Is Gender-Neutral AI the Correct Solution to Gender Bias? Using Speech-Based Conversational Agents

First, the overall status of speech-based conversational agents, such as devices and service types, average usage time, and main tasks, was identified. In addition, the voice of the agent was characterized by 1) tone, 2) color, and 3) appearance of gender, and investigated in detail.

The voice tone (1) was identified by dividing it into three physical representative features of sound and three types of tone. In the study of Feigen (1971), loudness and pitch denoted the physical characteristics of sound, and also speed is a factor to consider. The specificity of sound, i.e., timbre, suggests the mood of the sound, such as the attitude and emotion of the speaker, noting that it depends on the complexity (Feigen, 1971). The voice tone defined in this study refers to the physical characteristics of each sound pressure or the mood of the sound at several frequencies. As presented in Table 1, specific voice tone characteristics can be investigated through six parameters for the two significant categories of voice tone.

Table 1

Details of the voice tone

We defined voice color (2) as the voice characterization of the agent and voice tone. The voice color defined in this study refers to the degree of intimacy between the listener and the psychological distance (cold color) and close (warm color). Table 2 lists the specific voice color characteristics. Eight detailed items can be studied for the two significant categories of voice colors.

Table 2

Details of the voice color

Voice gender (3) was inferred using the voice tone and color as defined above. In this study, we described the voice gender using gender dichotomy(Male or Female). We expected that specific descriptions can be given using the surveyed voice tone and color.

Second, we checked the gender expected by users in the voice of speech-based conversational agents. Reich-Stiebert and Eyssel (2017) investigated the voice of gender for each task to specifically describe gender bias. In the survey, six task lists were formulated for speech-based conversational agents (play and search music, search weather information, search traffic information, search the internet, receive and send phone calls, and operate TV functions). The six task lists, mainly performed by speech-based conversational agents, were formulated, and it was found that “music playback and search” was the task most performed by speech-based conversational agents. In the process of utilizing agents for the task most performed by speech-based conversational agents(music playback and search), we tried to identify the gender of the agent that users prefer, feel stable, and are satisfied with.

In the survey, we provided video data containing the process of “music playback and search” between users and agents. In this video, a user requested a task via text as shown in Figure 2. The agent who received the command was represented by the voice of ‘NAVER CLOVA Dubbing¹⁾’ and it progressed the conversation. The female and male voices were that of “Ara Joy” and “Minsang,” respectively.

Figure 2

Study 1: A capture of video for second survey (Scene that user request a task via text and scene that AI Speaker response via the voice of NAVER CLOVA Dubbing from left image).

We checked significant differences in preferences and stability between genders (male/female) according to functional performance (positive/negative performance). In addition, the degree of functional performance (positive/negative performance) was investigated through satisfaction to confirm that it had a significant effect on the experience. We conducted a pre-survey to determine if the participants were correctly aware of the gender of the provided voice. Only participants who recognized the gender were used to collect the evaluation data. Table 3 compares the voices of male and female gender agents who performed the “music playback and search” task positively or negatively (2×2 within-subject design).

Table 3

Study 1: Research structure model (function performance x voice gender)

According to this survey, users used speech-based conversational agents present in smartphone services such as Galaxy Bixby (34.78%) and iPhone SIRI (23.91%) at least once a day (40.00%). The agents mostly performed tasks such as searching for weather information (22.08%), music playback (20.78%), and schedule management functions, such as timer (18.18%). Most participants (67.57%) answered that they were real users who used speech-based conversational agents in their daily lives, but now they have no need for agents (41.43%) and have stopped using them. Also, non-users confirmed music playback and search (30.86%) and weather information search (25.93%). In the results of these data, the participants (regardless of actual or non-user) want the speech-based conversational agents to perform simple and daily tasks such as “music” and “weather search” instead of them.

To find the answer to “What voice characteristics do users prefer in AI agents for performing a task?”, we set the focus on the voice tone and color. First, as shown in Table 4, the voice tone was investigated considering five parameters (loudness, pitch, speed, attitude, and emotional state). The results were classified according to a 6-point scale, wherein weak, low, slow, commanding, negative, and cynical tones denoted one point and strong, high, fast, convincing, positive, and optimistic denoted six points.

Table 4

Study 1: Category of voice tone

Participants, on average, indicated “strong (M = 3.20, SD = 0.87)” and “high (M = 3.34, SD = 0.97)” for the voice tone (n = 72) among the existing speech-based conversational agent, and the speaking speed was “difficult to see fast (M = 2.77 and SD = 0.84).” Most participants stated that the voice of the agent was “convincing (M = 3.6, SD = 1.03),” “positive (M = 4.20, SD = 1.05),” and “optimistic (M = 3.51, SD = 1.07)”; the three items demonstrated the individual difference of opinion.

Based on the voice tone of the existing agent, we examined the difference in expectations regarding the agent’s voice tone in real/non-user groups. We used one-way ANOVA to determine if a significant difference was present between the existing agent voice tone and the agent voice tone expected by real/non-users for each of the six items. Bonferroni’s post hoc test was used to check for differences. No significant difference was found in terms of strength, pitch, two attitudes, and emotions. However, there was a significant difference between the voice tone of the existing agent (M = 3.23, SD = 0.73) and that expected by actual users (M = 3.85, SD = 0.67) only in the “slow-fast” items (F(2, 52.36) = 5.05, p < .05). The users considered the “speed” at which the speech-based conversational agent delivered information and determined the necessary changes.

In addition to these results, “I do not use AI when I am in a hurry. I guess the AI agent is slow. (P7),” “The car navigation voice has become too slow and mild since its recent update. Because I have to concentrate on the road, I wish to feel a little more determined and emphasized. As shown in (P6),” It can be inferred that the speech-based conversational agent as a “conversator who needs to wait because the response speed is slow” is fixed. “It does not feel like ping pong like talking to people, so I think it is a little slow. However, this is not very reliable. (P2).” As such, it is difficult to believe that any task performed by a speech-based conversational agent, a “conversator who feels insufficient due to the slow response speed,” so we found that speech-based conversational agents find it difficult to believe that they will perform any tasks and this point should be improved.

Second, we studied the voice color considering eight elements of speech-based conversational agents (cold and hard, artificial and businesslike, intelligent, clear, calm and stable, soft and Gentle, friendly and kind, and cheerful and bright). Table 5 classifies the eight items in the form of a spectrum from the upper “cold” tag to the lower “warm” tag. We collected data through duplicate selection for eight items and determined the results.

Table 5

Study 1: Category and tag of voice color

Table 6 summarizes the ranked results, which can be expressed in the lower voice color. These lower voice color results denoted items selected by participants at least twice as many times as other upper color values. Compared to the voice color of the actual speech-based conversational agent, what participants expect is overlapping items such as “friendly and kind” and “calm and stable.” Considering only these results for agents, unlike cold voice colors, “it felt like a human feeling (P2).” It provided a warm and convenient voice color.

Table 6

Study 1: Results for comparison between the actual voice color and the expectant color of speech-based conversational agents

However, the “warm” convenience voice color does not just mean “optimistic.” For the agent voice color expected by real users, “clear” and “cheerful and bright” received the highest rank. This implies that actual users want a voice color that seems to perform functions “more reliably and clearly” than the existing agent. For the agent’s voice color expected by non-users, it is like the actual agent’s voice color, but “cheerful and bright” are ranked at the top together. “I think the voice provided to the existing voice service is a bright image. I do not think I have ever felt much trust, even though it was easy to hear (P10)” and “Because the existing products or services are optimistic, I did not have much confidence. As can be seen in (P4)”; thus, it can be inferred that non-users who do not currently use the agent want the agent voice color to be “intelligent and clear.” In other words, all users expect a trustworthy agent that can provide precise answers and efficiently perform the requested tasks.

Combining the results of the expected tone and color of the speech-based conversational agent, participants displayed minimal confidence in the ability of the agent to perform a task efficiently owing to slow reaction speed. Thus, it can infer only the routine and simple tasks that it would expect to perform explicitly, such as playing/searching for music and providing weather information (see Section 4.5.1). The follow-up interviews confirmed that limited awareness and use of speech-based conversational agents are responsible for the gender assigned by users to AI agents. “I feel that a woman’s voice has a light tone, so I do not think there is any resistance to requesting such simple tasks. (P1),” and “I feel that the female voice conveys a lot better, so I keep using it for everyday tasks. After all, AI is only used to find music or weather. (P6)”; accordingly, we found that the specific gender of a “female” voice and “belief” about task performance was mentioned together.

According to the survey, users continuously used the voice of the speech-based conversational agent without changing the initially set “ female” voice (82.86%). In contrast, some participants said, “if the voice is too mechanical, it bothers me. Choose a voice that is close to what a person says. (P8),” and “When I heard the alarm, the high notes (specific to women) bothered me. (P7)” Similar to P7, there was an opinion that participants were concerned about the initial set-up “female” voice owing to the somewhat artificial characteristics of the voice (17.14%). In addition, some participants questioned the limited and consistent gender of the agent, such as “Because I only use a gentle female voice (P25).” In other words, although there is a psychological basis for the voice characteristics and gender of the speech-based conversational agents, the participants did not take any action to change the initially set “female” voice. Also, non-users also answered that they recognized speech-based conversational agents as women (81.08%), so we can infer that most users give female gender to agents.

To further check why users do not change the initially set “female” voice of the speech-based conversational agents, for each task, we investigated the expected gender of agents (Table 7). The survey helped to determine the gender of the speech-based conversational agent that performed the six tasks (music playback and search, weather information search, traffic information search, internet search, phone reception and call, and TV power).

Table 7

Study 1: The expectant gender considering the tasks of speech-based conversational agents(n=72)

A male and the gender indistinguishable between males and females (gender neutrality) can be compared with females by integrating into the ‘other’ gender. Consequently, we found that in all tasks except searching for traffic information for real users, more than half of real and non-users said that they expected a female voice. To determine if there is a significant difference in the “traffic information search” task for actual users with relatively large “other” gender values, an independent two-sample t-test was conducted to confirm the difference in results between actual and non-users based on tasks. No significant difference was found between the actual/non-users in all six tasks (p > .05). In other words, we confirmed that all participants generally expect a female voice of the speech-based conversational agent regardless of the task.

We investigated how the expectant gender eventually affects user experience. Participants observed the interaction between the user and the agent performing the task requested by the user through the video. Based on their experience in performing these tasks, participants evaluated each gender in terms of preference, stability, and satisfaction.

Table 8 summarizes the results of the frequency distribution analysis of three items. It can be checked that the “female” agents received a high experience of task performance for positive/negative functions. Also, regardless of the experience of task performance, participants answered that they felt more preferred and stable when using the “female” agents. The items of “satisfaction” were investigated on a 5-point scale to obtain results on the overall experience of using speech-based conversational agents. We used the paired samples test to investigate the difference in satisfaction with the experience of task performance(positive/negative). The results showed that satisfaction with experience performing positively (M = 3.80, SD = 0.82) was significantly higher than that with experience performing adverse (M = 2.70, SD = 0.88; t(39) = 6.58, p < .001). In other words, the user experience was positive when a female speech-based conversational agent performed a positive task performance.

Table 8

Study 1: Evaluation of voice gender by the experience of task performance

The speech-based conversational agent with gender-neutral voice “G” was produced as the voice of “NAVER CLOVA Voice.” According to the website of Naver CLOVA Voice, they provide the hybrid form of Unit-selection and Deep Neural Network technology, providing the best synthetic voice in various domains. They provide a hybrid form of unit selection and deep neural network technology, providing the best synthetic speech in various domains. These synthetic voices in various areas are classified into contexts such as roles and occupations rather than gender. We could find “gender-neutral” voices that were not distinguished as male or female, and among these voices that included “Diversity.” As mentioned in the study by Feigen (1971), the voice tone of the test tool “G” can be classified as the most accessible pitch element to distinguish physically. We divided the gender-neutral voices of “G” into voice tones (high pitch or low pitch) and colors (warm or cold), as shown in Table 9, so that participants can intuitively evaluate different gender-neutral voices. Four versions of the gender-neutral voice of “G” with low or high pitch and cold or warm color were produced, compared, and evaluated (2×2 within-subject design).

Table 9

Study 2: Details of the test tool's voice

We controlled the two features (reasonably fast response speed and precise task performance) most expected by users to evaluate the four versions of the gender-neutral agent.

First, the test tool “G” adjusted the a) response speed and b) delivery speed to maintain a moderately fast conversation pace with the participant. When a participant requested to reduce the speed at which the agent responds, a Wizard of OZ test was designed, a format in which the experimenter manipulates the agent who responds to the user without the user knowing it. Because the test evaluates participants who believe that they are testing actual speech-based conversational agent prototypes, we can expect the evaluation data to remain unchanged. After the evaluation, only participants who answered through contextual interviews, i.e., “I thought/believed that I was talking to an actual AI agent,” were used as evaluation data. Moreover, because there was a result that the delivery speed for the performance result felt somewhat slow (see Section 4.5.2), the speech speed selected for the study was adjusted slightly faster. Likewise, only participants who answered that the agent speed was “appropriate” or “not a problem” used these evaluation data.

Second, the test process was controlled as a single scenario to ensure that “G” provides more definite answers to the participant's requests. The scenario task was selected as the most preferred task (schedule management and weather and traffic information search) when performed by another gender voice, not female, in the first study. By synthesizing the three tasks, we constructed a conversational scenario between the participants and agents to check the schedule and current weather before leaving the house for an appointment and receiving recommendations for transportation and routes to the appointed place. As presented in Table 10, participants can request or respond to the agent by selecting one of the four responses in each scenario and naturally moving on to the next step to continue the request or respond to the agent's response. After the evaluation, only the participants who answered “the story flow of the scenario was natural” or “it did not interfere with achieving the real goal of the test” through contextual interviews were used as evaluation data.

Table 10

Study 2: Test scenario

Thus, for usability evaluation of speech-based conversational agents of gender-neutral voices, we produced “G,” an agent of four gender-neutral voices, adjusted in tone (high or low pitch) and color (warm or cold). In addition, we matched the “speed” that will affect usability in the non-gender testing process to the user's expectation level. Finally, we have controlled user scenarios with “tasks” that gender-neutral voices are familiar with performing from the findings of Study 1. Based on these study designs, we tried to confirm the usability of gender-neutral speech-based conversational agents in terms of 1) preference, 2) stability, and 3) satisfaction (Table 11).

Table 11

Study 2: Test scenario

Participants commonly said, “It feels neutral overall. (P10),” and “It is not easy to distinguish between gender.” (P2), thus confirming that G is a gender-neutral AI agent, which was manufactured for this test and did not encounter any problems during evaluation. We attempted to understand the overall voice image perceived by the participants, including tone, color, and gender, for the four neutral-gender voices of G. Accordingly, a 4 min graph was used for the x-axis, divided into (left) low and (right) high pitch, and the y-axis into (left) warm and (right) cold. Each denominator was divided into 7 × 7 compartments so that the participants could express the relative difference in voice in more detail. Figure 3 shows the results of using a graph with voice tone pitch and color (warm-cold) as an axis and displays the corresponding graph denominator and the exact position as participants listen to each version of the voice of G.

Figure 3

Study 2: Voice keyword for four versions of "G" (G -ver1, ver2, ver4, and ver3 clockwise from the left-top graph). The number of participants displayed on each scale is written; the larger the number, the darker the background color of the scale.

We studied voice keywords using a semantic differential scale consisting of eight parameters (loudness, pitch, speed, attitude (command-advocative, negative-positive), emotion, color, and gender). We arranged the keywords of each item in the opposite manner and checked the results using a 5-point scale at both extremes. Each scale ranged from -2 to +2, and we calculated the score for each item. The left-side items indicated negative attributes, such as weak, low, slow, imperative, pessimistic, cynical tone, and cold color; the right-side items indicated positive attributes, such as strong, high-pitched, fast, persuasive, positive, optimistic tone, and warm color. We calculated the average scores for each of the eight items. The corresponding version of the voice keywords is displayed on the number line in Figure 4.

Figure 4

Study 2: Average of voice keyword for four versions of "G."

Through Figure 4, it was difficult to grasp the tone and color patterns of the version 1, version 2, and version 4 voices of the participant's thoughts. In addition, the voice keyword scales of version 1, version 2, and version 4 voices were also close to zero, making it difficult to identify distinct features. However, we confirm that, unlike the other three versions, the tone and color patterns of version 3 speech are relatively dense in one spot. In addition, it was confirmed that the scale of the voice keyword was characterized by +12.50.

In the follow-up interview conducted to review the usage experience of the four versions of G, we attempted to investigate how the participants perceived the identity of each version. Table 13 presents a collection of adjective modifiers that express the perception of each version. The adjective “neutral” appeared at least twice for all four voices. However, there were slight differences in the adjectives of each version’s list (Table 13). Version 1 voice received the highest number of neutral expressions; however, it contained more harmful or cold color expressions than other versions. For version 2 voice, several general expressions such as “the degree to which it was relatively indistinguishable”; it is expressed like “youth boys” were used. Similarly, the version 3 voice was expressed as “young” and “cheerful.” In contrast, the version 4 voice was perceived to have a negative or colder expression than versions 2 and 3, but it was confirmed to be relatively “calm.”

Table 13

Study 2: A collection of adjectives for four versions of "G"

In summary, participants recognized each version of G’s voice as follows: Most participants noted that Version 1 voices were “cold” and “distinguished” and answered that a voice tone and color matched the “machine.” For this reason, it was confirmed throughout the interview that participants described Version 1 voice as the most “neutral” voice. In addition, the version 3 voice goes beyond expressing a high tone and warm color, and specific characteristics such as “age” are described.

We confirm G’s usability (preference, stability, and satisfaction) with four versions of neutral voice applied through observations of the overall testing process and responses from follow-up interviews.

• Preference

Most participants answered the following when asked to select their preferred voice among the four versions of G. “It is hard to choose because it is ambiguous to distinguish them all, and it does not sound as comfortable as the existing voice. (P1)”. Conversely, we have confirmed the opinion of which of the four versions of voice would be the least preferred when using a real speech-based conversational agent.

We confirm that version 1 and version 4 are the least preferred. We received feedback regarding the version 1 voice: “It was creepy, but it felt like a neutral voice among the four. (P7)”; “Because the concepts of women and men were ambiguous, as the version 1 item felt that way, the other three voices felt relatively like boys. (P10).” Similar opinions to P10 prevailed. Participants chose Version 1 as their least preferred speech based on these opinions. We received feedback regarding the version 4 voice: “In fact, version 4 was not memorable, but I thought it might be frustrating because of the low tone. (P3)”; “Version 4 was not preferred in other respects compared to version 1. Version 1 was a creepy voice mixed with the voices of the two men and women, and Version 4 was uncomfortable to talk to. (P7)” In other words, we can infer that version 1 represents a human aspect. Also, we can infer that version 4 represents a human aspect, but no unique self-characteristics could be ascertained. As a result, in terms of “preference,” participants feel G negatively with a neutral voice and have not formed any meaningful connection or influence between the agent and the participant.

• Stability

During the testing process, participants confirmed that they would continuously use the agent’s voice when it felt “stable.” We analyzed why participants chose Version 3 and Version 2 among the four versions of G as ‘stable’ and ‘trusted’ voices through the following feedback. “It's the fastest one. I thought it was that natural because it was done quickly and adequately, as a real person said. (P8),” “It felt similar to my voice, so I felt strange and friendly right away. (P1),” “It felt like a smart lower-grade elementary school kid was talking. (P4)” Participants express natural, stable, and reliable when somewhat like the characteristics of a voice they know, such as the speed of most humans speak, my familiar voice, and the voice of an elementary school student (Table 13). In other words, participants mentioned the possibility of smooth communication with the agent only when they judged that the characteristics and conversation method of human voice were applied among versions of G with a neutral voice in terms of “stability.”

•Satisfaction

We confirmed that among the versions of G, the participants were most satisfied with the version 2 voice. Participants mentioned version 2 voices that felt like human objects due to their “clear identity” rather than “neutral that felt like a machine.” In addition, participants evaluated Version 2 Voice as “the most preferred voice (preference side)” and, at the same time, “similar to my voice (stability side).” Through this, we confirmed that the participants felt a certain degree of “satisfaction” with the voice that fully felt the previous “preference”and “stability.”

In other words, if the “preference” and “stability” are not satisfied with the neutral voice, it also means that participants have difficulty feeling “satisfaction” with the gender-neutral voice. Through the previous results, the “preference” and “stability” side, participants generally did not prefer G's neutral voice. In addition, the participants said they felt stable in G's gender-neutral voice only when it has the characteristics and conversation method of the human voice. Therefore, it is difficult to say that the participants were satisfied with the four neutral voices of G.

In Study 1, we were able to find the answer to Study Question 1. Through Study 1, we confirm that users of current speech-based conversational agents perform simple tasks mainly utilizing “female” voices. In addition, we confirmed that they wanted “female” voices in various voice tones and colors and that they expected “female” voices regardless of their experience in performing tasks or tasks.

In Study 2, we were able to find the answer to Research Question 2. To enable participants to evaluate “gender-neutral” voices consistently, we utilized four neutral voice “G”s with different tones and colors as test tools. Through study 2, we confirmed that users of speech-based conversational agents did not prefer “gender-neutral” voices with ambiguous identities and no human nature. In addition, we confirmed that users could feel stable regardless of gender if it is a voice similar to humans or a conversation method. Finally, we confirmed that the user could only be satisfied with the “gender-neutral” voice when this preference and stability were satisfied.

In study 2, which evaluates the use of four neutral voice versions of G, all participants first provided answers for the keyword “gender.” “I don't know why, but I was the first one to check men and women.” Based on P10, it is evident that the participants attempted to unconsciously classify the gender of agent G. This reaffirms the results of Costa and Ribas (2019), who identified gender according to the agreed criteria of gender intelligibility for smooth interaction in the study by Butler (1990). Only gender (male or female) in dichotomy classification, which is the existing gender rule, is a significant factor, owing to which, AI agents are perceived as humans. In other words, it suggests that rejecting the expectant gender from speech-based conversational agents may lead to negative results between human and AI interactions.

By exhibiting gender distinction, the participants also presented essential rejection reactions such as “neutral voices are difficult and unfamiliar to distinguish. (P1)” and “the first neutral voice seems to be more awkward than other voices. (P8)” However, if a user has sufficient experience with AI agents, despite the ambiguity of gender distinction, the agent can be identified using factors other than gender. “First, I thought (all versions of the voices) were difficult to recognize as an object, but I thought that just the way of speaking and atmosphere, like the items on the evaluation paper, could function well enough. (P12)” Even if an AI agent is ambiguous in terms of gender, if the voice tone and voice color are evident in the tone, then users can be expected to perceive the agent as a conversation partner. Considering the long-term usage time, even if the impact of the gender of AI is small, the impact of characteristic factors such as conversation method, content and intuitive distance can help in understanding the identity of the AI.

Eight out of the 12 participants from study 2 answered that if they considered the currently used AI agents, including speech-based conversational agents, to be gender-neutral, the agents would be perceived as an object to be respected with honorifics instead of a classic woman. There was an opinion that neutral AI agents would be significantly more intelligent and reliable than before. Moreover, “if neutrality is applied, I think it will go well with the word artificial intelligence. (P4)” There was another opinion that the neutrality of the machine would be well utilized.

However, the opposition sharply opposed. Regardless of the gender applied to the AI agent, some participants unconsciously sought after feminine features or functions (expectant gender). “Unless devices such as AI speakers are dramatically developed considering high-dimensional functions in areas such as tone and speech, or even so, I do not think we can change our existing image. Rather, I think a more elite female secretary will come to mind. (P11)” In other words, applying “gender-neutral” instead raises the opinion that applying the existing “female” can be more effective for usability. However, this does not break away from AI agent design that only applies specific gender (female) but can instead continue “gender bias.” Thus, it can harm human-AI interaction by recklessly rejecting users’ expectations of the gender (Expectant Gender) of AI agents, including speech-based conversational agents.

Eleven out of the 12 participants from study 2 stated that the existing AI agents, including speech-based conversational agents, learn the bias inherent to humans. They agreed that if humans continue to use AI agents without further guidance for future development, awareness regarding gender cannot be raised and the gender bias may be intensified. However, as mentioned in Section 6.2.1, gender is a significant aspect for AI to be recognized as a conversational partner and it cannot be removed recklessly. In addition, it has been pointed out that the use of gender-neutral AI agents for reducing the gender bias may have adverse effects (refer to Section 6.2.2). How should we design to eliminate/minimize the “gender bias” of applying only one gender to AI, which is increasingly used in everyday life as an essential tool for humans?

The gender assigned to AI can be distinguished from that of humans and viewed as a third gender (yet unnamed). “I think the neutrality of the machine and humans is different. If I consider it as same area, I think the confusion will be too great. It is necessary to consider and develop differently AI gender from human gender standards. (P10)” Considering the results of this research, AI designers suggest that it is important to take measures to develop a third gender for AI that is different from the classification of human gender. As mentioned in Section 6.2.1, it is necessary to reinforce characteristic elements, such as conversation method, content and intuitive distance. Strengthening these characteristic factors implies that users can replace and minimize the time taken to consider whether the gender of an AI is female or male, thus eventually reducing the influence of the existing human gender classification on AI identity. In other words, in AI, efforts are needed to develop genderless AI that focuses on roles and contexts rather than gender boundaries from a human perspective.

What does “genderless” imply? It can be considered as the third gender, including diversity, which does not require boundaries or standards that distinguish gender(male or female). In study 2, it was difficult to distinguish the voice characterized by “young age” into male or female; most participants described it as a “neutral” voice. Users felt more comfortable and showed high preference to this voice because it seemed “relatively young,” “younger than me,”and “high and cheerful.” Similar to the results of Jackson, Williams & Smith (2020), when the voice characteristics of the user and robot match, a more favorable evaluation can be obtained. It was also found that in this case, users tend to feel more familiar with the voice tone and they believe that the voice color of the AI agent is similar to theirs. Because the use of characteristic elements such as closeness and familiarity emphasizes the identity of AI, it can be a potential future step toward developing genderless AI.

Nevertheless, it is necessary to assign gender to intelligent machines in certain situations and context. For example, the leading navigation service T-Map provides version 3 of G like voice in the child protection zone near elementary schools to minimize gender impact and emphasize the surrounding environment to the driver. Because users have high expectations for voices that match the appearance and image of the device in the case of speech-based conversational agents (Carpenter, 2019), the consideration of these “functional elements” that can be extracted by AI from the context of users can also be a prospective direction for designing genderless AI. In other words, designing functional elements that emphasize the context and roles of AI is expected to allow users to perceive AI as a conversation partner with the minimum influence on the existing dichotomous gender classification.