From Major-Countries Coordination to Global Mechanisms: an Analysis of the Global Governance of Large AI Models

The global governance of large artificial intelligence (AI) models calls for urgent actions. Large models, representing a type of technological innovation, is advancing global technological progress and is also compelling the international community’s attention to the role of global governance system in coordinating the development of AI. In building a global governance mechanism for large models, both traditional governance system and new international system should be explored to construct Regime Complexes for the governance of large models.

Definition and Risks of Global Governance of Large Models

The concept of global governance of large models is interpreted in two ways. For one, it refers to the application of large models in addressing challenges facing the global governance. As a new development of AI technology, large models can be applied in various aspects of global governance. For another, as a new technology, the large models themselves are bringing about many emerging global problems and global governance should be employed to solve such problems. This paper adopts the second interpretation. Large models could pose three risks: unemployment, disorder, and loss of control.

The first is unemployment. Large models inevitably exacerbate unemployment, which may trigger economic or political crises as unemployment expands. In early 2023, researchers at OpenAI released an article analyzing the impacts of GPT on labor market. If we examine current technological progress, we can see this article has accurately foreseen the impacts of large models on the job market, especially on such professions facing a fairly high rate of job replacement as translators, designers, programmers, financial practitioners, and writers. Further development and employment of large models will continue to exacerbate the problem of unemployment.

The second is disorder. Order of knowledge is the foundation of political order. As a new-type of knowledge generator, large models will undoubtedly impose significant impacts on the traditional order of knowledge. In other words, as large amounts of content generated by large models keep emerging, they are highly possible to create a purely virtual world, where it is extremely difficult for people to establish trust and may even trigger a more serious political trust crisis. Such a political trust crisis, if combined with an economic crisis, will exert more serious social impacts.

The third is loss of control. The loss of control mainly stems from the capacity and consciousness of large models themselves. Large models indicate the possibility of the artificial general intelligence (AGI). Once AGI becomes reality, it is highly likely to create two types of problems: the first is super-capacity. GPT-3 now has 175 billion parameters and GPT-4 using the Mixture of Experts model may reach whopping trillions of parameters. According to the Scaling Laws, the super-capacity of large models will suddenly emerge at a certain scale of parameters. If such super-capacity is employed to make nuclear weapons or biological viruses, effectively controlling it will be difficult for humans. The second is super-consciousness. Large models, if having the capacity of AGI, may develop self-consciousness. Relevant research at the moment indicates that large models possess the consciousness of human individuals at a certain stage, which can be proven by the Theory of Mind. Once large models develop self-consciousness, they may adopt “jailbreak behavior” to get rid of human control. In addition, large models might also mislead humans into making wrong decisions, further resulting in bitter conflicts.

In this sense, the key to the global governance of large models is classified governance. In other words, it is necessary to distinguish between two types of models. One is super-large models. The fundamental principles behind the governance of super-large models should first be non-proliferation and non-malicious use. Due to its kind of super-capacity, the large models, once being used for malicious purposes, will have huge negative impacts. The principle of minimum proliferation should be applied for such super-capacity. But in real scenarios, according to the technological logic adopted the United States now, the country must possess such super-capacity to secure its leading edge. Some countries, as late-comers, tend to adopt the approach of making technological breakthroughs. Therefore, making breakthrough in building super-large models is merely a matter of time, but such models should not proliferate. In other words, if major countries will inevitably be able to build super-large models, it is still necessary to prevent widespread proliferation. Because only when the number of countries possessing such models is limited could the probability of their malicious use be limited to a certain extent. It is also naturally necessary to formulate a set of strict rules to regulate developers and users of super-large models. Users should not employ super-large models to carry out malicious activities, such as conduct dimensional collapse strikes on other countries or overthrow regimes.

The other is the general models. The fundamental technical logic behind the governance of general models is that such models are allowed to proliferate and be used in peaceful ways. With relatively small parameters, general models have fewer chances to develop super-capacity and super-consciousness and thus may impact human society mainly by bringing unemployment and disorder. Large models, essentially as new productive tools, will improve human’s economic well-being to a certain extent and humans cannot completely stop the adoption of such large models. The governance of general models is mainly carried out by building more precise pricing mechanisms to control the speed of the use of large models to a certain extent, so that they can play a more positive role in human society.

Management and Control of Super-Large Model

The precondition to manage and control super-large models is to clarify the definition of super-large models. Widely-adopted approach at the moment is to define based on parameters, within the range of approximately ten trillion to hundred trillion. The focus of the governance of super-large models should be controlling of super-consciousness and super-capacity. Since super-large models are positioned to develop super-capacities, the logic of limited development should be applied, which refers to non-tolerance of unlimited development, similar to the case of nuclear weapons. At the initial stage of the development of nuclear weapons, a basic logic was to save ammunition of conventional weapons, falling within the scope of taking nuclear weapons as certain kind of productive forces. But by the end of the Cold War, the world was equipped with so many nuclear weapons that were enough to destroy the Earth several times. The basic logic of major-countries’ nuclear strategy is “Mutual Assured Destruction”. Therefore, limiting the quantity and scale of nuclear weapons is needed.