A shallow defence of a technocracy of artificial intelligence: Examining the political harms of algorithmic governance in the domain of government

Artificial intelligence (AI) has proven to be superior to human decision-making in certain areas. This is particularly the case whenever there is a need for advanced strategic reasoning and analysis of vast amounts of data in order to solve complex problems. Few human activities fit this description better than politics. In politics we deal with some of the most complex issues humans face, short-term and long-term consequences have to be balanced, and we make decisions knowing that we do not fully understand their consequences. I examine an extreme case of the application of AI in the domain of government, and use this case to examine a subset of the potential harms associated with algorithmic governance. I focus on five objections based on political theoretical considerations and the potential political harms of an AI technocracy. These are objections based on the ideas of ‘political man’ and participation as a prerequisite for legitimacy, the non-morality of machines and the value of transparency and accountability. I conclude that these objections do not successfully derail AI technocracy, if we make sure that mechanisms for control and backup are in place, and if we design a system in which humans have control over the direction and fundamental goals of society. Such a technocracy, if the AI capabilities of policy formation here assumed becomes reality, may, in theory, provide us with better means of participation, legitimacy, and more efficient government.


Introduction
Artificial intelligence (AI) has proven to be superior to human decision-making in certain areas. This is particularly the case whenever there is a need for advanced strategic reasoning and analysis of vast amounts of data in order to solve complex problems. Few human activities fit this description better than politics. In politics we deal with some of the most complex issues humans face, short-term and long-term consequences have to be balanced, and we make decisions knowing that we do not fully understand their consequences. One step at a time, AI has conquered realms of great complication, such as chess, Go and the game of StarCraft. While these are games, some now suggest that the same techniques of machine learning used to conquer such games could also be used to conquer politics. One example of this is the suggestion that an 'AI Economist' based on reinforcement learning and multilayered agent-based simulations could make better tax policies, improving productivity while simultaneously reducing income inequality (Zheng et al., 2020).
In this paper I examine an argument in favour of involving AI in political decision-making. This idea is based on the premise of AI's stipulated ability to make better decisions than humans in certain areas, and the premise that we ought to implement the best policies possible. This is a hypothetical situation in which technologies such as the AI economist just discussed are mature, ready for real life applications, and functional on a large scale. This is not a technical treatise about AI but an analysis of the political theoretical implications of AI in politics. As such, I introduce a thought experiment called the magical decision box, which is used to test a set of political theoretical arguments for and against an AI technocracy.
There are several approaches to the analysis of the effects of computer decision-making in politics. Algorithmic governance is not an entirely new phenomenon, and various harms following the use of AI to govern various aspects of human affairs have been examined in much detail. There are, amongst other issues, concerns related to issues of privacy and surveillance, bias and inequality, transparency and procedure, and freedom and autonomy (Danaher, forthcoming;Sadowski and Selinger, 2014). Sadowski & Selinger (2014) provides a taxonomic tool for technocracy, in which they propose categories for the domain, means, and harms of technocracy. I limit my examination to the political harms of technocracy. Furthermore, I consider only a subset of the phenomena that are analysed in the literature on algorithmic governance, as I focus on the technocracy in the domain of government, and mainly through the means of mandates. In addition, my case is one in which the technocracy of AI is based on systems with low transparency and high degrees of automation (Katzenbach & Ulbricht, 2019). This is a consideration of an extreme case of a technocracy of AI, and this case serves well to highlight the political harms to which I limit my analysis. These choices make this article a partial contribution to the overall consideration of the dangers of technocracy. It is, however, a central part, as these dangers have not been sufficiently examined, particularly not in the setting of an extreme case of AI technocracy as I consider here.
The political harms I examine are mainly based on ideas from the discipline of political theory. The first is that human beings might require political activity in order to thrive. If that is the case, the drawbacks of technically 'better' decisions might outweigh the benefits. The next problem is that legitimacy might suffer if people are taken out of politics. I will here examine several approaches to this problem, and how it might be possible to overcome it through regulation and popular participation in guiding computer politics. We could also argue that computers should not make decisions that affect citizens' lives and wellbeing. The fourth problem is that we lack transparency when we rely on decisions that are beyond the capacity of human reasoning. Without such transparency we will be left to conduct politics on the basis of trust in machines. A final issue is that of accountability when AI is involved in making political decisions.
However, none of these objections are strong enough to fully debunk the argument in favour of employing AI in politics, if a technocracy of AI is implemented with a satisfactory system of control, mechanisms of backup, and a participatory element. Such a technocracy, if the AI capabilities of policy formation here assumed becomes reality, may, in theory, provide us with better means of participation, fair and impartial political outcomes, and more efficient government resulting in benefits for most individuals and society in general.
Should we actually choose to implement a technocracy of AI, politics will, and must, change in ways that might both shift its focus and, perhaps paradoxically, revitalise democracy and popular participation. In an age where algorithms are increasingly governing human action, there is a need to openly and carefully consider a) whether we as societies wish to allow the various forms of algorithmic governance and b) how such forms of governance lead to a need to discuss fundamental political institutions and arrangements. My main goal with this article is to highlight the implications of developments already underway. As Sadowski & Selinger (2014) point out, governments are increasingly opening the door for a) private companies solving public problems, and b) the use of AI in government. de Sousa et al. (2019) also shows how AI is now employed in all areas of governments, while recommending that implementation of AI in the public sector should be preceded by a 'debate with society' about such applications. AI creep -in the sense that the AI slowly and almost unnoticeably is applied to ever new areas -if left unchecked, might take us to a situation in which we have technocracy of AI without either political design or democratic input. I propose that we examine the key questions in advance, and prepare our societies for a future in which AI has the role we desire it to have in politics.
I will first develop the concept of traditional technocracy, in order to make sense of the logic and potential benefits involved, and the problems associated with traditional technocracies of humans. I then show how AI can, and in many cases already have, taken the role of a technocrat. Such use of AI is associated with somewhat similar, but not identical, risks, to those of traditional technocracy. The combination of the concept of technocracy and the use of AI lets me propose a set of objections to an AI technocracy based on the idea of political harms, and these are first presented and discussed before I move on to the discussion of the implications they have, and the potential for a legitimate and workable technocracy of AI.
2. Technocracy, expert rule and democracy Schumpeter (2003) argues that if we want results that are satisfactory to people at large, and judge our governments by such a criterion, government by the people 'would often fail to meet it'. For us to understand what an AI technocracy means, we must understand the basic ingredients of technocracy. Sadowski and Selinger (2014) are right to call for an improved understanding of what technocracy is, and they intend for their taxonomy to be a 'conversation starter' on the topic of technocratic uses of technology. While they argue that it is an ill-defined concept, I posit that the solution to this problem is to draw upon the existing literature on the subject form the field of political science. Furthermore, I reject their implicit position that technocracy is by definition a bad thing, associated with harms only, and no benefits. This is based on one of the fundamental points in this article, which is that democracy, and specific forms of democracy, cannot be assumed to be ultimate goals at the outset of an evaluation of the effects of democracy. If we start with such an assumption, anything that challenges democracy is by necessity bad. A more fruitful approach is to take an agnostic stance to the value of democracy. Firstly, this lets us perform an honest and open evaluation of alternatives. Secondly, it enables us to strengthen and revitalize democracy by showing why it is good, and not merely stating that it is.

Technocracy and expert rule
Technocracy is a term that in daily parlance refers to the rule of experts. More specifically, I use the term to refer to a state whereby experts in science and technology -not experts in politics -wield power. Technocracy is thus differentiated from other forms of rule by the few, such as epistocracy and meritocracy. Epistocracy is a form of government whereby those with a particular knowledge of politics rule, and is often considered a rule of the wise -those with knowledge of 'politics, history, economics' (Estlund, 2003(Estlund, , 2008. This idea is old, and famous political philosophers such as Plato (1968) and John Stuart Mill (1985) have proposed variations of it. Meritocracy, on the other hand, is often linked to mental and cognitive abilities, regardless of specific knowledge of a particular subject (Arrow, Bowles, Durlauf, 2000;Young, 2017).
Technocracy has been defined in various ways, one being Meynaud's (1969) description of 'a system of governance in which technically trained experts rule by virtue of their specialist knowledge and position in dominant political and economic institutions' (Meynaud, 1969). The experts, then, become technocrats, and have authority based on technical expertise (Bell, 1973). Putnam (1977) uses the word technocrat to describe proponents of technocracy, whilst I use it to describe technicians in power.
There are several possible arguments in favour of technocracy, and I split them into positive and negative arguments. The negative arguments mainly focus on the shortcomings of people and democracy -usually implying that the people are unfit to rule, or that democracy itself produces bad outcomes. The positive arguments chiefly focus on the benefits that could be gained from expert rule, e.g. efficiency, rationality and optimisation. It is worth noting that technocracy usually flourish in times of panic and great crises, as the need for efficiency and optimal policies in such times tend to be more pressing, and obvious (Sadowski & Selinger, 2014).
The negative arguments are mostly based on the idea that human beings are the main problem of politics, and that they are unfit to rule. People have personal interests, and they tend to pursue them, even when they are tasked with pursuing the public interest (Mueller, 2003). Furthermore, they have limited capabilities, in terms both of how much they can do and how well they can do things. Some of them have a greater capacity than others, however, and this is the foundation of elitist arguments for technocracy, for example as seen in Schumpeter (2003). Another class of negative arguments focuses on democracy as a system, as it is portrayed as inefficient, unjust and prone to instability.
The positive arguments are mainly based on achieving some form of rational optimisation in politics, and partly the belief that most problems, when properly understood, are technical problems amenable to the logic of statistical analysis and optimisation. However, there are also arguments in favour of non-democratic systems based on, for example, the idea that markets are the best way to organise everything. At least almost everything. And if we do need some form of government, it had better be small and kept out of the way (Hayek, 1960).
A point of great importance is that the technicians do not have all power in a technocracy. While a technocratic society delegates much authority to their technicians, they are not given the power to determine the goals of society. The technocrats are not political or moral experts, and I do not consider a situation in which morality and the question of political values are reduced to technical questions amenable to the calculus of mathematics and logic. This means that the technocracy I discuss have a political apparatus in place for providing the direction for society, while the technocrats have the authority to determine and effectuate the policies deemed necessary to proceed in the desired direction. This is akin to the process described by Peters (2014), where technical and political decisions are separated, and the technical issues are transferred to "independent agencies, bureaucracies and technocratic elites". Technical issues are also political, however, which is why I argue that we have a technocracy of AI when we delegate power to make technical decisions to computers. This is the argument made by Naess (1989), who urges us to see the political nature of all things technical and to submit the technical to an evaluation in light of our ultimate values. Fischer (1993) writes that technocracy revolves around the desire to use experts with specific knowledge in positions of political power, for the sake of building the 'good society'. This, then, is not the technocracy I discuss, as politicians -or the people -are charged with determining what the good society is, while the technocrats are charged with making it reality in the most effective manner possible. This is in line with Sadowski & Selinger's (2014) picture of technocrats as influential, but still largely controlled by elected politicians. This implies a break with the understanding of technocracy as an approach in which 'political realities play no role' (Janssen & Kuk, 2016).

Arguments against technocracy
Technocracy may have a bad reputation, but it is here given a blank slate and the chance to prove any worth it may have. While Sadowski & Selinger (2014) aim to arm us to 'combat technocracy', I claim that we must first understand if, and why, it should be combatted. Before moving on the use of AI in politics, I present the main traditional arguments against technocracy. However, as some of these arguments become less relevant if human technocrats are replaced by AI, I do not examine these in great detail, unless they remain relevant, and thus become part of the objection in section 4.
The arguments against technocracy are also based on various principles, and I briefly consider arguments based on the inherent values of democracy, human beings' needs and finally the inherent flaws of technocratic government itself. Estlund (2003) discusses objections to epistocracy, which are partly applicable to technocracy as well. The demographic objection is based on the fact that giving those with, say, university degrees more voting power means giving privilege to certain groups. The second objectionthe bias objection -is a version of the former, adding the claim that the privileged classes, races etc. will exert a certain bias that is unacceptable (Estlund, 2003;Lippert-Rasmussen, 2012). These arguments constitute the first class of arguments against technocracy, and I label them the technocrats are people, too arguments. One of the arguments in favour of technocracy is that people are flawed, but some are less flawed than others. No-one is free of flaws, however, and people can never be assumed to have totally overcome irrationality and bias. But can computers? I return to this issue in Section 5. It is also worth noting at the outset that problems of bias and discrimination are not exclusive to the use of algorithms, and in evaluating algorithmic governance, we must examine whether algorithmic bias is in fact worse than existing problems associated with human bias.
A second argument against technocracy is based on a concern for legitimacy. It posits that technocracy is bad because the process itself does not provide legitimacy (only participatory politics does so). This constitutes one of the objections discussed in Section 4.
Thirdly, people might object to technocracy because they believe a deliberative process will provide better outcomes. This particularly comes from the deliberative-democracy camp (Elster, 2005). Opposed to these arguments would be ones inspired by Janis (1982) and Hobbes (1946), who both emphasise the dangers inherent in democratic, or group, decisionmaking. It must be added that human technocrats are also liable to group-think.
The fourth argument against technocracy might be called a decentralisation argument. It is the Hayekian notion of dispersed knowledge (Hayek, 1960). No human being or small group of human beings can ever amass the kind of knowledge that each individual has about their own affairs. The only effective way to utilise such knowledge, Hayekians would argue, is to give power to the markets (Hayek, 1960).
The problems of technocracy are not insignificant, but as this examination shows, there is reason to ask if a technocracy of AI may in fact be better than a technocracy of human beings. This stems from the fact that several of these arguments more easily applies to human beings than to computers. However, as Janssen & Kuk (2016) shows, human faults are often refound in the workings of our algorithms, in the form of, for example, bias and discrimination.

The beginnings of a shallow defence of a technocracy of AI
The title of this article promises a shallow defence of a technocracy of AI. This refers to the argument I will construct in the first part of the article, which consists of three proposition that lead to the conclusion that we should employ AI in politics, and erect a technocracy of AI. This argument will then be subjected to a set of objections to such a technocracy in section 4. The first two components of the shallow defence follows form the preceding consideration on technocracy and politics.
First of all, politics can be understood as a process aimed at implementing the best possible policies. But what policy is considered best can, of course, only be decided once we know what criteria to apply. Thus, the first and fundamental purpose of politics is to develop and elucidate what fundamental moral values a society is based on. Only then can politics as we know it in the day-to-day workings of society take place. And only then can we properly assess the value of technology (Naess, 1989). This is the first building block in the defence of an AI technocracy: Policies should be evaluated on the basis of the fundamental moral values of the society in question, and finding these values is the first purpose of politics.
Furthermore, if politics revolves around the question of finding the best policies, as Schumpeter (1946) implies, we also have a second premise in the establishment of what will become the defence: The best policies in accordance with the evaluation discussed in the first premise should be implemented.

Artificial intelligence and political decision-making
Artificial intelligence is superior to human intelligence for analysis of large and complex problems involving the need for strategy, prediction of long-term effects and analysis of vast amounts of data. One example is playing chess and Go (Campbell et al., 2002;Kurzweil, 2015;Chouard, 2016;Google, 2018). However, AI can beat humans at more than fun and games. Politics is complex, and it involves many considerations of notoriously uncertain short-and long-term consequences. As the literature on algorithmic governance shows, AI is already employed in many facets of government, as I show in section 3.1.
I will note at the outset that the potential benefits from AI in politics is heavily contested (Katzenbach & Ulbricht, 2019). I agree with the idea that AI as of today is not some silver bullet that can be employed in order to create flourishing societies. However, as the many existing applications clearly show, there is undeniably a potential for the beneficial use of AI in politics. I will assume that some of this potential can be realised, and that we will continue to see certain improvements in the technologies involved. This means that I consider possible near-future technologies, and not speculative developments related to some form of singularity, etc. (Kurzweil, 2015).
I will not discuss the technologies involved in great detail, but rely on the assumption that much of the future applications of AI is based on the technologies currently employed. De Sousa et al. (2019) chart the technologies most often employed in public sector AI; various forms of machine learning are used, while artificial neural networks (ANN) is the most popular technique in terms of usage. Artificial neural networks (ANNs) imitate biological information-processing systems, and consist of artificial neurons that transmit signals to each other in reinforcement systems, for example (Bishop, 2006). Deep learning is our term for deeply layered neural networks, and this is the machine learning approach used by the AI economist I use as an example, as well as by Google's AlphaGo and AlphaZero (Sutton and Barto, 2018;Zheng et al., 2020;Google, 2020aGoogle, , 2020b.

AI and expertise in questions of policy
Algorithmic governance is a term describing the utilisation of algorithms both in ordering human action in general, and in traditional political structures (Danaher et al., 2017;Katzenbach & Ulbricht, 2019). As such, it is a very broad term, and as discussed in the introduction, I only focus on a subset of what is referred to in general as algorithmic governance. Using the taxonomic tools of Sadowski & Selinger (2014), I limit my discussion to the use of AI in the domain of government, and mainly to a situation in which AI is given the authority to govern by mandates (not more informal nudging, or even less direct governing through technological mediation).
According to the same typology, I mainly focus on a specific set of harms, labelled political harms. These are distinguished from existential and discursive harms (Sadowski & Selinger, 2014). Existential harms relate to the reduction of individuals through the datafication and behaviouristic approach to the control of human beings. Discursive harms involve the restriction of discourse through a domination by technocratic ideas. In this article, I do not focus on harms related to privacy and surveillance or bias, unfairness and inequality. A potential loss of freedom is another potential harm that falls beyond the scope of the article (Danaher, forthcoming).
Issues of privacy and surveillance are of great importance, but I will in the following assume that AI can be built on data which is not individualised, and that it does not govern by individualized nudges or micro-directives, etc., based on detailed personality profiles (Saetra, 2019;Ma, 2020). Hypothetically, policy could also be based on machine learning in combination with simulations, such as in the example of the AI economist (Zheng et al., 2020). Related to issues of privacy are issues of agency, autonomy and freedom (Katzenbach & Ulbricht, 2019;Danaher, forthcoming). Another point is that increased used of data, and the translation of all aspects of human action into data is that such quantification by necessity involves multiple stages of human interpretation, challenging the objective veneer of big databased decision making (Janssen & Kuk, 2016;Saetra, 2018;Ma, 2020). Such issues are only partially considered in this article, and will thus remain important philosophical issues for further research.
Some of the potential harms that I do not focus on, fall in the categories of unforeseen and unintended consequences (Katzenbach & Ulbricht, 2019). One such problem is algorithmic bias and resulting issues of fairness and inequality (de Sousa et al., 2019). Furthermore, algorithms, partly because of problems associated with the training data, could also favour particular political ideologies, and they could reinforce discriminatory and other undesirable practices (Janssen & Kuk, 2016). These problems are real, and important, and any technocracy of AI must implement mechanisms of algorithmic oversight in order to address such problems. However, I do not consider these challenges in detail in the present article, and employing the thought experiment presented in section 3.2 enables me to isolate the political harms I do focus on from the issues of algorithmic bias.
According to Sadowski and Selinger (2014), political harms relate to people being disenfranchised and deprived of political power and influence. This relates to the potential depoliticisation effects of algorithmic governance (Katzenbach & Ulbricht, 2019). In addition to this, questions of transparency and procedure are relevant to the analysis of political harms (Danaher, forthcoming). These concerns are examined in more detail in section 4.
With regard to the applications of AI, I posit that we have already moved a long way towards AI decision making, and this is an important point to keep in mind when we consider the technocracy of AI. As I will show, a condemnation of the technocracy of AI that I propose will in many respects simultaneously involve a condemnation of many current practices.
AI is now used in all areas of society: driving cars, guiding our missiles, trading stocks, helping us navigate new places, playing chess, recognising speech, speaking to us, finding our mates, forecasting the weather, etc. (Danaher, 2016). AI is even used to determine who gets bail and who gets loans and identify likely tax evaders (Tashea, 2018;Danaher, 2016). All these applications of AI are relevant to an examination of algorithmic governance in general, but not all of them are in the domain of government.
de Sousa et al. (2019) provide a thorough account of current applications of AI in the public sector. They find that it is employed in all areas of government already, while general public service, economic affairs and environmental protection is the areas most emphasized in the literature on AI in government. Examples of current applications are found in most functions of the public sector, such as public health, transportation, education, security, communications, and the actual examples involve, amongst others, systems of identification of risks, optimization systems, systems for response, analysis and prediction (de Sousa et al., 2019). Predictive policing is another area that has gotten a lot of attention, and AI is also used in, for example, the administration of immigration and in calculating social benefits (Katzenbach & Ulbricht, 2019;Janssen & Kuk, 2016) Algorithmic control of traffic lights and speed limits is one low level example in the domain of transportation, where the utility of having dynamic systems based on machine learning could increase efficiency and utility (Ma, 2020). If the objective of our traffic lights and speed limits is to a) optimise traffic flow while b) minimising casualties, AI could most likely improve a static system of pre-determined speed limits -allowing for increased speed in low danger situations and not having drivers wait unnecessarily at red lights. The idea is that similar benefits could be reaped in other domains of law and politics, such as tax policies (Zheng, 2020). Another example would be the use of AI to implement optimal strategies for pandemic response, which could involve using AI to predict and identify outbreaks and dynamically apply the optimal response based on a society's tolerance for risks weighed against other factors, such as economic costs.
In general, when we face optimisation problems, AI could have the potential to aid us. Most problems of politics are highly complicated optimisation problems. As long as we have a goal, AI can in theory take us there in an effective manner. Furthermore, it could continuously experiment, and could update its policies. Policies would be dynamic, always responding to changes in human preferences and behaviour. One recent example is the AI economist which can make tax policies, in a simulation, that both improve productivity and decrease inequality (Zheng et al., 2020). Another is the use of WeBuildAI, which through participatory algorithmic design managed to increase efficiency of a food donation transportation service while simultaneously improving the perceived fairness of the outcome -adjudicating equity and efficiency (Lee et al., 2019). Such examples are limited and hypothetical, but, in combination with the many existing applications, they suggest that we could in time find us in a situation where such systems are applicable to real-world large-scale situations.

The magical decision box and whether or not to use it
One way to approach this issue without getting lost in the technical aspects of AI is to consider an analogy. I invite you to replace AI as we know it with a hypothetical device called the magical decision box (MDB). This thought experiment requires you to accept that a society has discovered a box -origins unknown -that makes decisions. The box accepts input and constraints, it accepts goals and it provides answers. Through rigorous tests the society in question has found that the box significantly outperforms human experts in making policies that optimise the goals it has been given. Thus far they have found no major errors. Despite their very best efforts, they have not been able to understand how the box works, and they cannot recreate (even in retrospect) the ways in which the box reasons. The decisions, however, are good, so should they employ it?
The MDB is, of course, different from AI in many ways. However, there are certain similarities that might elucidate the core issues involved in the objections to AI in politics. Like the MDB, AI is a 'black box' that we do not always fully understand. Even if we understand the principles involved in making advanced machine learning models, they will produce outcomes that we often cannot predict (Janssen & Kuk, 2016). I will later return to the notion of explainable, or explicable, AI -also referred to as XAI (Gunning, 2017;Robbins, 2019).
The realist theory of democracy and both cognitive and behavioural science have shown many of the shortcomings of human decision-making. This brings me to the next proposition in the argument leading to the preliminary conclusion that artificial intelligence should be used to improve political decisions. This could come about in several ways, ranging from computers aiding human decision-makers to their replacing them. I here examine the latter possibility, which I call an AI technocracy.
We now have the option of letting computers make expert judgements for us. With the use of advanced machine learning, computers could make better decisions than human beings. Vogl et al. (2019) argue that computers might improve decision-making by overcoming the human limitations of bounded rationality and information processing. Machines are better than us at chess and Go, and if this ability also applies to other well-defined strategic problems, should we not let computers decide for us? If we had the MDB, and knew that its results were good, would we use it?
3.2.1. The third premise -AI supremacy The possibility of AI superiority in solving complex issues related to human beings and their societies is not out of the question. As my goal is to analyse the possible future impact of AI on politics, the exact probability of this happening, or the exact scope of AI superiority, is of less importance. For these reasons, we can consider AI to be the MDB for our present purposes. What matters is that for now we should accept the premise that AI might in the future play the part I am proposing here. If this does not happen, or until it happens, this paper can be considered null and void, and no harm should have occurred. To be clear, I do not believe AI can currently be the basis of a full technocracy such as the one I here describe.

Artificial Intelligence is better than humans at finding and enacting the best policies in certain areas concerning science, engineering and complex societal and macroeconomic issues
With this, the shallow defence of an AI technocracy is complete. Firstly, the purpose of politics is to find the best policies, in accordance with some set moral basis. Secondly, we should enact the best policies we have identified. Thirdly, AI is better than human beings at creating and identifying the best policies. All this leads to the preliminary conclusion that AI should be given the power to discover and enact our policies.

The potential arguments against a AI technocracy
It seems non-controversial to assume that such a technocracy could in certain areas give us better decisions in terms of efficiency, utility or whatever other criteria we set as the computer's goal. Lee et al. (2019) and Zheng (2020) provides proofs of concept for the idea that AI could provide both efficiency and still be amenable to popular control and direction.
However, there are also several potential downsides, which will have to be weighed up against these benefits. I will here provide a schematic overview of some of the obvious objections to an AI technocracy, based on the political harms it may cause. They are: a) people might need politics, b) legitimacy is linked to democracy, c) AI is not capable of morality, d) we have an issue with transparency related to AI and e) AI decision-making involves problems assigning responsibility.
Note that I am not considering the objection that such a technocracy would not lead to better policies, as this very premise is part of the argument I am testing. If this premise fails, the argument in favour of technocracy crumbles, and there is little need for any of these objections.
Accepting the first three premises for now, I proceed to the political theoretical objections against the shallow argument in favour of an AI technocracy. In this brief examination of the objections I will first present each objection, and then some considerations concerning its possible shortcomings.

Human nature and homo politicus
According to Aristotle (1981), humans are political animals. The best that they can achieve is to exercise their political capabilities and live a life of self-determination in a political society. Only then will they realise their true nature -their telos.

People need full political participation in order to be satisfied
This is, however, only one of the interpretations of Aristotle's view of humans and politics. Mulgan (1990) shows that Aristotle can also be understood as supporting 'the withdrawal from politics, or at least reluctant participation'. In both The Politics and Ethics the philosopher's lifestyle is portrayed as being superior to the life of the statesman (Mulgan, 1990, p. 195).
Another immediate objection would be to ask: Are people politically active in today's political systems? What will really change if AI makes our decisions, instead of bureaucrats? The 'political animal' argument is against any move from direct democracy towards democracy by representation or republicanism, and this move could be said to be a necessary evil if we are to have large-scale, complex societies.
This objection also relies on an agreement with Aristotle's hierarchical ordering of human activities. Other philosophies, such as hedonism, could easily portray a life of increased wealth and leisure as a potential improvement to a life material hardship and political activity. Philosophers such as F. A.
Hayek have proposed that involvement in politics is not that important, and that economic liberty is what is essential to good societies (Hayek, 1960). Following this argument, he states that democracy is not necessarily the best way to rule, and that '... a democracy may well wield totalitarian powers, and it is conceivable that an authoritarian government may act on liberal principles' (Hayek, 1960).
However, a crucial point of this article is that a technocracy of AI may in fact lead to a revitalization of popular participation and the accessibility of the political domain. Through new ways of participation, such as algorithmic co-creation, we could find ourselves in a situation in which most people have, and experience, more political power than they do today. This is examined in more detail in the next objection.

Legitimacy and democracy
Another objection is that people need to be involved, and that they need deliberation to take place if political systems are to have legitimacy (Rousseau, 1968;Manin, 1987). Even if the best decisions were made by a dictator, his decisions would not be legitimate if the citizens had no political power. We thus distinguish between legitimate authority and effective authority. AI might be extremely effective, but might still lack legitimacy.
We can see this objection as consisting of two parts: a) policy should be decided by the populace, and b) all citizens should have equal political rights and a realistic chance of exercising them. The objection to technocracy would be that legitimacy is lost when the population is deprived of crucial political rights and the ability to take part in decision-making. This objection is based on a procedural view of legitimacy (Peter, 2017). According to such a view, due process is the source of legitimacy, and deliberation is one aspect of such a process (Manin, 1987).

The second objection -legitimacy based on participation
People will not deem a government in which they do not participate to be legitimate The counterobjections are similar to those regarding the previous objection. We could first argue that an AI technocracy makes people more equal than they would have been in any democracy -equally free to participate in the moral debates about the direction of society, and equally deprived of participation in policy formation. We might also argue that such a system would enable people to understand more about politics, and that the barrier to understanding politics and taking part would be lowered. When human politics only revolves around fundamental moral values, it is easier to both understand and take part in this process.
Furthermore, not everyone agrees that legitimacy is gained through participation. For Hobbes (1946), legitimacy is achieved through the social contract, which ensures that a governmentby one, few, many or all -is created to satisfy everyone's fundamental need, this need being security -sometimes labelled liberty (Manin, 1987). König (2019) even discuss how algorithmic governance bears some resemblance to Hobbes's apolitical Leviathan. If democracy contributes to better political outcomes, it is conducive to legitimacy. If it does not, it is not necessary (Raz, 1994). The latter argument is the instrumental view of democracy.
Even in today's societies participation in politics is restricted to a select few, despite everyone having a theoretical chance of taking part. The size and complexity of our societies have led to the growth of bureaucracies, expert rule and indirect democracy. Going from this to an AI technocracy is perhaps not as radical as it might seem at first sight.
One obvious potential benefit of digitalisation and the increased capacity to communicate using digital tools is the possibility of directly involving far more people in political processes. The Pirate Party is an example of an actual attempt at this, and they have done very well in several countries (Beyer, 2014).
If a technocracy of AI is to overcome this objection, participation in some form must be ensured. One way to implement this is to promote co-creation and 'democratizing algorithm' (Janssen & Kuk, 2016). One such example is the WeBuildAI example, which shows several methods of popular participation in the creation and design of algorithms, and how this could lead to both increased participation and increased algorithmic awareness (Lee et al., 2019). This also addresses a potential objection based on the idea that only democracy has the pedagogical effects ensuing from participation in politics. People learn from taking part in politics (Rousseau, 1968;Tocqueville, 2004). According to this argument, the beneficial effects participation has on people could outweigh the negative effects of having sub-optimal policies. However, if systems such as WeBuildAI allows for participation and efficiency, there is little reason to choose sub-optimality.

Machines, morality and human wellbeing
Another possible counterargument against the use of AI in politics is based on the argument that politics always involves questions of values and morality. This is not necessarily a problem per se, but it creates an obstacle for AI as soon as we run into people arguing that morality is solely a human endeavour. According to some people moral machines do not exist, and human qualities such as compassion and wisdom are crucial prerequisites for important decisions (Weizenbaum, 1976).
Resolving this question entails coming to some sort of agreement about how we define morality, and whether or not our definitions exclude machines. What we are concerned with here is whether or not machines can act morally and make moral decisions. In the terminology of ethics: whether or not they can be moral agents (Nolt, 2015).

Computers should not make decisions that affect people's lives and wellbeing
Some people believe that we can train a machine to become moral, e.g. by having a lot of people answer various questions on the morally superior decision in various instances of the trolley problem. The Moral Machine is one such endeavour that has attracted much attention (Awad et al., 2018). It is an attempt to crowd-source the morality of people around the world in order to teach a computer to distinguish right from wrong, in order to find ethical principles to guide machine behaviour. This is an approach which could be related to the approach of WeBuildAI, in which social choice theory and popular participation is used to train algorithms (Lee et al, 2019). Kurzweil (2015) predicts that machines will gain human proficiencies and capabilities, including morality-related emotions. Maldonato and Valerio (2018) and Scheutz (2016) also discuss related topics, e.g. machines with value systems and the creation of machines with moral competences. Even if they do not gain anything akin to human morality, machines are becoming better at understanding human morality, as it can quantify, track and compare moral bias, for example, 'across cultures and over time' (Schramowski et al., 2020).
However, we could also approach this objection more pragmatically. If we decide that decisions that affect people's wellbeing require moral capabilities, we are already employing AI in situations where it does not belong. Self-driving cars could not then be allowed. Robots trading in markets would become highly problematic. Use of AI to determine whether or not criminal offenders should receive bail would be highly dubious.
AI is already involved in moral decisions in countless ways, and our current approach to this issue is to make someone other than the computer itself morally responsible for the decisions made. We will include the premise in the tentative discussion, but I will argue in the conclusion that this objection is better understood by way of handling objection 5 and the attribution of AI decisions.
Another obvious counterobjection would be that if machines cannot be moral, neither can they be immoral. To some people this situation would appear to be a huge advantage compared with the capacities of today's politicians. If indeed power corrupts, we should see it as an advantage that AI has no morally corruptible nature (Dalberg-Acton, 1907).
A different kind of objection would be that the important issue is not morality but whether or not machines are able to act ethically. If morality is right action based on some internal value system, then ethics are actions based on some external system. Machines' ability to act in accordance with humans' codes of ethics might be sufficient. Following this objection, one could also question the nature of human morality itself, and portraying morality as some vague and unscientific idea, plagued by no consensus, which is thus of little interest. We might even object that we have no test for the morality of our human politicians, so why should we implement one in order to keep robots away from power?

Transparency and the problem of understanding what we cannot produce ourselves
Human experts can explain their choices to the populace in a way that computers cannot -at least not yet. And even if they become better at explaining why certain choices are made, will we be able to understand? The reason for giving computers the authority to decide is that they are able to make evaluations and calculations that are beyond human capacity.
Transparency would then be an issue, and we might be left guessing why our AI technocrat made the decisions it made, in the same way as chess experts guess why the computer calls moves good or bad when analysing the world's best players. Santiago (2019) emphasises that humans must have the last say in decisions made by AI. However, if we cannot understand the reasoning behind the decisions, transparency and control are hard to achieve (Janssen & Kuk, 2016). Opacity is one of Danaher's (2016) main objections to algorithmic governance.

AI is not transparent and thus not fully amenable to human control
Pressure is mounting, Wachter et al. (2017) state, to make sure that AI becomes explainable, and thus transparent. Explainable AI (XAI) is the quest for machines that can make us understand their decisions and reasoning. Such a development could make us trust, understand and manage AI decision makers (Gunning, 2017). This is a field that receives a lot of attention, but until we have achieved this goal we could imagine other ways of dealing with the problem.
Like with the magical decision box, which is by definition unfathomable, we must perhaps rely on control by goal achievement and observable result parameters with AI technocracy. If AI or the MDB make decisions that take us closer to our goals, do we really need to understand how they manage to do so? Similarly, if the bureaucrats make decisions beyond politicians' understanding, do we complain? We might simply recognise that the very reason politicians rely on experts, or AI technocrats, is because the latter make decisions that are unfathomable, owing to their being too advanced. Lack of understandability is the price for better decisions. In the words of Robbins (2019), 'a principle of explicability for AI makes the use of AI redundant'. Janssen & Kuk (2016) is right in arguing that a full understanding of how complex algorithms operate is most likely restricted to the 'happy few'. However, so is a full understanding of how many current systems of government operates.
Danaher (2016) does not see this counterobjection as a valid reason for ignoring transparency issues. The fact that we have similar problems today, he argues, is not a reason to allow them (in more serious form) in algorithmic governance. I am, however, not arguing that an AI technocracy is an ideal system. I merely argue that it could be argued to be better than what we currently have. As such, the pragmatic counterarguments pointing out that an AI technocracy does not lead to a worsening compared to current affairs is of a great interest. If AI technocracy leads to a pareto improvement where efficiency is improved, while other aspects of political life remain the same, I will consider this desirable.
A related point is that opacity could provide certain benefits. If the exact workings of a system is not known, it is difficult to game the system (Janssen & Kuk, 2016). However, the AI economist of Zheng et al. (2020) allows for such gaming of the tax system, and find their policy maker capable of coping with it.

Who is responsible when a computer makes decisions?
The final objection is based on the problem of accountability. If a computer makes a decision, who is held accountable for this decision? Until we are willing to consider robots legal persons with rights and obligations, someone else must be held accountable for their actions. But who? Their authors, or perhaps their owners, who are responsible for using them?
The accountability gap is a term used to describe such issues, and it is a problem that becomes increasingly prominent as our technologies become more advanced and capable (Wachter et al., 2017). This complexity creates situations in which the creator of a machine cannot foresee the actions of his creation (Matthias, 2004). When this happens, some perceieve a gap between who acts (for example, a complex machine), and who is perceived as responsible (the creator).

Accountability regarding the consequences of political decisions must be clear, and it becomes less clear when AI makes decisions
Santiago (2019) states that when businesses employ AI in decision-making, humans must have the final say. But which humans, when we are discussing AI technocracy? We are not discussing a political system in which bureaucracy is simply replaced by AI, but one in which the technocrat machines are given greater autonomy in terms of policy formation. Thus, the current system, whereby politicians are responsible for the decisions of their bureaucracies, will not work, as they will not be assumed to be in control of their technocrats as they now are of their bureaucrats.
We might not have any perfect solutions to this problem, but one might also argue that neither are popular modern political systems perfect. We do, however, have several options. One is to hold the producers of AI legally responsible for their artificial technocrats.
Another possibility could be to create a human council for machine control, tasked with making sure our technocrats do as we wish. Algorithmic 'constituents', Binns (2018) argues, have a right to scrutinise and hold someone accountable for the exercise of algorithmic governance. As Janssen & Kuk (2016) point out, few are able to exercise such control.
Educating the population and organising such a council for machine control is a vital precondition for an AI technocracy, as I will return to later.

The pros and cons of an AI technocracy
The shallow argument in favour of an AI technocracy is based on three premises. Where, then, does this lead us? Firstly, if O1 or O2 is true, the argument in favour of technocracy is weakened. We might, however, introduce a counterobjection, to remove some of the sting of these two objections. If politics is really about deciding which moral values are important, and what we see as the good society, having a policy process whereby human beings take part in the formulation of these issues is compatible with a technocracy.
I will even argue that it is necessary in order to avoid a technocracy with no goal or direction. If this counterobjection is accepted O1 and O2 can be disregarded. However, O1 states that people require full participation, and this condition might not be satisfied by my proposed policy process involving only fundamental values. However, objection 1 could conceivable also strengthen the initial argument in defence of an AI technocracy. While Danaher (2016) argues that algorithmic governance will tend to limit and reduce the possibilities of participation, my argument in favour of reorienting the goal of politics to fundamental values could make it more accessible.
O3 is an objection that sounds both true and reasonable, but let us consider the implications before accepting it. If we take it as it stands, AI must not be used in any area where human wellbeing is affected -not in determining loans, not in determining bail, and not even in helping us identify criminals, or in areas such as financial fraud. Furthermore, positive use of AI affects people's wellbeing just as much as use that we more often perceive to be problematic. This means that autonomous vehicles are most certainly out of the question, and even brake assist etc. must be got rid of.
The implications are absurd, which implies that we must subject the objection to close scrutiny. Some use of technology is unavoidable if we are not to reject all the assistance it currently provides us with. When we pursue this issue, we find that all technologies have moral and political implications, which means that the issue at hand is not really restricted to the use of AI. The important issue, I argue, is accountability, which will be discussed shortly.
The proposition that AI should not make decisions because the decisions are not transparent (O4) is a universal problem associated with letting those with the most knowledge make decisions. It is, however, possible to imagine a situation whereby decisions are explained by virtue of the results they lead to. How economists arrive at a new and better tax system is perhaps of less interest than their proposal showing what a new tax system will lead to. Politicians will not understand the technical aspects of the reasoning, just as they will not understand the processes that lead the AI from input to ideal policies. This problem is not restricted to the use of AI.
I argue that AI is transparent, but that we simply have a hard time understanding how known methods and known input lead to decisions that are better than those we could make ourselves. If nothing is hidden or secret, a decision is by definition transparent, and whilst further work on explainable AI will be beneficial, Objection 4 does not appear to be particularly strong.
This issue of transparency is a problem in two respects. It is a challenge to the legitimacy of an AI technocracy, and people might not be willing to take the risk involved in being ruled by something one cannot understand. With human politicians we realise we may be taking a degree of risk, as human fallibility is well known. It is, however, harder to identify the risks involved in computer decisions, and the fear that it will err in ways we will not be able to uncover, or correct, is different from the risk of being ruled by fallible leaders who are just like us. Furthermore, if we develop and apply methods of participation and co-creation, we could increase general algorithmic awareness. This might hypothetically lead to most people having a better understanding of how our societies function than they do today.
Accountability issues are the fifth objection, and I have proposed that either a council for oversight or the producers of the AI systems could be held accountable. It is not a problem per se if we refuse to give AI moral agency, and thus moral responsibility and accountability. An AI technocracy need not be unlimited and arbitrary. It could be a constitutionally limited AI whereby a council -small or large -of men is tasked with giving the AI both a direction and a goal.

Discussion
AI is in certain respects like a modern-day version of Leibniz's God -the good architect who must be assumed to have created the best possible version of the universe (Leibniz, 1989). Had we known what God knows, we would realise that the universe is as good as it can be, despite occurrences of what we perceive as evil. The argument works as long as you believe in an almighty and good God, and this kind of trust must thus be transferred to AI.
This takes us back to the magical decision box. If we knew it had always produced good decisions in the past, should we employ it, even if we did not understand how it worked? We would be in a situation similar to the one Leibniz considers when examining whether or not the world is good. God is good, and God created it, thus the world must be good, even if it may not always seem that way to us.
If God created the magical decision box, I argue that we should employ it as our technocrat. However, as my analysis suggests, we should perhaps not discard the use of AI as our technocrat even if it is not a magical and perfect decision box. There are, however preconditions that must be satisfied for such a technocracy to be able to overcome the objections I have presented. Furthermore, the defence of a technocracy of AI here presented is shallow also in the sense that not all objections and considerations have been considered. The future capabilities of AI are somewhat speculative, I have not gone into detail on issues of bias, discrimination, agency, and liberty, and it remains to be seen if the world of wicked political problems is truly amenable to the optimising hand of artificial intelligence (Head, 2008). Nevertheless, I do propose a defence that is not without value, as it suggests that the political harms of an AI technocracy are not insurmountable.
One of the preconditions for a technocracy of AI is that there are backup and control mechanisms in place, in case the AI systems should fail, or if we desire to scale back the use of AI. Control systems are necessary in order to supervise the performance of the systems, and particularly to monitor potential issues related to bias, discrimination, and various other unintended consequences. Furthermore, times of crises and rapid change have shown that machine learning systems can easily become confounded, and in such cases, there must be mechanisms in place for human intervention. This was shown clearly during the Covid-19 situation that led to problems for many machine learning models trained on 'normal behaviour' (Heaven, 2020).
Putting AI in power involves, in a certain sense, putting the companies and individuals that create these systems in power (Danaher, 2016). Some people might argue that the creators are simply creating algorithms -just doing maths and creating abstract and valueless systems. Dwork & Mulligan (2013), however, point out that any such system is bound to reflect some of the implicit ideas and values of its creators. Whilst this might occur by pure accident, despite the intentions of the designers, Gillespie (2010) points out that most of the creators of such systems have clear profit-maximisation goals. Such goals and accompanying incentives might make it difficult for us to regard AI as an equivalent to Leibniz's God. Griffy-Brown et al. (2018) notes the profit-driven technocratic ideology, which is different from the disinterested and scientific technocratic ideal I previously described. These considerations show that a robust system of algorithmic supervision is required whenever the AI in question is a human creation, and not a magical decision box.
Another precondition is some form of political control of the core values and direction of society. One possibility is to simply scale back, but largely keep, the current system of representative democracy. This involves a limitation of the scope of politics, as the technocrats have authority over many issues of policy. Another, and more radical possibility, is to pursue some of the new modes of popular participation and popular government provided by new technologies.
the shaping of algorithms and the potential of increasing popular understanding of AI (Lee et al., 2019). This latter option, while more radical, also shows how a technocracy of AI could in fact increase people's opportunities for political participation, rather than the opposite. While doing so, it could also meet the demands for involving various stakeholders and social groups in political decision making. If this is not enough, increasing broad participation in political also allows us to reap the benefits of dispersed knowledge and decentralised markets, as emphasised by Hayek (1960).

Conclusion
The benefits of better political decisions are considerable. I have shown that AI is already being employed in the public sector in order to improve political analysis and decisions, and that it will become ever more capable. If we agree that politics should have good policy as a goal, and if we do not consider our current democratic order as a given and inviolable good, we start opening the door to the idea that AI should have political power.
In this article I have focused on the potential widespread application of AI in the domain of government, and I have examined an extreme case in which AI is given the authority to form policy. This is done through systems characterised by low transparency, high complexity and high degrees of automation. As such, the situation is rigged to trigger the alarms of those concerned with democracy, participation and the political harms of algorithmic governance.
I have chosen to focus mainly on a set of harms perceived as vital from a political science stance, while acknowledging that other important harms are both important and necessary for acquiring a complete picture. However, all the harms from and objection to an AI technocracy cannot not be discussed in the required detail in one article. As such, my account is a partial one which must be seen in combination with detailed accounts of the harms related to bias, discrimination, autonomy, etc.
However, the political harms are vital, and not as often discussed in detail as some of the other harms, which makes this an important part of a full understanding of the consequences of algorithmic governance. Through five objections, based on participation, legitimacy, nonmorality of machines, transparency and accountability, I have reached the conclusion that an AI technocracy is not derailed by such objections. If, that is, certain preconditions are met.
An AI technocracy capable of overcoming the objections discussed must involve some way of making sure that humans retain the power to define the fundamental goal of politics -that of deciding what our fundamental values are, and how we imagine the good society. If humans retain this role, and make sure that the process is a democratic one, both participation and legitimacy might be preserved. Furthermore, I have argued that the non-morality of machines is not a problem if we assign responsibility either to the proposed democratic council of values and AI control or to the producers of AI.
AI is in some respects like the magical decision box. If we had such a device, we could employ it, and over time we would might to trust that its decisions were good -like Leibniz's God. If the MDB were a mystical device, not one created by humans, there would be, I argue, no reason for any particular group of people to object to its rule. While there are many reasons to be wary of an AI technocracy of human origins, it is time we take the possibility seriously. Firstly, because we are already seeing the application of AI in many areas of the public sector, without a full and open debate about the desirability and implications of such a development. Secondly, we must take the possibility seriously, as I have shown that it has the potential to revitalise democracy, increasing public participation in politics, and provide more efficient outcomes.