[Special Feature: AI Society and Public Space] AI in Politics

2019/02/05

British political philosopher Michael Oakeshott viewed conservatism as a skepticism toward rationalism. Furthermore, British politician Edmund Burke, often cited as the "father of conservatism," was wary of progressivism. Both were concerned that tradition—wisdom accumulated by predecessors through trial and error that falls through the cracks of formulation and verbalization—would be destroyed in the name of "rationality" and "progress." In other words, the conservative stance can be summarized as exercising humility in social reform based on the premise of the limits of human knowledge.

In this article, I will discuss Artificial Intelligence (AI) in politics, particularly in the policy-making process. This is not about AI as an object of governance (industrial policy or regulatory theory), but rather AI that carries out governance or influences legislation. I want to discuss not only the extent to which AI can replace politics, but also whether such replacement is desirable, and whether humans have achieved the standards we demand of AI. In doing so, we will examine AI not just as a program that performs inference and learning based on data, but as a symbol of rationalism and progressivism, or as an object of people's expectations and desires. This is because what is currently being questioned is our own attitude as humans. Constitutional scholar Joji Shishido has diagnosed that a significant portion of the problems expected to arise from AI lie with the humans who use it, and this also applies to AI in the policy-making process.

However, this topic may sometimes seem like science fiction. Therefore, below, I will divide the application of technology into several stages of the policy process to demonstrate its feasibility and limitations.

"RegTech" is currently attracting attention in the use of AI in politics. A coined term combining "Regulation" and "Technology," it refers to companies and other entities using advanced technologies, including AI, to respond to regulations more effectively and efficiently.

RegTech has seen remarkable development primarily in the financial sector. This is because regulations were rapidly strengthened after the 2008 financial crisis, placing a heavy burden on risk analysis and reporting obligations, and increasing compliance costs. The streamlining of risk data collection and analysis, as well as real-time monitoring of payment information, are seen as promising. Companies providing these services for a fee have appeared mainly in Europe and the US, and the business is also emerging in Japan.

While RegTech refers to the response of regulated parties such as financial institutions, implementation also requires the cooperation of regulators and supervisors. Furthermore, the UK government gives it a more proactive meaning, emphasizing not only the reduction of regulatory compliance costs but also the aspect of using technology to improve productivity. Additionally, authorities themselves are working on the development and implementation of supervisory tools.

This is an example of AI utilization in the implementation and operation of policy, a stage close to the "exit" of the policy process. Compared to stages such as policy formulation, decision-making, and public opinion formation discussed later, discretion is relatively small. In this stage, the "rule of law" applies, and the aim is to exclude arbitrary or capricious actions by rulers ("rule of men"). Therefore, there is likely little resistance to the introduction of AI in policy implementation.

Regarding the "rule of law" and AI in the judiciary, constitutional scholar Keigo Komamura has already discussed this. Since criminal law scholar Hiroki Sasakura is also analyzing AI in criminal justice, I will omit those topics here.

Next, in the stage of creating the policies to be implemented, how can AI be used, and what challenges exist?

To begin with, "legislative facts" that form the basis and premise for enacting laws include not only past and present facts (existing facts) but also future predictions (future states and consequences) based on them. However, it is difficult enough just to accurately grasp past and present socio-economic facts. In addition, making predictions from existing facts is a more complex task than fact-finding.

Therefore, it is conceivable to collect and analyze various statistics, materials, and administrative business data, then use AI to select and aggregate information, recommend information to be referenced or issues to be considered, and present future prediction scenarios. This would allow for more efficient use of the scarce human and time resources of the executive branch and the Legislative Bureau of the Houses that support the Cabinet and members of parliament who have the right to submit bills.

Formulation involves relatively more discretion compared to the implementation stage. Therefore, it is a stage where the difficulty of control increases. The aforementioned efforts contribute to promoting data-based judgment and making the judgment process transparent, allowing for a certain level of control without compromising flexibility. This can also be said to resonate with the trend of "Evidence-Based Policy Making (EBPM)." In EBPM, seeking evidence can check "distortion" and "sontaku" (acting on unspoken cues) caused by political pressure, and is said to lead to the deterrence of "policy-based evidence making." Expectations for improving the "quality" of legislation can also apply to the use of AI in policy formulation.

However, there are challenges to proceeding with implementation. First, as a general rule, high-quality datasets are not available, requiring a lot of pre-processing. Furthermore, it can be pointed out that policy evaluation and effect verification are not necessarily being conducted sufficiently at present. To improve the accuracy of AI, data related to policy evaluation would be very meaningful as (additional) training datasets. However, for many civil servants who tend to aim for the infallibility of administrative activities and base their work on a "demerit system," policy evaluation may be seen as a risk to career development. Also, considering that some results are not suited to quantification, focusing only on easily measurable indicators as performance may lead to over-adaptation, potentially diminishing effects that cannot be captured by performance evaluation indicators.

In that case, it might be more realistic to start with introduction in the stages of control and verification after formulation, rather than justification during formulation. This would involve AI utilization in parliamentary questioning, accounting audits, and "Administrative Project Reviews." However, attention must be paid to the power dynamics of verification itself. Therefore, after carefully examining mechanisms where checks and balances function and functions that are acceptable at the site of policy formulation, it is necessary to proceed with gradual introduction alongside reviews of statistical work, changes to personnel evaluation systems, and thorough implementation of the spirit of the Statistics Act and the Public Records and Archives Management Act.

Even if AI utilization in policy formulation is expected and acceptable, AI utilization in the preceding stages might not be permitted (even if technically feasible).

Let's look at policy decision-making in more detail. The point of the section before last was that when there are options such as methods a1, a2, a3... for a certain policy issue A, better options might be presented by making issue organization and future prediction more sophisticated and efficient. However, in policy decision-making, there are situations where one must compare issue A with other policy issues B, C, D... For example, comparing how to allocate a budget of 50 trillion yen among medical care for the elderly, support for young unemployed people, and the promotion of science and technology.

If the policy issues are common, they can be compared to some extent with the same yardstick, but if not, one gets lost in a maze of incomparable values. Yet, there are still moments when a decision must be made. Such decisions are supposed to be made in the legislature (after adjustments within various ministries and the ruling party).

To further enrich these adjustments and discussions, it has been proposed to collect and analyze preference data inferred from opinions and behavioral histories scattered across mass media and the internet, and to present public opinion in real time.

Regarding this proposal, a conflict with the representative system (Article 43, Paragraph 1 of the Constitution) can be pointed out. If one takes the view that the reason the Constitution adopted a representative system is to ensure two processes of discussion: (1) discussion between the people and their representatives, and (2) discussion among representatives in the legislature, then a situation where the latter discussion becomes hollow is hard to overlook. In other words, it would be impermissible for AI to replace discussion and decision-making beyond decision support.

Next, if we imagine AI as a facilitator to encourage deliberation among representatives in the legislature, the hypothesis that agenda-setting is the source of political power will immediately come to mind. It is also necessary to consider the current situation where mass media, which is supposed to handle the agenda-setting function, is in a predicament of declining credibility. The difficult challenge of examining ideals such as "fairness/justice" and "impartiality/neutrality" and building trust lies ahead.

In addition, attention should be paid to the relationship between deliberation and decision-making. Deliberation is a method of preference formation rather than a method of decision-making. In other words, it is a process of searching for "reasons" to justify one's own claims, and sometimes preferences change upon hearing the "reasons" of others. Of course, consensus on preferences may not be reached (in fact, empirical research shows that clashes between extreme views and the steering of discussions by a loud minority are common). Therefore, a decision by voting is required as a system to terminate deliberation. Nonetheless, it is said that voting and deliberation should be connected. This is because the decision can be verified by having the preferences and reasons formed through deliberation explained (simply put, when a policy ends in failure). The responsibility of the Cabinet mentioned in Article 66, Paragraph 3 of the Constitution can be understood as this kind of accountability.

In that case, if the ensuring of accountability can be prompted by AI, it might contribute to the realization of a better policy-making process.

However, the problem becomes even more complex because the very premise of encouraging discussion is being questioned. Have we not been giving "acclamation" to the breaking of "politics that cannot decide"?

If a better policy-making process only means the maximization of performance, then democracy as a form of government is subordinate to elitism or authoritarianism. And now, as an alternative to democracy, the "rule by AI (or AI-supported technocrats)" is newly emerging. As pointed out in public choice theory and quantitative political science, democracy is not epistemically superior to other options. The current situation where the economic progress of China and Singapore is often spoken of with a gaze of envy (despite knowing the numerous social issues) compared to the sense of stagnation in the US and EU seems to corroborate this.

Of course, democracy can be defended if one adopts a position that emphasizes the normative value of procuring legitimacy from deliberation and democratic processes. However, it is not necessarily self-evident to what extent individual and collective self-determination, and by extension freedom, should be emphasized in governance. The problem setting of which should be emphasized—self-determination/freedom or performance/happiness—and for what reason, "Freedom, or else happiness?", remains an open question.

And it was fake news that made us realize once again that freedom, one of the principles supporting the system, is a kind of fiction.

Fake news became a social issue triggered by events such as the 2016 UK referendum on leaving the European Union (EU) and the US presidential election that produced President Trump. At that time, the focus was on fabricated text and photos becoming articles and spreading through sharing on SNS and other platforms. However, it is expected that video will become the main battlefield in the near future. This is because a group of technologies called "deep fake" is advancing rapidly. Deep fakes use image processing that also utilizes AI to quickly synthesize human mouth movements and voices, creating high-quality fake videos without much cost or time. Since public speaking by politicians can be forged, there are concerns that it will be misused as fake news.

Of course, fabricated articles and propaganda are long-standing problems. However, there is a possibility that quantitative increase will turn into qualitative change. The "marketplace of ideas" theory, which is the basis of democracy, was based on the idea that false information and low-quality speech would be weeded out through free competition. However, if fake news is mass-produced and circulated by AI, the verification of truth and falsehood will fall further behind, and as a result of exceeding human cognitive limits, there is a risk that the "marketplace of ideas" will become dysfunctional.

As a countermeasure to such situations, Germany's Network Enforcement Act (NetzDG), which attracted international attention as a fake news regulation law, was enacted in 2017 (for details, see Hidemi Suzuki, "News and Legal Regulation in the Web Era: From the Case of Germany," Mita-hyoron (official monthly journal published by Keio University Press), June 2018 issue), and voluntary regulation by companies is also progressing. Technological development is also increasing, and various methods such as fact checking of individual claims, measuring the reliability of information sources, and detecting the presence or absence of audio/video processing are being tried. Here too, the use of AI is expected, and a competition between AIs is emerging between those who create fake news and those who detect it. AI can be used not in a direction that endangers the foundation of democracy, but also in a direction that maintains democratic values.

As we have seen so far, exploring the possibility of better policy-making through AI also serves as an opportunity to reflect on whether the policy-making process to date has ensured rationality and accountability. Not only that, it can extract values that cannot be exhausted by rationality and efficiency alone. And the two are not necessarily contradictory. Constitutional scholar Tatsuhiko Yamamoto preaches "binocularism," arguing that we should aim to implement AI well in a way that contributes to the better realization of constitutional principles, without being trapped only by the logic of economic rationality and efficiency.

At the beginning, I stated that what is currently being questioned is the human attitude. Even if it is uncertain and clumsy (as in this article), the act of verbalizing what we have sought from politics and what we want to seek from politics in the future, and sharing images through conversation, is not only about thinking about the product specification of "AI in politics," but also a practice of creating "the shape of this country."