Designing AI Policy in an Age of Uncertainty: Adaptability and Principles

One of the greatest difficulties facing AI policy is that, at the moment when policy decisions must be made, it is difficult to know sufficiently the direction of technological development and its social impacts. It is difficult to predict precisely how fast a new technology will develop, in which industries and services it will spread first, and which risks will actually materialize.

As the rapid diffusion of generative AI has shown, technological change can exceed the scope and speed anticipated by policy. Even after a new technology emerges, it often takes considerable time before its economic and social impacts become clear.

The problem is that policy cannot simply wait until such uncertainty disappears.

If intervention occurs only after risks have been sufficiently confirmed, harms may already have occurred on a broad scale. Conversely, if overly strong regulation is introduced when risks are still unclear, it may constrain the development of technologies and markets that have not yet fully formed.

AI policy is therefore less a matter of traditional decision-making, in which policymakers select the optimal policy after securing sufficient information, and more a matter of making decisions under incomplete information, observing their consequences, and continuously revising policy.

This characteristic requires a change in the very method of policy design.

Traditional policy design has often assumed a relatively linear process: set policy goals, select instruments to achieve them, implement the policy, and evaluate the results. In the field of AI, however, where technological change is rapid and uncertainty about future impacts is high, there are limits to the idea of designing a complete policy from the beginning. If overly detailed rules are created based on current technologies and risks, the gap between the rules and reality may widen quickly as the technology changes.

Therefore, AI policy should be designed not as a completed institutional arrangement, but as an adaptive system that can be revised in response to new information and experience. Rather than trying to regulate every possible situation from the outset, policy should intervene where necessary, observe both intended and unintended effects, and adjust policy instruments when new risks or technological changes emerge.

Regulatory sandboxes, pilot projects, phased regulation, regular policy evaluation and review, and mechanisms that reconsider whether a regulation should remain in force after a certain period are all possible instruments for implementing this approach. What matters is not any particular instrument itself, but the fact that the policy is designed from the beginning to be revisable.

This approach also requires changes in how policy instruments are selected.

Rather than trying to solve complex policy problems through a single regulatory instrument, the policy mix becomes important: the combination of regulation and support, information provision and technical standards, incentives and self-regulation. Policy instruments do not operate independently. One instrument may strengthen the effect of another, or it may weaken it.

Good policy design, therefore, does not mean using as many policy instruments as possible. It means combining different instruments so that they work consistently toward the same policy goal. This is why the perspective of policy mixes, which emphasizes coherence and complementarity among policy instruments, is important (Howlett & Rayner, 2007).

In AI policy, such policy mixes do not need to remain fixed.

The intensity and combination of policy instruments can change according to the maturity of the technology and the level of risk. In the early stages, when the potential and risks of a new technology are not yet sufficiently understood, policy may focus on disclosure, monitoring, voluntary standards, and experimental institutions. As risks become clearer or social impacts grow, stronger policy instruments such as certification, audits, and legal obligations can be added.

Conversely, if existing regulations lose their necessity because of technological change or unnecessarily restrict innovation, they should be revised or abolished. The capacity to change becomes as important a policy capability as policy stability itself.

However, there is also a risk in emphasizing adaptability and flexibility alone.

If policy can continue to change depending on circumstances, firms and citizens may find it difficult to predict the standards by which governments make decisions. The discretion of policymakers may also become excessive. If even the basic direction of policy changes frequently in the name of responding quickly to technological change, market uncertainty may increase and social trust may weaken.

Therefore, policies that can change also need standards that should not change.

This is where the importance of principles-based policy emerges.

In a field such as AI, where technological forms and modes of use change rapidly, it is difficult to define every possible situation in advance through detailed rules. If rules are made based on the form of a particular technology or service, regulatory gaps or overlaps may arise whenever new technologies appear.

By contrast, basic principles such as human safety and rights, accountability, transparency, and human control can provide direction for policy even when the concrete form of technology changes.

A principles-based policy does not mean eliminating rules or leaving everything to corporate autonomy. Rather, principles provide the direction that policy must follow, while rules and technical standards play the role of implementing those principles in concrete situations.

For example, if human control is established as a principle, the question of when humans should intervene and what decisions can be delegated to AI may be designed differently in healthcare, finance, education, and national defense. Privacy protection and transparency can also be translated into different obligations and technical standards depending on the level of risk and the context of use, rather than applying the same procedures to every AI system.

This approach shows that principles and rules need not be seen as opposites.

Principles alone may leave the standards for actual behavior unclear, while detailed rules alone may be unable to respond to a rapidly changing technological environment. Therefore, basic values and responsibilities should be presented as relatively stable principles, while the rules, technical standards, and policy instruments that implement them should be designed so that they can be revised in response to technological and social change.

Principles stabilize the direction of policy.

Adaptability enables that direction to be implemented in response to changing realities.

Ultimately, what matters in designing AI policy in an age of uncertainty is not only the ability to predict the future accurately. More important is to accept from the beginning that predictions may be wrong. Instead of trying to create perfect rules from the outset, policy must observe changing technologies and risks, learn from the effects of policy, and revise policy instruments when necessary.

At the same time, in order to prevent such changes from becoming arbitrary policy shifts, policy must clearly establish principles that should continue to be protected, such as human rights and safety, accountability, and transparency.

Good policy in the AI era is therefore neither the policy with the most rules nor the policy with the greatest flexibility.

It is policy that clearly maintains the principles that must be protected while adjusting the instruments that implement them in response to changing circumstances. In other words, it must be policy that is both principles-based and adaptive.

Rather than trying to eliminate uncertainty, AI policy must build a system that can learn and revise itself within uncertainty.

That is the core of policy design required in a rapidly changing AI era.