Artificial Intelligence

The term “AI” carries a breadth of meanings that have evolved alongside its advancements. The essence of AI, as pointed out by AI pioneer John McCarthy, morphs as its applications become ubiquitous in everyday technology (Vardi, Reference Vardi2012). At the center is the notion of “intelligence” itself. Although definitions vary, they commonly highlight abilities in reasoning, problem-solving, and adapting to new challenges (Sternberg, Reference Sternberg2003). In an effort to integrate these recurring themes, AI researchers Legg and Hutter (Reference Legg and Hutter2007, Reference Legg and Hutterp. 9) propose defining intelligence as “an agent’s ability to achieve goals in a wide range of environments.” This perspective suggests that intelligence, fundamentally, is about an entity’s adaptability and its proficiency in navigating a spectrum of scenarios to achieve its goals.

The “artificial” aspect of AI lies in its deliberate design, contrasting with biological intelligence that naturally occurs in living organisms (Gabriel, Reference Gabriel2020). As such, AI research focuses on building intelligent agents that choose actions to maximize performance based on received inputs and inherent knowledge, where agents perceive their environment through sensors and act upon it through actuators (Russell and Norvig, Reference Russell and Norvig2022, pp. 54–58).

Building on this understanding of AI, it is essential to recognize that AI systems exist within a complex sociotechnical context. Dobbe et al. (Reference Dobbe, Gilbert and Mintz2021) highlight the frequent discrepancy between the promised benefits of AI systems and their actual consequences, termed the “sociotechnical gap.” This gap arises from the divergence between socially necessary outcomes and what AI can technically achieve. For example, while a recommender system may aim to provide valuable suggestions to users, in practice, it could inadvertently promote misinformation or polarization.

To address this challenge, various scholars have proposed understanding AI as a sociotechnical system that encompasses not only its technical capabilities but also its limitations and the governance structures surrounding it (Dean et al., Reference Dean, Gilbert, Lambert and Zick2021; Dobbe et al., Reference Dobbe, Gilbert and Mintz2021; Krippendorff, Reference Krippendorff2023; Selbst et al., Reference Selbst, Boyd, Friedler, Venkatasubramanian and Vertesi2019; Stray, Reference Stray2020; van de Poel, Reference van de Poel2020; van der Maden et al., Reference van der Maden, Lomas and Hekkert2022; Vassilakopoulou, Reference Vassilakopoulou2020). For instance, ChatGPT should be considered in terms of not merely its underlying model but also its user interface, the company behind it, public perceptions, and the various use cases and purposes it serve. As such, some of these scholars advocate for adopting a cybernetic perspective, which emphasizes the importance of feedback and adaptation in managing the inherent complexities of sociotechnical systems, giving rise to the inclusion of non-technical and natural entities (Dobbe et al., Reference Dobbe, Gilbert and Mintz2021; Pangaro, Reference Pangaro2021; van der Maden et al., Reference van der Maden, Lomas and Hekkert2022; Krippendorff, Reference Krippendorff2023).

Cybernetics: AI as sociotechnical system

Cybernetics, which emerged in the 1940s, is a transdisciplinary field focused on communication, control, and circular causality in systems (Mindell, Reference Mindell2000). The term “cybernetics” is derived from the Greek infinitive “kybernao,” meaning “to steer, navigate, or govern.” A core concept in cybernetics is the feedback loop, which creates circular causality between a system’s past, present, and future states (Wiener, Reference Wiener1961). At its core, cybernetics presents an alternative perspective to traditional AI design by emphasizing the symbiotic relationship between humans and machines within complex sociotechnical systems (Mead, Reference Mead, von Foerster, White, Peterson and Russell1968; von Foerster, Reference von Foerster2003). It focuses on the dynamics of feedback loops, communication, and control mechanisms that underpin both biological and mechanical systems, proposing that understanding these can enhance the design and function of AI (Beer et al., Reference Beer, Chiel and Sterling1990; Sato, Reference Sato1991).

By viewing AI through a cybernetic lens, designers are encouraged to consider AI not just as isolated algorithms but as part of an interconnected web of social, technological, and environmental factors (Dobbe et al., Reference Dobbe, Gilbert and Mintz2021; Scott, Reference Scott2004). This perspective offers a holistic framework for understanding the challenges around designing AI across various systemic layers. As such, the design of positive AI interventions can go beyond the algorithm and even the platform itself. For example, companies may use the method to make adaptations to the recommender systems that govern their platforms, while smaller design firms may develop a third-party add-on that alters the interaction with a platform to support wellbeing. Interventions can even take place in the broader ecosystem, such as the development of institutional guidelines for use of ChatGPT in education – thus lowering student anxiety in using these tools while bolstering its potential educational impact.

Challenges of designing AI

When discussing the design AI systems, what specific process are we referring to? The development of AI is often conceptualized as a multi-stage life cycle, traditionally segmented into seven key stages,Footnote ² with the design phase being integral in translating business cases into engineering requirements (Morley et al., Reference Morley, Floridi, Kinsey and Elhalal2020). This phase is critical, as it lays the groundwork for how AI systems will function and interact within human contexts. Norman and Stappers (Reference Norman and Stappers2015) advocate for the involvement of designers throughout the entire development process of sociotechnical systems. However, this article primarily addresses the design phase. This emphasis does not detract from the importance of a holistic approach – which we strongly support – but rather aims to provide a detailed examination of the unique challenges and opportunities within this specific phase. By concentrating on the design phase, we aim to address the nuances that shape the early and critical decisions in AI development, understanding that these decisions have far-reaching implications for all subsequent stages.

Returning to the question at hand, Yang et al. (Reference Yang, Steinfeld, Rosé and Zimmerman2020) highlight that designing effective AI systems presents unique challenges distinct from those encountered with traditional software systems. They argue that the uncertainty surrounding AI capabilities and complexity of possible outputs makes it difficult to ideate, prototype, and evaluate human–AI interaction using standard HCI methods. As they point out, AI systems continue adapting after deployment, so designers struggle to anticipate changing behaviors across contexts. Additionally, the near-infinite output possibilities, especially for adaptive AI, mean that traditional prototyping fails to capture the full range of behaviors and experiences.

Furthermore, as we will address later, effectively incorporating wellbeing into AI design demands engaging with user communities. However, integrating user communities into the AI development process is challenging because of technical complexities, the unpredictable evolution of AI technologies (Sadek et al., Reference Sadek, Calvo and Mougenot2023b), significant communication gaps (Piorkowski et al., Reference Piorkowski, Park, Wang, Wang, Muller and Portnoy2021), and lack of relevant expertise (Hsu et al., Reference Hsu, Verma, Mauri, Nourbakhsh and Bozzon2022). These difficulties are exacerbated when designing for values such as wellbeing, as they are complex, multifaceted (Schwartz et al., Reference Schwartz2012), and interpreted differently across individuals (Graham et al., Reference Graham, Haidt and Nosek2009) and cultures (Sachdeva et al., Reference Sachdeva, Singh and Medin2011). As such, there are many possible interpretations of values like fairness, trustworthiness, and empathy, as well as disagreement over their relative importance (Jakesch et al., Reference Jakesch, Buçinca, Amershi and Olteanu2022).

While there is broad aspiration toward high-level AI ethics principles like fairness and transparency, translating these into practice remains challenging (Morley et al., Reference Morley, Kinsey, Elhalal, Garcia, Ziosi and Floridi2021; Schiff et al. (Reference Schiff, Rakova, Ayesh, Fanti and Lennon2021b). For example, a review of guidelines on AI ethics found extensive discussion of principles like transparency and fairness, but very little on technical explanations for achieving them (Hagendorff, Reference Hagendorff2020). Similarly, Schiff et al. (Reference Schiff, Borenstein, Biddle and Laas2021a) underscore the complexity of applying ethical principles like fairness and transparency across sectors, highlighting a gap in consensus on practical implementation, which directly impacts the integration of values such as wellbeing into AI systems. Bridging this divide between principles and practice remains an open research challenge. It requires developing methods that reduce the indeterminacy of abstract norms while retaining adaptability to diverse contexts (Jacobs and Huldtgren, Reference Jacobs and Huldtgren2021).

In this regard, we may look to VSD as a promising methodology for embedding abstract values such as privacy into concrete design specifications, thereby guiding AI systems to better serve and reflect the diverse needs of stakeholders while promoting inclusivity and human-centricity in technology (Zhu et al., Reference Zhu, Yu, Halfaker and Terveen2018b). For instance, Umbrello and van de Poel (Reference Umbrello and van de Poel2021) present a case study in which they successfully translated crucial values like non-maleficence into actionable design criteria for a novel AI system.

However, Sadek et al. (Reference Sadek, Calvo and Mougenot2023b) note that a significant shortfall in current VSD practices is their inability to effectively assess whether these values are genuinely reflected in the outcomes of AI systems, highlighting a lack of impact assessment mechanisms (both qualitative and quantitative). It is this gap that our method tries to fill for two reasons. First, for any impact-centered method (which arguably any value-oriented design project is), it is essential to establish causal links between interventions and system fluctuations (Fokkinga et al., Reference Fokkinga, Desmet and Hekkert2020) – mere good intentions do not cut it. Second, as Schiff et al. (Reference Schiff, Ayesh, Musikanski and Havens2020) point out, impact measurement leads to evidence-based decision-making and promotes accountability, thus fostering iterative improvement. Now that we have an overview of the challenges related to designing AI, let us turn our attention to the additional challenges introduced by a focus on wellbeing.

Challenges of designing AI for wellbeing

Designing AI systems specifically to enhance human wellbeing introduces additional complexities. That is, wellbeing is inherently multifaceted, is variable across individuals, and manifests differently across cultural contexts, making it difficult to define and design for in a measurable way (Halleröd and Seldén, Reference Halleröd and Seldén2013; Huppert, Reference Huppert, White, Slemp and Murray2017). AI systems often optimize narrow objectives, making it hard to ensure that they improve wellbeing holistically. Instead of promoting broad human flourishing, they tend to target limited metrics. This narrow focus in AI systems’ design and implementation is recognized as one of the six grand challenges for human-centered AI (Ozmen Garibay et al., Reference Ozmen Garibay, Winslow, Andolina, Antona, Bodenschatz, Coursaris and Falco2023). To address these challenges further, a recent article identified seven key challenges for designing AI for wellbeing specifically (van der Maden et al., Reference van der Maden, Lomas, Malak and Hekkert2023b). They used a cybernetic framework to group the challenges into four categories, listed below and mapped to a simple schematic in Figure 1.

• • Conceptualization of wellbeing: the challenges around choosing the appropriate theoretical paradigm for conceptualizing wellbeing and modeling wellbeing contextually given its complexity and unclear relationships between system components and wellbeing facets.

• • Operationalization of wellbeing: the challenges of measuring wellbeing contextually with adaptive instruments, translating small-scale qualitative wellbeing data to large-scale metrics suitable for optimization algorithms, correlating self-report and behavioral data collection, and reconciling the different paces at which wellbeing changes versus AI optimization occurs.

• • Designing AI actions that promote wellbeing: the lack of mature methods and examples for putting wellbeing at the core of AI system design, beyond just avoiding harm.

Figure 1. It shows a schematic representation of a cybernetic system. The different categories of challenges can be mapped onto this framework: (1) understanding the system context which entails modeling the relation between wellbeing of the systems constituents and its various components; (2) operationalizing said model of wellbeing; (3) designing interventions to actively promote operationalized model of wellbeing; and (4) retaining alignment with the overall goal. The latter refers to both challenges of algorithmic optimization and scrutinizing the objective (e.g., Is the wellbeing objective still aligned to needs and desires of all relevant stakeholders?) Used with permission from van der Maden et al., (Reference van der Maden, Lomas, Malak and Hekkert2023b).

Most tools focus on alignment but lack concrete guidance on promoting human flourishing.

• • Optimizing for wellbeing: the challenges of making trade-offs between competing objectives (e.g., individual versus communal wellbeing) when optimizing AI systems for wellbeing, and dealing with the fundamental constraint that wellbeing changes slowly while AI optimization is rapid.

Several frameworks have been developed to guide the design of wellbeing-supportive technology, such as the framework by Wiese et al. (Reference Wiese, Pohlmeyer and Hekkert2020) that maps wellbeing-enhancing activities (Lyubomirsky and Layous, Reference Lyubomirsky and Layous2013) to digital technology design and the “METUX” framework by Peters et al. (Reference Peters, Calvo and Ryan2018) that supports wellbeing in digital experiences. However, these methods were not specifically developed for AI and do not directly address all four challenges mentioned earlier. The IEEE-7010 standard (Schiff et al., Reference Schiff, Ayesh, Musikanski and Havens2020), however, provides a more holistic approach tailored for the development and assessment of autonomous and intelligent systems (A/IS) with human wellbeing at the forefront. This standard offers an iterative wellbeing impact assessment (WIA) process, stakeholder engagement, and a focus on wellbeing indicators across various domains. Through these facets, the standard addresses three of the four challenges by supporting the conceptualization, operationalization, and optimization of wellbeing. However, as a standard rather than a design method, it does not directly guide designers in translating insights from the assessment into concrete design interventions.

Thus, our goal is to develop a method that not only builds upon existing frameworks, such as those outlined in the IEEE-7010 standard, but also harmonizes with recognized design and innovation approaches, including design thinking (Dorst, Reference Dorst2011; Cross, Reference Cross2023) and the double diamond model (Design Council, 2007). By doing so, we hope to invite human-centered designers to the field of AI and bring human-centered design principles to the development of AI systems.

Development of the method

The positive AI method was developed using a cybernetic approach as an organizing framework following earlier discussions. Cybernetics views systems as cyclic processes of sensing states, comparing to goals, and taking action (Mindell, Reference Mindell2000). This perspective enabled organizing the design challenges into distinct phases, with each phase addressing a different category of challenges (Fig. 1), while acknowledging the inherent entanglement present in complex sociotechnical systems (Dobbe et al., Reference Dobbe, Gilbert and Mintz2021).

We developed this method following a research-through-design process that drew inspiration from existing frameworks such as the earlier mentioned IEEE-7010 standard (Schiff et al., Reference Schiff, Ayesh, Musikanski and Havens2020). The development involved collaboration between designers, researchers, and students over multiple projects. An initial two-year project designing a cybernetic system for institutional wellbeing during COVID-19 informed the first version of the method (van der Maden et al., Reference van der Maden, Lomas and Hekkert2023a), which incorporated elements of the IEEE-7010 process, such as stakeholder engagement and a focus on wellbeing indicators across various domains. This early version was then refined through an iterative process that involved scrutinizing the method’s efficacy and incorporating community feedback. The refinement focused on streamlining the method’s steps, improving its understandability, and enhancing its relevance. These various versions were then tested in five design courses given at the master’s level, where student teams designed AI systems aiming to support wellbeing. These findings were then consolidated to present the version that is evaluated in this article.

Further, the positive AI method is intended to complement and enhance typical design processes. For example, it parallels the empathize, define, ideate, prototype, test, and implement phases of design thinking (Dorst, Reference Dorst2011; Cross, Reference Cross2023), with a specific focus on wellbeing and AI. Furthermore, the phases of our method align with the convergence and divergence characteristic of the double diamond framework (Design Council, 2007), while also emphasizing the iterative process inherent in most design strategies and frameworks.

Taking these considerations together, we arrived at five phases in the proposed method. This choice mirrors phases in other design cycles and highlights the importance of restarting the cycle, often less evident in typical AI development and psychological research. To emphasize that continuous improvement of the wellbeing model is central to positive AI, we included restarting the cycle as a separate step. The phases correspond to the four categories of challenges (conceptualization, operationalization, design, and optimization), with an additional stage for implementation.Footnote ⁵ We will discuss these phases in depth next.

Phases of the positive AI method

In short, the method involves ensuring that AI systems are sensitive to factors of human wellbeing and enabled to support them. Consequently, the five phases should help the designer to understand wellbeing in context (phase 1), to make it measurable (phase 2), to design systems (inter)actions that promote wellbeing (phase 3), to implement the designs (phase 4), and to sustain alignment (phase 5). Figure 2 shows an overview of the methods phases with brief annotations of the content of each phase. A useful checklist with the activities and outcomes of the method can be found in the Appendix.

Figure 2. Diagram illustrating the positive AI method’s cyclical approach within a wellbeing context. Phases 1 (contextualization) and 2 (operationalization) primarily contribute to developing the sensors of the AI system, while phases 3 (design) and 4 (implementation) focus on developing its actuators – reference the section on the background of AI for a discussion of sensors and actuators. The cycle culminates in phase 5 (continuous alignment), demonstrating an ongoing feedback loop between all stages and its environment.

Method applied to fictional example of a streaming platform

To illustrate the positive AI method in action, this section presents a fictional scenario involving a streaming platform seeking to align with wellbeing goals. In this hypothetical case, the process would begin with a review of literature on video platforms and human–AI interaction to compile a list of key features and hypothesize their impacts on wellbeing. For example, studies suggest personalized video recommendations can sometimes limit users’ openness to new perspectives and may create filter bubbles (Ferwerda and Schedl, Reference Ferwerda and Schedl2016; Dhelim et al., Reference Dhelim, Aung, Bouras, Ning and Cambria2022; Areeb et al., Reference Areeb, Nadeem, Sohail, Imam, Doctor, Himeur, Hussain and Amira2023). In contrast, features like custom video playlists may boost users’ feelings of autonomy and control over their viewing experiences (Möller et al., Reference Möller, Trilling, Helberger and van Es2020). At this stage, our initial theoretical model incorporates hypothesized relationships between specific AI features (such as recommendations and playlists) and key aspects of wellbeing (like openness and autonomy).

Next, we could refine our preliminary conceptual framework through comprehensive user studies. This step would involve conducting interviews, organizing focus groups, and administering surveys. The aim of this research phase is to gather direct feedback on how the platform’s functionalities impact user wellbeing, balancing the identification of obstacles with potential improvements. By engaging a diverse participant group, we seek to capture a wide range of insights and user experiences.

We would then operationalize wellbeing by selecting validated scales like the Personal Growth Initiative Scale (Robitschek et al., Reference Robitschek, Ashton, Spering, Geiger, Byers, Schotts and Thoen2012) to quantify growth and openness to new ideas.

Additionally, we would develop context-specific metrics, such as an aggregated playlist complexity score. This local metric could be calculated from factors such as breadth of topics, diversity of creators, and degree of organizational structure in users’ video playlists. It would serve as an indicator of the level of perceived control over viewing experiences.

Equipped with this contextualized model of the platform’s wellbeing impacts, we could propose targeted interventions to optimize the scales and metrics. For instance, one could suggest an algorithmic adjustment that sporadically introduces unexpected video recommendations, motivating users to explore content beyond their regular preferences, thereby potentially elevating personal growth metrics.

To implement such proposals, collaborative sessions with designers, engineers, and users would allow iteratively developing and refining features based on observed wellbeing impacts and user feedback. Designers would facilitate participatory design workshops to envision algorithm tweaks and interface changes. Engineers would build required components and monitor the system. Users would provide perspectives to ensure changes align with their values and needs.

To evaluate intervention effectiveness, we would employ a two-phase process: qualitative methods in the local design context, followed by large-scale A/B testing for system-wide verification. This approach ensures locally observed positive effects translate to the system level. By continuously revisiting the contextual model and indicators based on these evaluations, the platform could incrementally adapt its AI systems to support multidimensional wellbeing objectives. This balanced set of global and local metrics enables holistic progress tracking, potentially moving the platform beyond user satisfaction toward genuine wellbeing alignment.

Having discussed this hypothetical example, we will now proceed to our three real-world case studies.

Procedure

Three student projects utilizing the positive AI method were initiated over a three-week period (March 2023). Each student received a personalized introduction to the project and was provided with reference materials including an overview document of the method, recommended literature, and the “Positive Design Reference Guide” (Jaramillo et al., Reference Jaramillo, Pohlmeyer and Desmet2015) to support their research. When the third student commenced their project, a collaborative kick-off meeting was held to establish a shared understanding of the method, address questions, and align expectations across the projects.

Over the remainder of the project, the students met weekly with the supervisory team for guidance. This structured approach aimed to sufficiently equip the students with the necessary understanding and resources to effectively apply the positive AI method within their individual projects. By providing one-on-one introductions, reference materials, a collaborative kick-off, and ongoing supervision, the aim was to support the students in successfully utilizing the positive AI approach.

Then, to gather information on the method efficacy and usability, multiple sources of information were consulted. These included observations taken during the weekly meetings, progress reports and presentations, the final design outcomes and reports, and three recorded and transcribed one-on-one interviews with the students. The weekly meetings provided a platform for the students to share problems they encountered. Oftentimes, they faced similar hurdles, which were carefully documented. The progress reports and presentations served as useful post hoc data to examine how the students were dealing with challenges, how the process developed, and how the designs developed over time. Finally, the interviews aimed to substantiate key themes identified throughout the project period. The first author analyzed the various data sources and shared the findings with the students to ensure accuracy. Before presenting these results, we provide a brief overview of the final design outcomes.

Materials: design outcomes

In this section, we briefly summarize the final design concepts resulting from the three student projects applying the positive AI method.Footnote ¹³ We present the core functionality and wellbeing goals addressed by each design to provide context before examining the process evaluation findings.

MiHue

Student 1 designed a dating app called “MiHue” that leverages AI to enhance users’ experience of autonomy and relatedness. The core concept balances the needs for uniqueness (autonomy) and connection (relatedness) by highlighting individuality within similarity.

To identify which wellbeing aspects to focus on, Student 1 began by researching the user’s contextual wellbeing needs and experiences in using dating apps. Her literature analysis revealed autonomy and relatedness as salient wellbeing facets impacted by dating platforms. To further refine her contextual understanding of wellbeing, she also held multiple generative workshops with target users. During these co-design activities, participants also ideated potential design interventions focused on supporting self-expression and social bonds.

Synthesizing her findings, Student 1 operationalized autonomy and relatedness within her context as the ability for users to express their unique attributes (autonomy) and to find meaningful connections based on shared interests or experiences (relatedness). She then formulated her design directions aimed at enhancing social connection by highlighting individuality within similarity – also referred to in design literature at “Autonomous Yet Connect” (Blijlevens and Hekkert, Reference Blijlevens and Hekkert2019). This approach focused on promoting a shared connection through uniqueness and common ground, using the AIxD Ideation cards (Piet and MOBGEN, Reference Piet2019) to link technology capabilities with desired wellbeing outcomes. This led to new features that encouraged users to share more personal and diverse aspects of their identities, beyond physical appearance. One key feature was an improved profile creation tool that prompted users to respond to creative and introspective (AI-generated) questions, facilitating deeper self-expression. Another feature was an algorithm designed to match individuals based on not only mutual interests but also shared values and life goals, aiming to foster more substantial and meaningful connections.

To closely mirror real-world application, she developed a strategy for implementing the novel features either within an existing app or as a standalone platform. Subsequently, she conducted user testing with an interactive prototype (designed using Figma). These tests included questions based on her earlier operationalization of the contextual wellbeing model such as those related to autonomy (e.g., “How well does the app allow you to express your true self?”) and relatedness (“Can you describe any interactions you had through the app that made you feel understood or belonged?”). This process was aimed at gathering feedback to refine the recommendations for subsequent iterations, thereby embodying the method’s emphasis on continuous alignment. Her subsequent recommendations for the next phase emphasized expanding the focus to encompass additional aspects of wellbeing not initially covered but highlighted in the theoretical model, such as self-acceptance, positive emotions, and physical health. Moreover, she underscored the importance of including diverse user groups, particularly minorities, and considering gender differences, to ensure a more inclusive and holistic approach to enhancing wellbeing through the app’s usage. Figure 3 presents this design in three simple illustrations.

Figure 3. Three visuals used to illustrate key aspects of MiHue’s journey as presented in the expert study: (a) the protagonist’s frustration with current dating apps that focus on looks over personality; (b) the protagonist entering their interests during MiHue’s enhanced account creation process that encourages authentic self-representation; (c) the protagonist matching with someone who shares common interests, as highlighted by MiHue’s features that spotlight unique and shared traits between users to foster meaningful connections.

FoodVibe

Student 2 designed an app called “FoodVibe” that uses AI capabilities including facial recognition, natural language processing, and machine learning to provide personalized recipe and dining recommendations tailored to users’ specific social contexts and past preferences. The adaptive system aims to promote wellbeing through home dining by encouraging mindful eating and nurturing social connections with shared meals. The idea for FoodVibe originated from the fact that existing nutritional and dieting apps tend to emphasize calorie intake and nutritional intake rather than other aspects of eating that affect wellbeing.

To identify what wellbeing aspects to prioritize in her design, Student 1 conducted a literature study as well as experience sampling (Van Berkel et al., Reference Van Berkel, Ferreira and Kostakos2017). Her literature analysis revealed mindfulness, social connections, autonomy, and engagement as salient yet overlooked facets. She then combined this information with results from the sampling study to map a user journey specifically aimed at understanding when certain wellbeing experiences may occur. Next, to refine her understanding, she hosted two generative workshops where users emphasized the value of reflection, awareness, and social aspects around meals. This led her to operationalize wellbeing in her context (eating at home) as being present (e.g., engaging with your meal instead of the television) and having a sense of belonging (e.g., feeling related to your family when cooking a nostalgic dish).

This led to the design of FoodVibe which enhances wellbeing by encouraging mindful dining at home, giving users autonomy in their food choices, and deepening connections with dining companions. Utilizing AI, FoodVibe personalizes recipe suggestions by analyzing users’ dietary preferences, the people they eat with (identified through facial recognition), and past meal satisfaction. Its features focus on personalizing meal recommendations, enriching social interactions by connecting with friends within the app and aligning meal choices with the group’s tastes, and promoting self-reflection on dining experiences to boost wellbeing aspects like autonomy, positive relationships, and mindfulness.

The final design was evaluated through user testing with a high-fidelity prototype. The evaluation focused on whether the app achieved its design vision and goals, emphasizing the enhancement of eating experiences and perceived wellbeing of users of healthy-eating app. The effectiveness of FoodVibe was assessed based on metrics related to autonomy, positive relationships, mindfulness, engagement, fun, and the overall usability and desirability of the app. These metrics were derived from theoretical models and earlier phases of the research. This led to recommendations for subsequent iterations focusing on user experience improvements, like using avatars for privacy, broadening wellbeing theories to various dining contexts, boosting AI accuracy for tailored recommendations, and conducting thorough user testing for long-term wellbeing impacts. To illustrate this design, Figure 4 presents it through three simple diagrams.

Figure 4. Three visuals used to illustrate key aspects of the FoodVibe journey as presented in the expert study: (a) a user frustrated with nutritional limitations when deciding what to cook; (b) the user utilizing FoodVibe’s “Recipe Generator” by taking photographs of ingredients on-hand so the app can suggest customized recipes; and (c) two people cooking together in the kitchen, representing FoodVibe’s goal of promoting wellbeing through shared meals and human connections.

Explore More

Student 3 designed a new Spotify feature called “Explore More” that uses the platform’s algorithms and extensive music catalog to guide listeners through unfamiliar genres in a personalized way. This idea stems from the observation that existing personalization mechanics, such as “Discover Weekly” playlists, tend to converge on a type of music which over time can get uninspiring. The goal is to expand users’ musical tastes and perspectives to foster personal growth and empathy.

She began by analyzing the current landscape of music streaming service features and linked these to wellbeing literature. Her analysis revealed personal development, specifically through music’s potential to facilitate self-discovery, as an impactful yet underutilized application for enhancing user wellbeing. Synthesizing her contextual findings, the student recognized limitations of Spotify’s existing personalized discovery playlists driven by recommender systems, which can restrict users within narrow musical preferences over time.

Upon recognizing these limitations, she directed her efforts toward intentionally utilizing music’s capacity for perspective expansion and self-discovery to promote personal growth (Hallam, Reference Hallam2010). In other words, she operationalized wellbeing as increased engagement with unfamiliar music genres, hypothesizing that exposure diversity would lead to self-development. This inspired the design of Explore More, a feature that could either be integrated into a service like Spotify or as a stand-alone third-party interface. Features included an interactive genre map to visually navigate through unexplored musical territories, guided discovery paths offering sequences of new genres tailored to the user’s tastes, personalized recommendations within those genres to ensure a resonant listening experience, self-reflection prompts aimed at deepening users’ introspective engagement with the music, and a feedback and adaptation mechanism to refine future explorations based on users’ experiences and preferences.

She developed an interactive Figma prototype for user testing, revealing key insights into navigation ease, the effectiveness of discovery paths, and the resonance of music recommendations. Self-reflection prompts were particularly noted for deepening users’ personal insights and musical connections. Based on feedback, she recommended refining the UI for better navigation, enhancing the recommendation algorithm for tailored music exploration, and deepening self-reflection prompts for richer introspection. Additionally, she proposed adaptive feedback mechanisms to align the exploration journey with users’ changing tastes, ensuring Explore More effectively supports personal growth and musical discovery. The design is illustrated in Figure 5.

Figure 5. Three visuals used to illustrate key aspects of the Explore More journey as presented in the expert study: (a) a bored user unsure what to listen to; (b) an interactive map of music genres that lets users visually browse and see their tastes in context; (c) a user happily dancing after Explore More recommended an unfamiliar yet related genre, demonstrating how it aims to broaden perspectives and facilitate personal growth through personalized music discovery.

Results of the case studies

To reiterate, a key aspect of evaluating the method is assessing its efficacy by examining whether the designer understands the wellbeing impacts and thoughtfully considers them in their design decisions. Additionally, it involves grasping the relationships between wellbeing dimensions and system components. On the other hand, method usability refers to aspects such as avoiding unnecessary detours or delays, understanding the steps involved, and feeling confident in executing the method Figure 5.

The multi-case study revealed both strengths and weaknesses of the positive AI design method when applied by novice designers. In terms of method efficacy, students initially struggled to feel confident in their understanding of the wellbeing literature. This was partly due to their unfamiliarity with the field. That is, the breadth of literature was overwhelming, causing uncertainty about when enough research had been done to proceed. This resulted in hesitancy during key stages as students were unsure they grasped concepts well enough to move forward.

Additionally, the lack of familiarity with translating wellbeing goals into technical requirements or metrics affected designers’ ability to thoroughly address wellbeing aims in their solutions. The unfamiliarity with core wellbeing concepts led to doubts about properly executing the methodology. Overall, students lacked confidence evaluating wellbeing considerations throughout the process.

Initially, designers faced challenges in evaluating and addressing wellbeing in their design process. However, as the project progressed, the methodology compelled them to consistently test their approaches against the specific context and the people involved. This iterative process gradually sharpened their focus and deepened their understanding. By the project’s end, this rigorous application resulted in a notable improvement in their ability to integrate wellbeing considerations, as evidenced by their satisfaction with the final solutions. This evolution underscores the methodology’s effectiveness in facilitating a contextually relevant approach to wellbeing in design.

In terms of method usability, the process involved iterative transitions between research, ideation, and prototyping rather than a linear sequence. This is often the case for design processes; however, the framing of the method as steps and the initial visualization of them (not included here) gave them the idea it would be more linear. Further, the overall structure provided helpful guidance, but some inefficiencies occurred due to unclear context definition early on and lack of knowledge about wellbeing. Therefore, significant time was spent exploring literature not directly relevant. To some, this may have felt as a waste of time. In reality, it is likely designers will explore directions that in the end of their project may not be relevant anymore – this is what makes them a good designer, being open to multiple avenues and perspectives. The designers suggested to include an introductory wellbeing course, a better explanation of the context early on, and to spend more time planning upfront. Despite inefficiencies, designers were satisfied with their process and outcomes overall.

In summary, the positive AI method demonstrated efficacy in guiding novice designers to translate high-level wellbeing goals into concrete design proposals grounded in user values. Key steps like contextualizing and iterative development focusing on emergent wellbeing priorities were effective, as evidenced by resulting concepts aligned with community experiences. Despite its inefficiencies, designers expressed satisfaction with the methodology, as it allowed them to address complex tasks through repeated testing against user perspectives, gradually improving contextual understanding. Thus, there is potential to enhance effectiveness in combination with refinements aimed at improving usability.

In terms of method efficacy, key points of attention are as follows:

• • Navigating expansive wellbeing literature overwhelmed novices, causing uncertainty grasping concepts to advance confidently. Clearer guidance on scope is required.

• • Unfamiliarity translating qualitative aims into technical specifications made operationalization challenging. More extensive scaffolds are needed to aid comprehension.

• • Initially lacking familiarity with core wellbeing concepts hampered confidence assessing impacts. But this grew through repeated engagement as understanding increased over time.

Additionally, regarding method usability, certain inefficiencies emerged despite the beneficial structure:

• • Unclear initial scope caused detours exploring tangential literature. Signposting priorities earlier would help.

• • Absent examples induced difficulty judging step completion. Providing benchmarks would resolve ambiguity.

• • Operationalization demands proved taxing for novices. Enhanced support could alleviate strain.

• • Shifting between abstract and concrete perspectives around wellbeing felt jarring. Framing this dynamic approach as integral to the design process could smooth transitions.

Having addressed the intricacies of applying the positive AI method through novice designers’ experiences and identified areas for refinement to bolster both its efficacy and usability, we now turn our attention to the last research question. In the following section, we present a narrative-based study involving experts to evaluate the quality of the AI system concepts resulting from the application of the positive AI method.

Method

Results of the narrative-based study

Figure 6 presents a graph of the results from the narrative-based study involving expert evaluations across three distinct concepts: MiHue, FoodVibe, and Explore More. The concepts, received moderately high mean ratings on perceived realism, impact on wellbeing, business desirability, and business feasibility. The concepts, on average, received moderately high ratings across the metrics. Specifically, Explore More was rated highest in business desirability, while MiHue and FoodVibe showed similar ratings in perceived realism and wellbeing impact, respectively. Notably, there were no significant differences in ratings among the different expert groups. The associated standard deviations indicate a moderate variation in expert opinions. On the qualitative front, feedback from two participants indicated that the narratives were lengthy, while another two experts remarked that the stories leaned toward being overtly positive. However, they also noted their understanding of this positive skew, acknowledging the study’s context aiming to portray an ideal user experience. Further insights from the expert’s feedback will be addressed in the subsequent sections.

Figure 6. Barchart comparison of expert evaluation ratings across the three concepts: MiHue, FoodVibe, and Explore More. Metrics visualized include perceived realism, wellbeing impact, business desirability, and business feasibility.