The Core Challenge of AI Policy: Coordinating Conflicting Values and Objectives

As discussed above, AI policy involves multiple policy objectives and values operating at the same time. It must promote technological innovation while also responding to new risks. It must pursue economic growth while considering the social distribution of benefits and opportunities. It must maintain openness and cooperation in the global AI ecosystem while securing national technological autonomy and security.

The problem is that these objectives do not always move in the same direction.

Therefore, the core challenge of AI policy is not to choose one objective over another, but to coordinate the relationships among multiple policy objectives and values that are all legitimate, yet may sometimes come into conflict.

The Paradox of Technological Innovation and Risk

The first tension in AI policy appears in the relationship between innovation and risk.

In general, the development of new technologies brings productivity gains, the emergence of new industries, and greater consumer benefits. Governments therefore seek to support research and development, cultivate talent, build data and computing infrastructure, and improve unnecessary regulations that may hinder corporate innovation.

However, the development of new technologies can also generate risks and social impacts that did not previously exist. This problem has long been explained in technology policy through the Collingridge dilemma. In the early stage of a new technology, its social impacts and risks are difficult to predict sufficiently, but policy intervention is relatively easy. Conversely, once the technology has diffused widely and its impacts and risks become clearer, it has already become deeply embedded in society and industry, making it far more difficult to control or redirect its trajectory (Genus & Stirling, 2018).

AI illustrates this dilemma particularly clearly.

As the performance of generative AI improves and its areas of application expand, new economic and social opportunities are created. At the same time, new problems such as misinformation, deepfakes, and copyright infringement can emerge. As AI agents become more autonomous, productivity and convenience may increase, but so too may concerns about actions outside direct human control and uncertainty over responsibility. The development of high-performance AI models also requires enormous computing resources and capital, raising the possibility that technological progress may strengthen the market power of a small number of firms.

In policy terms, this dilemma appears as the problem of under-regulation and over-regulation.

If regulation is too weak, problems such as privacy violations, discrimination and bias, safety incidents, misinformation, and concentration of market power may expand, weakening social trust in AI. Conversely, if regulation is too strong, the costs and uncertainty associated with developing new AI services may increase, making market entry more difficult for startups and small and medium-sized enterprises. In particular, if smaller firms bear greater compliance burdens than large firms that can absorb regulatory costs, regulation intended to reduce risk may paradoxically strengthen the market dominance of incumbent firms.

The important question, therefore, is not simply whether regulation should be strengthened or relaxed. It is what level of regulation is necessary, for which risks, and for whom.

This is where the risk-based approach becomes important. Rather than regulating all AI systems in the same way, a risk-based approach differentiates the level of policy response according to the domain and purpose of use, as well as the scale and severity of expected harm.

For example, AI used for everyday information search or content creation does not need to be treated in the same way as AI used in medical diagnosis, financial lending, hiring, critical infrastructure, or public administration, where individual rights or safety may be significantly affected. In low-risk areas, a broad degree of freedom for innovation and use may be allowed. In areas where the possibility of serious harm exists, however, higher levels of transparency, safety verification, human oversight, and accountability may be required.

The European Union’s AI Act, which regulates AI systems differentially according to their level of risk, is a representative example of this approach. Rather than uniformly restricting the use of AI itself, it imposes different obligations depending on the nature and level of risk (European Union, 2024).

The important implication of the risk-based approach is that it does not necessarily treat innovation and safety as a zero-sum relationship. Appropriate safety standards and accountability mechanisms may create short-term costs for firms, but in the long term they can also increase social trust in AI and provide an institutional foundation for the stable diffusion of the technology.

Ultimately, the task of AI policy is not to choose between innovation and risk. It is to build a policy framework that preserves the possibility of innovation while managing the risks that arise in the process.

Coordinating Growth and Distribution

The tension among policy objectives does not exist only between innovation and risk. Even if AI generates high productivity and economic growth, how its benefits are socially distributed is a separate issue. In fact, the faster AI spreads across the economy, the greater the possibility that positive effects such as productivity gains and wealth creation will be accompanied by widening gaps in income and opportunity.

Firms and workers that can actively use AI may benefit from productivity gains, while those that cannot may fall behind in competition. Gaps may also widen between regions with AI infrastructure and advanced talent and those without, and between countries with access to advanced AI technologies and those without. An even more important issue is that AI and automation may not merely replace some existing tasks, but transform the labor market and employment structure itself.

If the new jobs created by AI do not sufficiently replace the jobs that disappear, or if a large gap emerges between the capabilities required by new jobs and the capabilities held by existing workers, then even if technological innovation leads to economic growth, it cannot be assumed that the benefits will be evenly shared across society.

However, restricting the use of AI and technological innovation itself is unlikely to be an appropriate solution. The key is not to suppress growth in order to solve distributional problems. It is to sustain technological innovation and economic growth while socially managing the inequalities and transition costs that arise in the process.

To do this, policy must expand access to AI and AI literacy, support workers’ job transitions through education and retraining, and create foundations that enable small and medium-sized enterprises and the public sector to use AI. At the same time, policymakers must prepare for the possibility that the speed and scale of technological change may exceed what existing employment policies or unemployment insurance systems can handle.

The Challenge of AI (2021) also suggested that, in response to large-scale changes in employment structures caused by AI and robotics, societies may need to consider new income-security systems beyond traditional unemployment insurance, such as a Guaranteed Minimum Income (GMI) or Universal Basic Income (UBI). The specific design and financing of such systems will inevitably involve significant social and fiscal debate. Nevertheless, as AI increasingly changes the relationship between labor and income itself, the question of how to use part of the productivity and wealth created by technological innovation for social transition and safety-net construction can no longer be treated as a peripheral welfare issue. It becomes a core task of AI policy.

Therefore, coordinating growth and distribution is not a matter of choosing one over the other.

The goal of policy should be to realize the growth potential created by AI as fully as possible while ensuring that the transition costs generated in the process are not concentrated on particular classes or groups. Opportunity-expanding policies through education and retraining, transition policies that support labor mobility, and, where necessary, new forms of income security and social safety nets must be designed together.

Ultimately, coordinating growth and distribution in the AI era does not mean merely redistributing the gains of growth after the fact. It means designing growth and distribution together within a single policy framework so that more members of society can participate in the process of technological innovation and share in its outcomes.

Openness and Strategic Autonomy

At the international level, another tension appears between openness and strategic autonomy.

The development of AI technology has been driven by global research collaboration, talent mobility, international supply chains for semiconductors and cloud infrastructure, data, and open-source ecosystems. Especially in frontier AI, which requires enormous computing resources, capital, data, and highly skilled talent, it is realistically difficult for any one country to build the entire technology ecosystem independently. Excessive technological protectionism and barriers between countries may weaken the very foundations of innovation and slow the pace of technological development.

However, as AI moves beyond being a mere industrial technology and becomes a core infrastructure of the economy and society, as well as a strategic asset directly linked to national security, the need to secure a certain level of technological autonomy is also increasing. If a country depends too heavily on particular foreign countries or firms for core AI infrastructure and technologies, external decisions such as export controls, supply-chain disruptions, or changes in global AI platform service policies can affect its industries, public services, and even national security.

Against this background, Middle Eastern countries such as Saudi Arabia and the United Arab Emirates have recently invested actively in national AI infrastructure and model development. In Korea as well, sovereign AI policies, including the development of domestic AI foundation models, have been pursued. At the same time, controversies have emerged in the process of implementation and evaluation, leading to discussions about reassessing existing development methods and seeking new implementation systems. As a result, the goals and methods of sovereign AI policy themselves have become important policy issues.

The important point is that sovereign AI policy should not be viewed through the lens of technological self-sufficiency.

As AI becomes the foundation of national security and essential public services, securing a minimum level of independent operational capability, control over core data, and choice over computing infrastructure and models becomes a form of national risk management. In order to prepare for the possibility that the global AI ecosystem may become fragmented by geopolitical conflict or technological controls, states need to maintain a certain level of strategic choice.

This perspective is connected to real options theory, which argues that in highly uncertain environments, maintaining options for responding to future changes has strategic value, rather than committing all resources to a single choice. In uncertain environments, strategic value lies not only in selecting the currently optimal option, but in preserving the flexibility to expand, redirect, or, if necessary, withdraw investments as circumstances change (Trigeorgis & Reuer, 2017).

Applied to national AI strategy, this means that the key issue is not the ability to develop all core technologies domestically or to become self-sufficient. Rather, it is the ability to choose among alternative technologies, suppliers, and infrastructures when the external environment changes, and, when necessary, to shift toward independent capabilities.

In this book, I refer to this capacity as strategic optionality.

Strategic optionality is important not only for current technological competition, but also for preparing for future AI development pathways. Until now, AI competition has largely been a first-stage race to improve the performance of large foundation models. In the future, however, a second-stage race may become more important: the diffusion of accumulated AI capabilities across industries, products, and services.

As model compression, on-device AI, and edge computing advance, AI can increasingly be implemented in various devices and industrial settings without necessarily depending on ultra-large models and massive data centers. This shift can create new business models and industrial opportunities for Korea, which has strengths in smartphones and electronics, semiconductors, automobiles, and manufacturing.

Therefore, maintaining a certain level of independent AI technological capability is not merely an investment in the current LLM race. It also means securing future options in preparation for the diverse applications and industrial diffusion of AI that may unfold in the future.

From this perspective, the goals and scope of sovereign AI should be set realistically.

Rather than aiming to replace the LLMs of global leading firms with domestic models or to secure all AI technologies independently, the priority should be to secure alternative options and independent operational capabilities in areas where they are necessary, such as national security, essential public services, and strategic industries.

If competition with global frontier models that require enormous capital and computing resources becomes the policy goal itself, the state may face serious problems in allocating limited national resources efficiently. Therefore, sovereign AI should be approached not as an end in itself or as a direct measure of global competitiveness, but as a means of securing strategic optionality in necessary areas while actively participating in the global AI ecosystem.

Coordination Matters More Than Balance

The three tensions discussed above show that AI policy is not a simple optimization problem. It is a matter of choice and coordination among different values and objectives. Innovation and risk, growth and distribution, openness and strategic autonomy are all important policy objectives, but their relative importance and priority are not always the same.

Therefore, there are limits to understanding this issue merely as a matter of “balance.” The term balance can give the impression that an optimal policy can be found by assigning appropriate weights to different values. In actual AI policy, however, the importance of each value changes depending on the stage of technological development, the area of application, the degree of risk, economic and social conditions, and the international environment.

In areas where safety is critical, risk management may have to take priority. In areas where new markets need to form, more space for innovation and experimentation may be necessary. If a serious AI divide emerges, access and inclusion may become more important. If geopolitical uncertainty increases, strategic autonomy may become a more urgent policy objective.

Good AI policy, therefore, is not policy that treats all values equally or seeks a fixed point of balance among them. What matters is the policy capacity to judge which values and objectives should be prioritized under changing conditions, and to continuously coordinate relationships among policy goals that are all legitimate but may conflict with one another.

In this sense, the core of AI policy lies not in finding a single balance point, but in continuously adjusting that balance point in response to changing circumstances.

This is where the problem of governance emerges.

When different values and interests collide, who should decide priorities? What roles should be played by government, business, experts, and civil society? How should authority and responsibility be divided between the state and the international community?

In the end, coordination in AI policy is not merely a technical matter of policy design. It is a question of who performs that coordination, and according to what principles and procedures.

This leads to the next issue: the multilayered governance of AI policy.