Keywords

1 Introduction

The number of devices in homes and personal lives of people using artificial intelligence (AI) seems to increase [1, 2]: Voice-activated assistants featuring AI, e.g. Alexa [3, 4], reach a wide range of customers and nearly everyone knows personalized suggestions [1, 5]. This is changing the way of how people interact with their devices.

People tend to assign a character to objects because it helps to categorize them and adjust the expectations on it. It also masks some of the complexity of the object, which supports the interaction process [6].

However, this cannot be transferred directly to the digital world. Most digital devices resemble each other in their physical appearance as well as in their usability. Data Processing is neither understandable to the average user, nor is it communicated through the device [7]. Because of that, most of today’s AI systems appear like a Black Box [8] to the user which leads to a lack of transparency in how the system works due to missing feedback. This leads to a lack of trust from the user towards the AI. To approach these issues, the following research questions were developed:

1) How can the character of an AI be embodied through shape, tactility and movement? 2) How can a combination of this embodiment and tangible user interaction enhance the relationship of trust of the user towards an AI?

Thereby, the scope was narrowed down to the concept and functional prototype of a sleep companion (“SleepMate”). It is backed by an AI concept giving context for the scenario of waking up and going to sleep. Tangible user interaction is achieved by using three different caps, which can be put on the body of SleepMate and influence the character of the AI. The study carried out with the prototype shows, that 76,47% of the people like the idea of participation in the decision-process of AI. Also, 58,82% of the participants trust SleepMate more than other digital assistants.

The main contributions are the results of a small-scale study on the prototype SleepMate. The results show, that the participants of the study see advantages in SleepMate at physical interaction and embodiment of AI compared to voice control. They also confirmed the tendency that embodiment can enhance trust relationships between users and AI.

2 Related Work

Actual assistants, regardless of their implementation, mainly communicate through voice or text interfaces. This communication is supported by sound notifications, visual feedback (e.g. lights, screens) and the design of the product the assistant is embedded in. The design mostly resembles to brand guidelines, with little differences between the tactile elements of different brands. Most devices feature a minimalistic shape, some are partly covered in fabric [9, 10].

There are a lot of devices and apps that promise to tackle the problem of bad sleep and sleeping behaviour [11,12,13,14]. Some of them are simple data trackers, others are AI-enhanced. For alarm clocks, haptic, visual and tactile feedback was the most desired features. Also, simplicity in use was an often-mentioned feature [15].

Researchers at Microsoft proposed 18 Guidelines which are applicable for Human-AI interaction and were tested and verified with AI professionals on various projects [16]. Also, patterns developed by Smarter Patterns concerning form, functional approach and ethics are used to develop the AI-concept of SleepMate [17].

Enhancing transparency is not inevitably enhancing trust. When systems are designed for transparency they give up on trust and focus on control. Marisa Tschopp distinguishes between reliance and trust over the feeling of betrayal [18]. In this work, this distinction is not made, when the word “trust” is used, it carries both meanings. This helps to focus on the aspect of how the embodiment and interaction can affect trust relationships, which includes both, reliance and trust. Trust can define the way people interact with technology [7].

3 Design Process

According to research question 1, a set of interaction patterns is developed and applied to a prototype in the design process.

The process is split into three main steps: 1) conduction of a preliminary study to gather insight on sleeping behaviour and habits to examine the field in which the prototype will be domiciled, 2) development of the interaction patterns and their merge into a functional prototype, 3) evaluation of the patterns using the prototype.

3.1 Preliminary Study

To gather insight about sleep behaviour, a preliminary study was conducted, using the diary method [29, 30]. The group of participants consisted of 22 people (M = 32.8 yrs.; SD = 14.9) with various socio-demographic backgrounds. Over one week, they were given the same five questions about their habits and sleep behaviour every day via a WhatsApp text message at the same time in the morning [19].

The main results were, that the average sleep time of all participants exceeded 7 h on just two days within the study period of one week. This is less than the recommended [20]. A major part does not sleep a continuous amount during the week with no indications for an established rhythm. 11 of 22 participants set more than one alarm clock or snoozed on at least 3 of 7 days.

3.2 Interaction Patterns and Prototype

Information used to develop the patterns and the prototype was gathered from the preliminary study on sleeping behaviour and habits, a market analysis of devices connected to sleep and behaviour tracking as well as smart devices with and without AI and relevant scholarly papers and books on AI, sleeping behaviour and HMI [21, 22]. Additionally, the lecturing notes and workshop-results of the courses “Design Methods” and “Into Things” of the Magdeburg-Stendal University of Applied Sciences were compiled and accounted.

Based on a first Low-Fi-Prototype a basic shape for the device was developed. It is a cylindrical object with an average height of about 7 cm and a diameter of 14 cm to fit the diameter of the palm of an average hand (Fig. 1). It is divided into two parts, a base part which contains the electronic components and a top part which features exchangeable heads with different surfaces, so-called “caps”. It is assumed that SleepMate can detect if a person is in the same room using appearance sensors.

Fig. 1.
figure 1

SleepMate with the AI-Cap, one of the three interaction-elements

Because the caps are the main instrument of tactile interaction, a variety of materials was gathered, and 16 samples were 3D-printed to research different haptic experiences of surfaces. For the final prototype 3D printing was used, allowing for the incorporation of different tactile experiences into the design that has proven particularly suitable.

In addition to the prototype, an AI-concept was developed. It involves the acquisition of the user‘s fitness, social, productivity, calendar, health, and sleep data, as well as data about the usage of technical devices and services the user already owns (e.g. Wearables, Smartphones, Online Services) as well as built-in sensors of SleepMate (appearance sensors, light sensor). Additionally, external databases (like public health databases) are used as a reference. Based on that, the AI generates personalized recommendations for going to sleep, the duration of sleep and waking up, which are improved by continuous machine learning [21, 23].

The interaction mechanism with the device and the character of the AI are focused on tactile interaction. The character trait of the AI is influenced based on the caps that are put on the top part of SleepMate [24,25,26]. This significantly affects the behaviour of SleepMate by shifting different scales of character traits (e.g. energy or strictness) and enables or disables certain functions e.g. when waking up its user [24]. To develop and test the interaction mechanisms, three different scenarios, as well as the corresponding instances of caps and interaction patterns were developed (See Table 1).

Table 1 Scenarios and corresponding interaction patterns.

Construction.

For the movement patterns, three servo motors with gears and gear racks were integrated into the base part. Also, a NeoPixel RGB LED Ring for the light patterns is located there. The 3D-printed parts are built of white matte and transparent PLA-Filament. This enabled to illuminate the gap between the top and bottom part (See Fig. 2). The NeoPixel Ring, the servo motors and a Bare Conductive Touch-Board for external speakers, are connected to an Arduino ESP-32. Through a Bluetooth connection between the ESP-32 and an Android device, it is possible to remotely control the prototype and execute the scenarios repeatedly for the Wizard-of-Oz (WOz) evaluation [27, 28].

Fig. 2.
figure 2

From left to right: Internal construction, middle layer with NeoPixel LED ring and AI-driven cap on the device.

For a video of the described expressions, please refer to: https://youtu.be/SXLBkJXhFH4

4 Evaluation

According to our research questions, a structured interview was conducted to 1) evaluate how users react to the execution of the prototype and the concept of tactility, movement, light and sound in the context of AI and 2) study how embodiment changes trust towards smart assistants.

4.1 Apparatus and Procedure

After an initial study (n = 6) at an exhibition, the evaluation was adjusted. The final survey form consisted of four parts: 1) questions about the socio-demographic background, the preferred waking-up device and the previous experience with AI; 2) specific questions regarding trust towards existing AI devices, their reliability and possible application in exemplary situations; 3) is based on the AttrakDiff 2 short questionnaire [29] which uses the semantic differential to determine the handling and appearance of the prototype; 4) questions about the participant’s opinions on SleepMate. Part 4 included five questions about trust towards AI that were already asked in part 1 and 2 of the survey, which allowed for a direct comparison between existing smart devices and SleepMate.

The study was conducted in a seminar room, where SleepMate with its nightstand (Fig. 1) was placed on a desk with the neutral cap on the top and the other caps beside. To allow for an easier analysis a digital survey form was used.

After filling part 1 and 2 of the survey, participants were encouraged to investigate the prototype and the caps on their own and try possible interaction. After that, the participants were given an explanation on the functions of the device and told to complete the survey (part 3 and 4) [30].

Seventeen persons, ranging in age from 14 to 55 years (M = 31.4 yrs; SD = 12.7) participated. Since the usage of an alarm clock (or comparable devices) is a normal behaviour, no further exclusion criteria were used.

4.2 Analysis

Simple “scheme-checking” was eliminated through alternation of the position of the positive and the negative word pairs. Also, the AttrakDiff 2 short questionnaire was used to reduce bias [29]. The survey forms were evaluated after all participants had given in their forms.

For closed questions 5- and 7-point Likert scales were used. In the evaluation Top and Bottom Box scorings were grouped, leaving out the midpoint, and converted into percentage values [31, 32].

4.3 Results

52.9% of the participants use digital assistants at least once a week. 70,59% use smartphone alarms for waking up. But only 29.41% could image using digital assistants in their bedroom. Reasons are, inter alia, bad control surfaces and privacy concerns.

The short AttrakDiff has shown mainly positive feedback towards the appearance of SleepMate, with ratings towards the adjectives elegant (M = 1.4), good (M = 1.5), simple (M = 2.2), practical (M = 2.3) and predictable (M = 3.2).

Participants stated the following reasons for using SleepMate in their bedroom: Better interaction than smartphones, no usage of smartphones in the bedroom, functionality is aimed at supporting sleep, aesthetic reasons.

All participants stated that light and movement clearly support the character of the AI. Furthermore, 64.7% see advantages in physical interaction and embodiment of AI compared to voice control. Stated reasons being more unambiguous control, no usage of microphones, no need to talk, possible usage with closed eyes. In addition, 58,82% of the participants would trust SleepMate more than other assistants, because it is more personal, more tangible and less concerning when it comes to privacy.

5 Discussion

The results of this work show, that an enhancement of trust from the user towards the AI is favourable. It also shows that there is mistrust and privacy concerns against companies that acquire personal data. A connection between privacy-related trust concerns and bedrooms is suspected. Still, character and embodiment can have an influence on trust relationships with AI. The user feels more involved in the decision-making process (Table 2). A similar effect of tangible interaction is described by the Interaction Design Foundation for interaction that focuses on “doing” instead of “receiving” [33].

Table 2. Results of comparable questions in pars 1/2 and 4.

Also, there are indications that the embodiment of AI can be attained through shape, tactility and movement. This is shown by mainly positive feedback on SleepMate with tendencies to the attributes “predictable”, “practical”, “simple” and “good”, as mentioned in the results. The radical approach of not giving control by design shows the need for a balance between transparency and control to earn trust. There remains a gamble between trust and control when transparency in design is focused [18].

58,82% of the participants of the study said that they would trust SleepMate more than other digital assistants. They also ranked SleepMate more positive than other digital assistants even though the rehearsals did not use it for quite a long time. It remains to be noted, that building up trust is a complicated and delicate process [7] and that over time relationships of trust can change in a positive and negative direction.

Also, the participants stated that tactility, light and movement support the character of the AI, with 64.7% seeing advantages in physical interaction and embodiment of AI compared to voice control. This shows, that a combination of embodiment and tangible user interaction can enhance the relationship of trust of the user towards an AI, as predicted in the research question.

6 Future Work

An evaluation in a sleep environment, either by using a WOz-approach or implementation of alarm clock functionality into the prototype, would enable to gain more insight into Human-AI relationships in the given scenario. To verify the evaluation and find effects on long-term trust-building [7] this evaluation could be extended over a longer period of time. Because the approach was narrowed down to one specific scenario and use-case there is room for improvement in the prototypic approach as well as a generalisation for a broader range of application. This allows for the design patterns to be transferred to other use-cases.

7 Conclusion

This paper contributed to the evaluation process on how the character of an AI can be embodied through shape, tactility and movement and how the combination of this embodiment and tangible user interaction can enhance the relationship of trust of the user towards an AI. The main contributions are the results of a small-scale study on the prototype SleepMate. The results show, that the participants of the study see advantages in SleepMate at physical interaction and embodiment of AI compared to voice control. Also, the embodiment is described with positive adjectives which indicates positive character-shaping effects. They also confirmed the tendency that embodiment can enhance trust relationships between users and AI, with 58,82% of the participants saying, that they trust SleepMate more than other digital assistants. The developed interaction patterns can be used in projects in the field of human-AI-interaction to build up trust and communicate character.

It can be assumed, that the character and embodiment of an AI, as well as tangible interaction, can increase the trust of the user towards the AI.