGPT will serve as a new-generation “Internet platform”, following on the heels of the Internet to become another tool to deeply connect different development fields and accelerating China’s process of building an intelligent society and information superpower.


The emergence of ChatGPT is not the endpoint of the development of this type of AI, but one of the “singularities” in AI development and application model research. We need to pay full attention to its ample development potential. As similar AI technologies continue to be injected into the development and architecture of digital applications, the software and devices we use in our daily work will become more intelligentized and humanized, and virtual assistants, virtual doctors, and other such things will develop more comprehensive and more innovative functions. In addition, this will break through some of the physical limitations of existing humans, presenting characteristics such as “round-the-clock service, high-precision operations, and multithreaded work.” We can predict that this will achieve substantial improvement in S&T research and development and social service efficiency. In the future, as training scales increase and algorithms are updated, this type of AI will present more comprehensive and powerful functions, with more access ports and more diversified application fields, and will continue to innovate the existing knowledge transfer networks and means of information acquisition in society. It will essentially become the “external brain” of human beings in the digital age, achieving a leap in productivity.


However, the development of generative AI such as ChatGPT also means that AI will pose unprecedented and complex challenges to the national security situation, which requires us to promptly and squarely face up to the role of similar AI and its status in societal development. Only in this way can we prepare for the high-tech competition of the future.


To top

Cite This Page

黄日涵 (Huang Rihan), 姚浩龙 (Yao Haolong). "A Reshaped World? New Features of Artificial Intelligence and National Security under the Rise of ChatGPT [被重塑的世界?ChatGPT崛起下人工智能与国家安全新特征]". CSIS Interpret: China, original work published in Journal of International Security Studies [国际安全研究], June 28, 2023

FacebookTwitterLinkedInEmailPrintCopy Link