Digital art therapy

Creative processes have a beneficial impact on mental well-being [8]. Reynolds and Prior’s study [9] supports this claim, employing qualitative methods and semi-structured interviews to explore how creative activities influenced subjective well-being among women diagnosed with cancer. Their findings highlighted that the creative process provided participants with positive experiences of success. Flow-like experiences had therapeutic value, as they allowed participants to temporarily put aside illness-related worries during creative activities.

Not only do traditional artistic activities hold such therapeutic potential [9]. Hacmun et al. [5] explored the potential of virtual reality art therapy (VRAT), where participants draw using VR technologies, followed by semi-structured interviews analyzing experiences qualitatively (thematic analysis). Participants reported losing track of time and external reality, experiencing profound present-moment awareness. In the study of Lyu et al. [10], participants generated images expressing their inner feelings associated with the theme “sweet home” using Midjourney, documenting their generation activities and reflections. An important finding was that text-to-image AI allows overcoming limitations of talent and affinity, enabling even non-artistic individuals to easily actualize their ideas and creativity, thus offering significant potential for art therapy contexts.

Self-representation and art

The self can be understood as a dynamic constellation of thoughts, feelings, and behaviors shaped by relations to others and differentiation from them, with its conscious aspect referred to as the self-concept [11]. Self-concept encompasses an individual’s beliefs, convictions, and emotions about themselves, containing all things an individual believes to be true about themselves [11]. Instruments examining the self-construction process seek insight into various aspects of the self [10]. Both qualitative and quantitative methods are utilized in these investigations. Measurement tools and methods include, among others, the Twenty Statements Test, the Self-Construal Scale, and art therapeutic self-portrait creation [12].

According to Muri [13], self-portraits throughout art history enabled artists to deepen self-understanding by facilitating profound reflection on themselves and their experiences. Through self-portraiture, artists can observe themselves from an external perspective, enabling easier expression and validation of emotions and recognition of personal complexity. Self-portraits also help therapists as informative resources, versatile for therapeutic use. They provide insights into a client’s self-concept and perception of their environment. This method can explore various mental issues (e.g., depression, post-traumatic stress disorder, body image issues), potentially aiding their management by supporting and encouraging clients to confront their problems. Creating self-portraits during therapy—when timed appropriately—allows individuals to experience themselves from a novel perspective, fostering personal development [12]. Esteban-Guitart et al. [14] found that university students produced more abstract, complex drawings expressing inner characteristics and metaphors. In contrast, younger participants with basic education created drawings focused on concrete activities and descriptions of their environments.

AI-assisted self-representation

This study forms part of a growing research initiative launched by Kellerwessel and Ujhelyi in late 2022, aimed at exploring the experiential and diagnostic potential of AI-assisted artistic self-representation. The core methodology combines Midjourney with elements from classical self-portraiture and open-ended self-description tasks, treating each text prompt as a condensed self-description [12]. The overarching goal of this research line is to investigate whether generative art can provide meaningful insights into personality, identity, and mental health.

In their foundational mixed-methods study, 11 participants generated self-representative images using Midjourney, followed by brief (~15-minute) semi-structured interviews. Prompt categories were analyzed quantitatively, while thematic analysis of the interviews revealed strong experiential engagement. Participants described feelings of creative freedom, playful exploration, and personal insight. Many reported a sense of creative agency and self-discovery, highlighting the unpredictability of the AI output as a catalyst for new perspectives on the self. A recurring theme was the image as self-representation: participants used personal stories to inspire characters, settings, or objects. One participant, for example, recognized different periods of their life in the images and assembled a kind of visual self-tableau. This core theme inspired the current study’s focus on self-definition.

Building on these foundations, large-scale quantitative research extended the protocol to a sample of 230 young adults, linking prompt sentiment and visual content to depression risk, as measured by the Beck Depression Inventory (BDI).

• The first quantitative study [15] demonstrated that negative sentiment in prompts was a significant predictor of depression scores (r = 0.39; p < 0.001). A ridge regression classifier achieved 79% sensitivity in identifying high-risk participants based on sentiment patterns in their prompts, suggesting that depressive cognitive-affective patterns can manifest even in the prompts themselves.

• The second, qualitative study [16] employed purposeful sampling and focused on the ten participants with the highest BDI scores. The thematic analysis of their interviews and images indicated therapeutic mechanisms akin to traditional art therapy: flow, externalization, and symbolic resolution of emotional conflict. Participants often described “sentiment switches”—abrupt shifts between positively-laden imagery and intrusive self-criticism—and generated unique “resource images” that offered a sense of calm or resilience. Some participants even had a “very intense therapeutic experience” (participant Leaf—p. 7).

These studies suggest that generative AI tools—when used in self-depiction contexts—may serve dual purposes:

• 1.

  As low-cost, scalable screening tools based on linguistic and visual cues, and;

• 2.

  As catalysts for therapeutic reflection, by surfacing emotionally charged or symbolically rich imagery.

However, these earlier interviews were brief and limited by predefined categories, leaving numerous psychologically meaningful experiences underexplored. Furthermore, participants’ image interpretation processes, a key component of any art-based therapeutic or diagnostic tool, were not systematically analyzed and surfaced only when spontaneously mentioned by the participants themselves. The need for a richer, more idiographic understanding of participants’ experiences thus became clear.

To address this gap, the current study adopts a deep phenomenological lens using IPA [17]. This methodology emphasizes idiographic, in-depth inquiry, enabling a richer engagement with participants’ meaning-making. By conducting longer (45–70-minute) interviews and focusing on image-by-image interpretation, this study seeks to explore the following questions in greater detail:

• How do individuals interpret and narrate their self-generated images?

• What forms of self-insight, emotional regulation, or symbolic expression emerge through the process?

• Can these generative outputs serve as tools for therapeutic reflection, particularly in identity-relevant contexts?

The move toward deeper qualitative engagement reflects the need to understand not just what participants create, but how they relate to it, interpret it, and integrate it into their ongoing narratives of the self. The current study, therefore, not only expands the methodological scope of the original research but also deepens its psychological relevance—offering novel insights into how generative AI can support identity work, emotional processing, and potentially even therapeutic change. Exploring participants’ experiences of the self-definition process and the image interpretation segment may contribute to assessing the psychological applicability of the method, particularly its potential for art therapy. The justification for further exploration of these potentials is supported by several studies discussed in the theoretical introduction [5, 7, 9, 10].

The aim of the current research is therefore to explore the practical applicability of the method and to understand what takes place cognitively, emotionally, and experientially in participants during image generation and interpretation. This is pursued through the following research question: What is the experience of self-definition and self-presentation through images during an AI-based image generation process?

Participants

IPA is an idiographic approach and is specifically designed to work with small and homogeneous samples [18, 19]. Participants do not have to represent a population but rather offer access to a specific phenomenon from a particular perspective [20].

In the current study, participants were eligible if they were over 18 years old, had at least B1-level English proficiency, and were not undergoing psychiatric treatment at the time of participation (sine morbo) (see Table 1).

Five participants aged between 18 and 58 were included in the analysis. Their educational backgrounds ranged from secondary school to master’s degrees. All resided in the capital city, were either single or in a relationship, had at least B1-level English proficiency, and were familiar with the AI technique used—some even used it regularly. Only one participant (Budapest) was an outlier in terms of age, but this was not deemed to significantly affect or bias the results.

Interpretative phenomenological analysis

The IPA is a qualitative method based on an interpretative, hermeneutic approach. Its primary theoretical foundation is phenomenology (the science of experience), which focuses on the individual and their subjective interpretations and experiential world [18, 21]. It examines phenomena as they appear from the first-person perspective [22]. Philosophically rooted in existentialism, IPA is particularly suited to exploring how individuals “find themselves in the world” [21], and is particularly effective in mapping identity, making it an appropriate method for investigating self-definition. Interviews created a dialogical space conducive to disclosure and meaning-making between the researcher and the interviewee, a position from which genuine understanding becomes possible. It is especially appropriate when we are interested in the intertwining of experience and self [21], which is precisely the aim of this investigation.

The analysis involved two researchers familiar with IPA and with the details of the study. Interviews were analyzed and group experiential themes were developed through consensus. Reflexivity was maintained via continuous dialogue, with both researchers sharing subjective impressions to identify and control potential biases. The interviewer kept detailed exploratory comments, capturing interpretative comments, emotional responses, and salient elements emerging from the text. The second analyst, who had no direct involvement in the interviews, reviewed these notes, questioned subjective-seeming impressions, and contributed observations from a more distanced perspective.

Theme development was an iterative, hands-on process: approximately 25 experiential statements were identified per interview, printed, and physically segmented to enable transparent categorization. Within each case, similarities and differences between themes were examined to define personal experiential themes (typically 5–7 per participant). In cross-case analysis, overlapping personal experiential themes were compared and merged to reveal group experiential themes present in all participants’ accounts. These were then structured into three group experiential themes that captured the shared aspects of the phenomenon in a nuanced manner. Every step of the process was conducted collaboratively, with final decisions reached by consensus. Where interpretations diverged, we resolved them by consensus, privileging transparency and reflexivity over numerical agreement. In line with IPA conventions, we did not compute inter-coder reliability coefficients, as the aim is interpretative depth and credibility rather than category agreement.

A further added layer in the present research is that, during the image interpretation phase of the interviews, participants essentially engage in self-presentation through visual stimuli. This can be methodologically aligned with qualitative approaches that incorporate visual techniques. For example, drawing-based assessments are increasingly used in qualitative research, as they can facilitate rapport between researcher and participant by alleviating the anxiety of the interview setting or by providing a starting point or support for conversation. They also ensure that participants gradually engage in the process [23]. Combining visual techniques with in-depth interviews can yield a profound understanding of an individual’s subjective experiential world [22].

The workflow of the current research was the following: Interview recordings → Verbatim textual transcription of interviews → Idiographic analysis: explanatory comments and experiential themes, followed by the consensus-based development of personal experiential themes → Cross-case analysis: merging of personal experiential themes, identification of common layers of experiences, consensus-based development of group experiential themes, which constitute the results of the research.

In this study, the generated images serve a similar purpose to drawings. Beyond their practical utility in facilitating the interview, they may also function as a kind of projective surface—through participants’ interpretation of the AI-generated unexpected elements—according to the findings of Kellerwessel and Ujhelyi [12].

Procedure

Data collection took place in a quiet, private room at one of the buildings of the Faculty of Education and Psychology at Eotvos Lorand University. Each session was conducted in a one-on-one setting. The image generation duration was 45 minutes in every case, except for participant Berry, who finished 10 minutes early. Interview durations ranged from 45 to 71 minutes.

After building rapport and providing a short demonstration of the Midjourney software, participants were asked to use keywords of their choice to generate images that best represented them within the following 45 minutes, following the standardized instructions (Supplementary material 1). Midjourney (MidJourney Version 6, MidJourney, Inc.), the software used for image creation, generates images based on prompts (textual descriptions of the desired image) provided by the user. For each of those prompts, the software produces four distinct images, each interpreting the description with a different visual concept. The user can then decide whether to request additional variations or more detailed versions (upscales) of the images, or they can start the process over by making slight or significant changes to the prompt or writing an entirely new one. The process does not require any programming knowledge.

Participants could ask the researcher technical questions at any time, and they were allowed to use an online translation tool (DeepL, https://www.deepl.com) to help with any language difficulties, because the base language of the participants was Hungarian, but the language of the program was English. However, they had no other assistance; for example, they could not use a mirror or any specific images on their phones as a reference.

After the image generation procedure, a deep, approximately 60-minute-long semi-structured interview followed (Supplementary material 2), focusing on participants’ experiences during the image generation process and the background of the created images. In the first part of the interview, participants discussed their images freely (image interpretation phase), while the other part involved more targeted questions about their holistic experience of the process. Compared to the previous studies, a new, longer, and more in-depth interview guide was developed and used, and image interpretation was emphasized during the interview: every participant guided the interviewer through all of their images, explaining their background and the thoughts associated with them one by one—this being the most important novelty in the current study, compared to the previous ones.

The audio files were transcribed manually. Speech disfluencies and repetitions were retained in their original form, and significant non-verbal cues (e.g., laughter) were marked in the transcripts. Each transcript is approximately 20 pages in length.

Participants were assigned pseudonyms of their choice (those having some meaning in Hungarian were translated into English), and they are referred to using these pseudonyms throughout this study. All ethical details were included in the participant information sheet and consent form, which every participant signed. The transcripts and recordings contain no personally identifiable information. Participants were allowed to request the deletion of their transcripts and generated images within one week following the study session, thereby withdrawing their data, but none chose to do so. Upon request, participants received copies of the images they generated.

The complete set of prompts from all participants is included in Supplementary material 3.

The image as a tool: functional images

In many cases, the images served as multi-purpose tools for the participants: they, for example, facilitated emotion regulation, self-reflection, contemplation on topics of high personal relevance, or even existential summation.

A space for self-insight

The interviews revealed that the processes of image generation and image interpretation served as a space for self-exploration for the participants. A space in which they could freely reflect and associate with themselves, sharing whatever thoughts came to mind. It also acted as a space where, through reactions to contextual elements and stimuli, they could gain new insights about their inner functioning.

Modes of self-representation

Two subthemes emerged under the broader theme of self-definition and self-presentation: the role of symbolism and symbolic interpretation of AI-added elements, and the dimension of added meaning arising from the relationships between images.

Dual mechanism of meaning making

Our findings suggest that meaning-making in the AI-assisted self-representation protocol may involve at least two partially distinct pathways. The first pathway appears to be intra-generative: participants sometimes ascribed meaning to images already during the act of generation itself, without requiring subsequent discussion. For example, one participant (Little Mole) followed a deliberate conceptual plan, visually representing successive life periods that “shaped” the current self. This process occasionally involved re-framing difficult life events as necessary elements of personal growth, suggesting that symbolic reinterpretation can occur in situ. Such intra-generative effects could, in principle, be accessible in a solo setting, particularly if accompanied by structured self-reflection tools (e.g., guided journaling prompts).

The second pathway is dialogical, arising primarily during shared narration with a facilitator. Here, meaning was consolidated and made explicit through verbal articulation, co-construction of narrative, and reflective questioning. For instance, Berry emphasized that for her, “talking through” the images was indispensable: it was the act of jointly reviewing the generated images that prompted key self-insight realizations. This aligns with previous literature on externalized narrative processing, where the presence of an attuned listener can scaffold cognitive and emotional integration.

Whether self-guided post-session reflection (e.g., privately revisiting images, keeping a reflective diary) could approximate the benefits of facilitator-supported dialogue remains an open empirical question. Future research could test these mechanisms systematically using a 2 × 2 design (generation: present/absent × guided narration: present/absent), and by adding a self-guided reflective journaling condition to explore whether structured private reflection can approximate facilitator-supported insight.

It is also worth noting that some effects, particularly those related to emotion regulation (e.g., visualizing comfort or aspirational imagery, externalizing problems into symbolic form), were reported independently of subsequent dialogue. For example, Berry described intentionally generating images that she would “like to see in that moment” to foster calmness. Such strategies might retain their benefit even in solo practice, and could potentially be developed into low-intensity, self-administered exercises for emotion regulation, provided that appropriate safeguards and user guidance are in place.

Self-awareness and change of perspective

A key experiential dimension was self-awareness. In the process of depicting and reflecting on themselves, participants had to think about themselves, bringing known self-knowledge to the surface and even generating new insights. In counseling, professionals are encouraged to use creative tools to facilitate self-awareness [24]. Participants chose their own important topics, what they wanted to share about themselves, much like in open-ended questioning [12]. The interpretation phase also had a clarifying function—participants had to adopt an observer perspective, reflecting on their images. According to Drujan et al. [25], such perspective-taking helps distance oneself from intense emotions, potentially easing the processing of emotionally challenging realizations. The images functioned as narrative anchors for self-disclosure and reflection. Participants stated that reflecting on their images helped make their self-image more comprehensible—highlighting inner dynamics and patterns that might otherwise remain unconscious. Several of them said they wouldn’t have reached such realizations without the opportunity to discuss their images.

Symbolization and symbolic interpretation emerged as major modes of self-definition and presentation. Symbol formation is fundamental to art, so its appearance in this creative process is unsurprising. In Lyu et al.’s study [10], participants used metaphors rather than direct depiction, e.g., portraying home as a harbor of love or oneself as a Samoyed dog floating in space. From a psychoanalytic perspective, symbols may reflect hidden desires or unconscious conflicts [26]. Jung and von Franz also viewed symbols as tools for understanding individual and collective psychological processes [26].

Images as tools

One of the central themes was the use of images as tools, for example, for emotion regulation. Comfort images helped participants soothe themselves, while aspirational images helped them to aspire. Visualizing positive emotions fostered self-affirmation, and portraying difficulties enabled integration and processing.

Similar to vision boards—a form of collage illustrating personal aspirations using magazine cutouts, drawings, inspirational words, etc.—many aspirational images in this study served as symbolic representations, rendering abstract dreams more concrete [27]. Vision boards are rooted in creative counseling traditions and aid problem-solving nonverbally [28]. In therapy, they offer a visual representation of clients’ thoughts, facilitating reflection, discussion, problem-solving, anxiety reduction, and identity development [29]. As some participants explicitly stated that the images could later help motivate them toward their goals, AI-generated aspirational images could serve these functions in counselling with appropriate guidance. Aspirational imagery also aligns with psychoanalytic concepts of wish fulfillment, where desires are satisfied through dreams or fantasies [30]. The question arises whether having a visual representation of a fantasy might enhance the feeling of fulfillment in a similar way.

Safe place imagery is a useful emotion regulation strategy, aiding in maintaining positive emotions and processing negative ones. It can be thought of as an emotional refuge that individuals can mentally return to in times of stress to regain stability. It is a relaxation technique that involves imagining a concrete or symbolic place that evokes feelings of safety and calm. It can reduce emotional involvement when reflecting on past events, thus serving as a self-help tool [25]. In the current study, without explicit prompting, participants spontaneously helped themselves in this way when negative memories or emotions were evoked. This aligns with findings from Tomán et al. [31], who used IPA to study the adaptation of individuals with chronic respiratory conditions during the COVID-19 pandemic. They found that safe place imagery supported coping. This raises the question of whether the image generation method could be used, with the right instructions, as a form of self-soothing intervention. The visualized symbols and moods of comfort and safety had a calming effect. Could these images continue to provide comfort when revisited in the future?

Time, self, and narrative

Another important theme was the time and space created by the process. Generating images through AI offered a chance to reflect on the past, interpret the present, and imagine the future, as with some other art therapy methodologies. Participants could summarize their lives, presenting their identity to the interviewer, “as a story to tell.” According to Ricoeur [32], through storytelling, the narrator attempts to interpret the world, and narrative has a constructive role in the emergence of self-awareness [32]. Narrative identity is the internalized and evolving story of one’s life, integrating reconstructed past and imagined future to give life coherence and purpose. Through this narrative identity, individuals communicate to themselves and others who they are, how they became that, and what they see in their future [33].

Self-understanding, however, not only requires listing those past, present, and future selves: it also requires integration both synchronically and diachronically [34]. Participants’ self-definitions reflected both: they mapped their contradictory present aspects alongside temporal self-constructs, attempting to reconcile discrepancies, creating an overarching narrative. The concept of a turning point refers to a significant life event that leads to a clear change in a person’s life, making it a valuable tool for identity researchers [35]. Strauss stated that after such a turning point, the person feels fundamentally different [35]. In this study, some images depicted such turning points—moments after which participants felt they had a “new life.” When reviewing their images, participants like Little Mole expressed that certain events were necessary for becoming who they are now, and the image generation process effectively catalyzed such processes. This further underscores the narrative nature of the image generation process and its potential to examine identity formation through visual turning points.

Limitations and potential risks

The study’s idiographic and exploratory design privileges depth over breadth, which is consistent with the methodological aims of IPA. Accordingly, the sample is small and relatively homogeneous, and the findings are not intended to generalize. Additional limitations relate to the novelty of the research area, the early-stage nature of the method, and the specific characteristics of the tool used (e.g., challenges in human-machine interaction and software constraints). Platform variability is also a non-trivial concern for reproducibility; results should therefore be interpreted with caution. Regarding the subjectivity inherent in IPA, we addressed this by practicing ongoing reflexivity and maintaining continuous analytic dialogue between researchers.

The procedure should be applied with caution in clients with pronounced AI-related anxiety or perfectionistic tendencies, where the unpredictability of generative outputs may provoke distress. It may be contraindicated or require enhanced monitoring in PTSD and dissociative conditions (risk of triggers from vivid, dream-like imagery), in paranoid states (misinterpretation of AI agency), and in severe major depression (potential reinforcement of negative cognitions). We recommend brief pre-screening, explicit informed consent (including data use), a safety protocol for crisis escalation, and careful case formulation to decide whether the method is suitable at a given clinical stage. In all cases, the practice should be framed as adjunctive, not as a substitute for evidence-based interventions.

Conclusions and outlook

The results revealed that the process of image generation can facilitate emotion-regulating function and may aid self-definition and self-presentation challenges through the use and interpretation of symbols and the construction of narratives. In the overall picture, the relationships between individual images also emerged as important tools for self-definition. The significance of the image interpretation phase of the interview became evident:

Participants expressed positive experiences regarding the reflective image-sharing processes, and many of the process’s benefits were realized during this phase through reflection and retrospection, as in earlier studies.

This exploratory study demonstrates the potential of generative AI systems—specifically text-to-image platforms like Midjourney—as engaging, user-driven tools for emotional self-regulation, narrative construction, and identity exploration. The integration of creative digital tools into mental health practice aligns closely with emerging frameworks in digital health innovation, which emphasize personalization, accessibility, and cultural adaptability. Moreover, the combination of qualitative self-report and machine-readable image features paves the way for scalable, hybrid approaches that merge narrative therapy with data-driven assessment. Technology is evolving rapidly, and science should take advantage of every new opportunity. The engaging, experiential, and unique nature of the method—reported by participants—may increase involvement in the process. Due to their natural relationship with digital media, younger generations may be particularly receptive to experiences involving technological tools [5] or populations that may feel less comfortable with traditional talk-based methods. This method may therefore offer a promising way to reach and engage them.

Future research directions

One possible future direction is to examine whether visualizing desires or a safe space in the form of generated images could have similar positive effects as vision boards or safe place imagery. It would be worthwhile to explore the method’s applicability in narrative therapy, including narrative construction, externalization of problems, and the representation of turning points, as well as its use as a tool to study narrative identity and life transitions.

Based on the earlier outlined projective mechanisms, it would also be worth examining whether the method could function as a projective technique. Continuing this line of thought, exploring its parallels with existing projective drawing tests could be an exciting area of inquiry. The method’s potential for art therapy-style applications is also promising, especially in its capacity to help recall and integrate life events and support self-definition and self-understanding.

It would be interesting to investigate to what extent the technique could serve as a reflective surface or conversational anchor during psychological sessions, and whether it facilitates communication between client and practitioner.