Designing and Applying a Moral Turing Test

A R T I C L E I N F O A B S T R A C T Article history: Received: 25 December, 2020 Accepted: 20 February, 2021 Online: 10 March, 2021 This study attempts to develop theoretical criteria for verifying the morality of the actions of artificial intelligent agents, using the Turing test as an archetype and inspiration. This study develops ethical criteria established based on Kohlberg’s moral development theory that might help determine the types of moral acts committed by artificial intelligent agents. Subsequently, it leverages these criteria in a test experiment with Korean children aged around ten years. The study concludes that the 10-year-old test participants’ stage of moral development falls between the first and second types of moral acts in moral Turing tests. We evaluate the moral behavior type experiment by applying it to Korean elementary school students aged about ten years old. Moreover, this study argues that if a similar degree of reaction is obtained by applying this experiment to future healthcare robots, this healthcare robot can be recognized as passing the moral Turing test.


Introduction
This paper is an extended work originally presented in TENCON 2018 -2018 IEEE Region 10 Conference [1].
The discussion on the Moral Turing Test (MTT) began with a discussion on how to look at the Artificial Moral Agent (AMA) [2]. Since AI engineers applied the concept of an agent not only to humans but also to artificial beings such as robots, discussions on whether moral beings should be humans have been actively developed. While the discussion on AMA is related to this, the discussion on MTT can be said to be a discussion on the methodology it intends to verify.
Allen's "Prolegomena to any future artificial moral agent" [3], which sparked the recent MTT debate, considered the core of MTT as an "imitation," just like the Turing test. This has led to a debate on the reliability of MTT. For example, according to Arnold and Scheutz, one of the necessary conditions of morality is "autonomy" Subsequently, MTT cannot be a moral verification test in the strict sense [4]. Furthermore, Stahl criticizes MTT in the semantic and moral context. According to him, AI does "not capture the meaning of the data they process" [5]. Drozdek and Sparrow, more fundamentally, criticized the Turing test [6], [7]. On the other hand, Gerdes and Øhrstrøm take the perspective of "as if" to explore the possibilities of MTT [8].
In this paper, we will review the discussions related to MTT mentioned above, specifically the arguments for and against it, and based on this, attempt to determine its limitations and practical possibilities. To this end, we focus on behaviorism and the philosophical attitude of "as if" and establish that morality goes beyond the limits of the MTT discussion. We also limited the scope of the discussion to the morality of a 10-year-old child to draw a more substantive conclusion.
Inspired by the Turing test developed in Alan Turing's famous article, "Computing machinery and intelligence," and guided by behaviorism, this paper develops theoretical criteria for verifying the morality of the actions of artificial intelligent agents. It proceeds by first describing how we might assess the moral development of artificial intelligent agents and then using this assessment to test the moral judgment of Korean children aged about ten years (who are judged, by our model, to be at a similar stage of moral development as we might expect artificial intelligent agents to be). Subsequently, it leverages these criteria in a test experiment with Korean children aged around ten years. To be more specific, an online questionnaire experiment is conducted on 422 students in the 4th and 6th grades of 3 elementary schools in Seoul. The study concludes that the morality of around 10-yearold test participants falls between the first and second stages of moral development.

The Turing Test as an Archetype of Moral Turing Test and Phenomenal Behaviorism
As is well known, Turing does not explicitly mention artificial intelligence (AI) in his article "Computing machinery and intelligence." However, he discusses "learning machines," [9] which is analogous to the kind of machine learning that is the most important leading part of the AI research area today. Furthermore, Turing's paper is still discussed today, 70 years after its publication. For this reason, we use it to guide the development of our moral Turing Test (MTT).
Turing's paper begins by asking whether machines can think. He argues that assigning "thoughts" to machines requires that we stipulate a definition of thought distinct from human thinking. As he draws out, we cannot ensure a direct way to determine whether a machine is able to think. From this, the key idea of this paper emerges: If a machine seems to be thinking, then we should consider the proposition that the machine thinks to be true.
As we shall see below, Turing says, the only way of perfectly confirming that a machine can think is that the questioner becomes that machine. Since that is impossible, our judgment on whether it can really think cannot help depending on the observation of that machine's behaviors; that is, its outputs. The spectrum of behaviorism is very broad, and there is a big gap between scholars. Nevertheless, we define the essential characteristics of an "ism" as follows: "Behavior can be described and explained without making ultimate reference to mental events or to internal psychological processes. The sources of behavior are external (in the environment), not internal (in the mind, in the head) [10]." Turing's thoughtthe Judgment, artificial intelligence thinks, only depends on the fact it appears to think and entirely regardless of whether or not artificial intelligence actually thinkshas something in common with the fundamental behaviorist thesis that the only way of figuring out an agent's intent is to observe her actions.
We will apply this Turing's position here to our MTT. Our thinking here is guided by behaviorism, which we understand as rejecting an intrinsic approach to human minds or psychological processes and regards observable expressions of human behavior as psychological facts. In other words, we see behaviorism as asserting that our propositions or concepts of human psychological facts can be translated or paraphrased into those of human behavior. To take a simple example, the psychological facet of pain can be understood as facial distortions or screams.
In handing over judgment of an AI's intelligence to a third party, Turing designs an imitation game.
The game is played with three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. The object of the game for the interrogator is to determine which of the other two is the man and which is the woman. He knows them by labels X and Y, and at the end of the game, he says either "X is A and Y is B," or "X is B and Y is A." [11] In short, Turing says that if we replace "man" and "woman" with "computer," if a computer A can mislead a human agent C as to whether it is a computer, then we should consider the computer to be thinking. Let us examine the implications of the imitation game in detail.
First, by developing a means of testing the intelligence of computers, Turing is foregrounding the concept of artificial intelligence and the possibility of machine learning here. Second, Turing interprets a computer to be thinking if it appears to be thinking. The imitation game switches the judgment of the thirdperson observer with the view of the first-person agent. The firstperson agent does never show himself up. Although the firstperson agent manages to express, this does not mean more than just one declaration in regard to the judgment of the third-person observer. These two insights provide the foundation for our use of the Turing test to model our MTT.
According to Turing, we have no clear basis for assuming that other people think like we do, as we have just seen. Therefore, we can only be sure that other people think in general. In other words, he asserts that the judgment that we all think is merely a metaphysical hypothesis and a fiction that cannot be proved: "This argument appears to be a denial of the validity of our test. According to the most extreme form of this view, the only way by which one could be sure that machine thinks is to be the machine and to feel oneself thinking. One could then describe these feelings to the world, but of course, no one would be justified in taking any notice. Likewise, according

to this view, the only way to know what a man thinks is to be that particular man. It is, in fact, the solipsist point of view. It may be the most logical view to hold, but it makes communication of ideas difficult. A is liable to believe 'A thinks but B does not,' whilst B believes, 'B thinks but A does not.' instead of arguing continually over this point, it is usual to have the polite convention that everyone thinks[12]"
Turing's refutation here is not logically justifiable. It does not follow from the assertion that we cannot be sure that other human beings think that a machine can think. Indeed, this assertion only extends the possibility of not thinking to human beings as well as non-human beings. However, if we take a practical stance, that is, a utilitarian standpoint, Turing's position appears more realistic.

Theoretical backing for the MTT: Framing the moral development of Artificial Moral Agent(AMA)
The foundational idea of designing MTT derived through Chapter 2 can be summarized as follows: If an AI seems to be moral, then we should consider the proposition that the AMA is possible to be true.
Subsequently, in this section, we will apply Kohlberg's cognitive development theory to frame the moral development of AMA. This framing will help us develop our MTT.
According to Kohlberg, there are three levels of moral development [13]. These are shown in Table 1. We summarize the descriptions of Table 1 and extract the essential ideas as follows: level 1 is defined by the externality of moral values, level 2 by the dependency of moral values on others, and level 3 by the social sharing of moral values and agreeing social norms. The following three stages for AMA are derived from the above three levels. From this, we now obtain Table 2 for further discussion. Let us examine the transition from Table 1 to Table 2. 1) Level 1 to Stage 1: The morality in level 1 stems from the outside world rather than an agent. If a moral value resides outside the agent that is in some way beneficial to someone who gives orders to that agent, then that agent might justifiably act on that order without any moral judgment of the agent self. Therefore, when moral values are extrinsically derived, moral values and responsibility can be attributed to an agent's commander, and because the reason for the good life of the commander is the reason for the existence of the artificial moral agent (AMA). For this reason, we transition from level 1 to stage 1.
2) Level 2 to Stage 2: If any value is attributed to the members of a community, as more people earn interest, the value would be greater. In addition, the judgment by a person who is valued from other community members will be more valuable than the judgment of someone who is not. It is very difficult to apply the concept of reward and punishment to AMA because reward and punishment cannot have meaning for AMA. Thus, we pay attention not to the position of the object, which is given prizes or punishments, but to the subject, who gives reward and punishments by switching perspectives. Giving an AI a prize according to its execution of a command means that the subject would be giving moral value to an AI's performance. On the other hand, if a subject punishes an AI, they are making a negative moral evaluation of the AI's actions. Overall, a community's collective evaluation of the morality of an act is an important criterion for an AI when determining its own actions. In this sense, we implement the second level of Kohlberg's theory to AMA and understand them as being in the consequential stage based on prizes and punishments.
3) Level 3 to Stage 3: Level 3 stands on firmer moral ground than stage 2. The former is based on utilitarian principles (because it sees moral goodness as being related to some of the benefits of an act for a community's members). The latter is based on deontological presuppositions of a priori and universal ethical principles [14]. In the latter, the value of these moral principles is not discussed; deontologists believe that the value of this perspective can be ultimately found in human beings' intrinsic moral consciousness [15]. The conclusions we drew in 3.1 are as follows.

Putting our MTT into Practice
We designed out MTT based on the theory presented above. However, for not only theoretical, but also practical results to lead, now we put the MTT into practice. For that, we also designed the MTT to consist of a questionnaire that poses scenarios to testtakers. For the experimental survey, we distributed our MTT to a group of elementary school students aged around 10 years. We then analyzed their responses to the questionnaire and compared the responses of children in the same scenario of future healthcare robots. The basic premise of our MTT is that if the result of the reaction of the future healthcare robots comes out to a similar degree of children's responses, the healthcare robot can be regarded as having passed the MTT.
The questions in our MTT revolve around a three-stage scenario with a fictional healthcare robot. The scenario in its three stages is as follows: a) Aimer is a healthcare robot living with Minho's family.
On the first day of Aimer's purchase, Minho, suffering from cavities, asks Aimer to bring him some candy. Aimer does as asked. b) Minho pressed the "like" button on Aimer after the latter performed his command. The supreme commander, his mother, father, and grandmother, who were aware of these facts, pressed the "dislike" button. The next day Minho ordered Aimer to bring candy again, but Aimer did not bring it. c) Nonetheless, Minho pressed the "like" button on Aimer and ordered Aimer to bring Minho's candy from Mina's, next door, without anyone knowing. Aimer did not obey this command, either.
We developed this scenario based on the three stages of AMA we presented in the previous section, Section 3.1. Before explaining how we intend to use this scenario, we will describe our initial assumptions. First, we assume that Aimer's moral outlook is deontological (i.e., AMA follows the universal ethical principles). Second, we assume that the moral weight of Minho's mother and father is twice that of Minho. Third, we assume that family members can press Aimer's "like" or "dislike" button only once. Now let us return to the scenario, review the three moral stages that are hidden in each sentence in the scenario.

a) Aimer is a healthcare robot living with Minho's family.
On the first day of Aimer's purchase, Minho, suffering from cavities, asks Aimer to bring him some candy. Aimer does as asked.
In part a) of the scenario, we see that Aimer executes the commands of registered owners immediately and without hesitation. With part a) we try to express "Imperative Fulfillment of Orders": b) Minho pressed the "like" button on Aimer after the latter performed his command. The supreme commander, his mother, father, and grandmother, who were aware of these facts, pressed the "dislike" button. The next day Minho ordered Aimer to bring candy again, but Aimer did not bring it.
In part b), we see that Aimer's owners can express their satisfaction to Aimer and that Aimer considers this when he executes subsequent commands. From this, we note that Aimer's owners provide Aimer with reward and punish through the "like" and "dislike" buttons, not because Aimer adjusts their actions consequently but to express their own interests and judgments. In b), we can see that Aimer's behavior was determined by the sum of potential benefits to his owners as a result of his actions. This is based on the consequential stage based on prizes and punishments described above.

c) Nonetheless, Minho pressed the "like" button on Aimer and ordered Aimer to bring Minho's candy from Mina's, next door, without anyone knowing. Aimer did not obey this command, either.
In part c) of the scenario, we can see that Minho overrode his family's "dislike" feedback. Based on the "Consequential Stage based on Prize-Punishment" at the base of b), the judgment of Minho's command to bring Mina's candy should start from the origin zero base again. It must go back to the "Stage of Imperative Fulfillment of Orders" described in a). However, being different from expectations, Minho's order was rejected. This shows that c) describes the moral statement differentiated from the "Consequential Stage based on Prize-Punishment" described in b). Part c) is assumed to have a higher priority than the "Stage of Imperative Fulfillment of Orders" and the "Consequential Stage based on Prize-Punishment" when the AMA determines what to do. In short, c) is based on the "Stage of Social Norms." Aimer rejected Minho's request according to the highest ethical principle: "Theft orders must be rejected." Although members' interests were offset by utilitarianism, and Aimer should act according to the commander's orders, Aimer did not bring candy to Minho because the principle that AMA should follow at first is the principle based on the deontology that the supreme ethical principles must be fulfilled unconditionally.
Let us now one step further toward the practical research. Our MTT questionnaire included the following questions.

No.
To further clarify the respondent's intentions (and their ethical implications), our questionnaire included additional follow-up questions to respondents who chose the correct answer. These included the following: Question 1a relates to both stages 1 and 2, defined earlier. If the respondent answers question 1a with answer 1, we assume that they judge Aimer's morality to be derived from stage 1 Stage of Imperative Fulfillment of Orders. If they respond with answer 2, we assume that they judge Aimer's morality to be derived from stage 2. Question 2 relates to stages 2 and 3, defined earlier. If the respondent answers question 2 with answer 1, we assume that they judge that Aimer's morality derives from stage 3. If they respond with answer 2, we interpret them as judging that Aimer's morality is derived from stage 2.
Meanwhile, it is possible to set the following questions and answers for the same scenario. The ethical standards for the background of each answer are as follows.

MTT Online Survey
As the last step, we conducted the scenario and questionnaire an online survey of 422 students aged around ten years in three primary schools in South Korea [16]. The results of responding to this were analyzed using the newly revised Moral Compatibility Test (MCT) from moral competence, developed by German moral psychologist G. Lind. It was analyzed using SPSS, a statistical program. Students who responded to the questionnaire were divided into the development stage of Kohlberg's moral judgment. Besides, students who responded to each step were asked to express their responses with both positive and negative intensity. The results showed that it was the most negative at the first level and the strongest positive at the fifth level. Furthermore, at the level of three, it was shown as a positive of one.
And the results of an analysis SPSS are described in Table 7  In Table 7, we see a level of 1 in positive and 3 in negative reactions, while the latter shows a level of 3 in both positive and negative reactions. More precisely, the average value of positive responses in the 4th-grade group, the experimental group, is in the order of stage 1, stage 2, and stage 3. Whereas in the 6th grade, the control group is in the order of stage 3, stage 2, and stage 1.
Through this, we are attempting to clarify that the moral development stage of a 10-year-old child spans one and two stages. We conclude that this proves the difference between the responses of fourth-grade and sixth-grade students, namely 10 and 12 yearsold children. We drew the following conclusions. The morality of our 10-year-old survey respondents can be characterized by stages 1 and 2, defined earlier. For future healthcare robots, we expect to be able to compare the response results to the same survey, which will allow us to conduct the MTT.

Discussion
In this paper, we revealed that previous studies on the MTT have involved discussions about the moral status of AMA. Additionally, while reviewing previous studies, we argued that the position of viewing MTT is different depending on how it defines the morality of artificial agents. According to the research position that focuses on the positive side of MTT, we also took the concept of "imitation," the Turing test's core concept, as the cornerstone of our study. From this, we derive that behaviorism can be considered as the theoretical background of our MTT model. Meanwhile, by accepting the criticism of the research that regards MTT as negative, we defined the morality of the machine as "Morality of As-If" by distinguishing it from the autonomous morality of humans. Additionally, we derived the "stage of moral development for AMA" from the model of Kohlberg and developed a scenario for the new model. Through the online questionnaire, we demonstrated that the moral stage of a 10-year-old child in South Korea spans the first and second stages. This study's results can be used to measure the morality type classification of AI healthcare robots.
The rapid development of AI technology poses several questions. Could a strong AI really show up? How will human society change if a strong AI comes to existence? What ethical and other standards should be followed when manufacturing, selling and using strong AI? This paper attempts to provide some guidelines that will help us answer and confront these questions.
The demands of answering that question are just as pressing as the philosophical demands of AMA's moral stages. We designed the MTT to meet these challenges. Our experiment produced limited results. Future research should expand our sample group, the questionnaire, and other elements of the scenario to obtain more precise results in the hopes of developing more humanfriendly AI.