Co-Creating GenAI-Generated Visualizations of People Living With Dementia: A Participatory Text-To-Image Method and Case-Based Evaluation

Abstract

Generative Artificial Intelligence (GenAI) and large language models (LLMs) are increasingly shaping public perception of dementia through visual representations. These technologies often reproduce harmful and stigmatizing stereotypes that misrepresent the lived experiences of people living with dementia (PLWD). This article introduces and evaluates a three-phase framework for GenAI-facilitated co-creation sessions that centers PLWD as active co-creators in generating more authentic visual self-representations using GenAI. Through multiple case studies, key methodological and ethical challenges were explored, including strategies for effective prompting, approaches to creating recognizable yet dignified representations, and methods for including non-verbal PLWD as co-creators. Finally, the broader implications of this co-creative approach for challenging GenAI bias, future research and developing more ethical GenAI systems are discussed.

Introduction

In 2021, OpenAI introduced DALL·E, a deep learning model capable of generating digital images from natural language prompts, followed in 2022 by ChatGPT, a conversational chatbot. The emergence of such generative AI (GenAI) tools opens new possibilities for alternative forms of expression, which is particularly relevant for people living with dementia (PLWD). Dementia is a progressive condition that gradually affects how people perceive, interpret and act upon the world around them (Zwijsen et al., 2016). This condition can diminish people’s word-finding abilities or coherence in speech (Kempler & Goral, 2008; Klimova & Kuca, 2016), making self-expression more challenging (Banovic et al., 2018). GenAI text-to-image generators could empower PLWD by enabling them to communicate their lived experiences visually, even when verbal expression becomes difficult (McConnell et al., 2019). While AI technology has already shown promise in dementia research through AI-based cognitive testing (Li et al., 2022), MRI image analysis (Kameyama & Umeda‐Kameyama, 2024), and socially assistive robots (Newby et al., 2023), the potential for supporting communication and self-expression in this context remains underexplored. However, realizing this potential requires careful consideration of the biases embedded within these systems.

GenAI text-to-image generators highlight problematic aspects, including the large-scale reinforcement of stereotypes (Bianchi et al., 2023). Because generative models are trained on patterns found in existing visual material, the images they produce inevitably reflect and reinforce dominant cultural biases (Jenks, 2025). This tendency is confirmed by research analyzing AI-generated images of PLWD, revealing stereotypical portrayals such as passive postures (Jintaganon et al., 2025), muted color palettes (Spencer & Spencer, 2025), and an overrepresentation of elderly, light-skinned, female individuals (Putland et al., 2025).

We observed similar biases during an exploratory image-generation test conducted with DALL·E 2 in 2023 (Figure 1). The aim of this small-scale evaluation was to investigate potential biases and stereotypes emerging from AI-generated images of PLWD. We produced sixty unique images using a simple prompt to generate an image of “a person with dementia” without additional instructions. Remarkably, all images featured the stereotypical “headclutcher” framing (Orton, 2022; Putland & Brookes, 2024; Van Gorp & Vercruysse, 2012) depicting individuals holding their heads, suggesting distress or confusion. Additionally, most (44 out of 60) generated characters appeared female, and all individuals were portrayed as older adults. Furthermore, only five images portrayed individuals with a non-white skin tone, highlighting a lack of diversity. These findings underscore problematic biases within image-generation algorithms that could inadvertently perpetuate stereotypes associated with dementia.

Figure 1.

Composition of various test images illustrating the recurrent “Head Clutcher”-Framing. Note. Generated by the authors with DALL·E2 in 2023

While AI-generated imagery and LLMs hold significant promise, researchers and other stakeholders must remain vigilant when utilizing these tools. Spennemann (2025) found in his analysis that both GenAI (DALL·E) and LLMs (ChatGPT) can produce biased outputs. Therefore, users are recommended to proactively counteract these biases by avoiding generic prompts and instead crafting more precise, diversity-aware prompts.

The objective of this methodological article is to examine the potential offered by GenAI to produce alternative, more inclusive visual narratives of PLWD. We argue that meaningful and ethically responsible image generation is possible when PLWD are engaged as active participants in a co-creative process, where their self-perceptions and lived experiences inform the prompts. Yet, this approach brings forward several methodological and ethical challenges, which we will address and illustrate through concrete case examples drawn from our own qualitative and participative research practice.

Literature Review

Research With People Living With Dementia

Until a decade ago, PLWD were rarely involved in research on their own condition (Webb et al., 2023), because it was assumed that their cognitive limitations prevented them from making meaningful contributions to research (Lepore et al., 2017). Recently, however, there is growing recognition of the value of involving PLWD as active participants or co-creators in research and design (Griffith et al., 2024; Hendriks et al., 2019), as their lived experiences offer first-person insights that secondhand accounts cannot (Eriksen et al., 2022). This offers a more holistic understanding of dementia, including both challenges and strengths (Wolverson et al., 2016), and leads to more relevant and person-centered research, which improves care and support services (Miah et al., 2019; Wehrmann et al., 2021). Participation also fosters a sense of empowerment and agency, as PLWD contribute to knowledge that directly affects their lives (Webb et al., 2023). Therefore, capturing their perspectives is essential. Qualitative research methods are particularly well-suited for this purpose, as they offer ethical and in-depth ways to involve PLWD in research (Hennink et al., 2020).

In-Depth Interviews With People Living With Dementia

In-depth interviews enable a nuanced exploration of various topics and help establish a safe and supportive researcher–participant relationship (Dempsey et al., 2016; Shannon et al., 2021). This method is highly adaptable to participants’ needs (Novek & Wilkinson, 2019), allowing flexibility in the pace of the interview and the complexity of the questions: short, clear questions often work better than long, open-ended or multi-part ones for PLWD (Suijkerbuijk et al., 2024).

However, in-depth interviews may present specific challenges when working with PLWD. Dementia-related symptoms, such as difficulties with word finding, sentence comprehension, and coherent discourse, can hinder verbal communication (Banovic et al., 2018; Klimova & Kuca, 2016). Discussing emotionally charged experiences related to dementia can also be stressful for participants (Bartlett & Martin, 2002).

Creative Methods for Expressing the Lived Experience of Dementia

Given the limitations of in-depth interviews with PLWD, researchers propose two complementary methodological premises: (1) a participatory research design and (2) a visual research method. Participatory design shifts PLWD from passive subjects to active collaborators who co-shape the research process and outcomes (Clarke et al., 2018). This approach empowers PLWD to shape how their experiences are represented, rather than positioning them solely as respondents to researcher-driven questions (Pain, 2012).

Combining a participatory research design with visual research methods further enhances this empowerment. Visual methods enable the expression of thoughts and abstract concepts through imagery, thereby reducing reliance on verbal articulation and memory recall (Evans et al., 2016). Images frequently serve as tools for reminiscence, revealing what is meaningful in a person’s life (Beuscher & Grando, 2009; Robinson, 2000). Studies on reminiscence therapy indicate that using communication technology and visual or auditory materials from the past can facilitate social interactions and boost self-esteem (Lazar et al., 2014).

Among visual research approaches, photo elicitation and photovoice are particularly prominent. Photo elicitation uses photographs in interviews to stimulate discussion, foster emotional connections and make memories more tangible (Harper, 2002), qualities that suit research with PLWD, whose cognitive impairments can affect verbal communication and memory (Bates et al., 2017; Kyololo et al., 2023; Phillipson & Hammond, 2018; Wang, 2023).

Photovoice, on the other hand, stimulates greater involvement from participants, as they take photographs themselves, then select and discuss them. This process, known as autodriving, allows participants to guide the conversation, often uncovering insights that (semi-)structured interviews might overlook (Shell, 2014). This method enables PLWD to express their experiences from their own perspective (Jenkins et al., 2025).

Besides these photography-based methods, collaborative illustration offers another participatory approach where illustrations are co-created through dialogue between participants and human artists. In this process, artists translate verbal descriptions into imagery via iterative feedback, producing personalized visuals without requiring technical skills (Dadkhahfard & Takeuchi, 2025). Unlike traditional photographic practices, illustration can depict experiences, emotions, and imagined futures for which no photos exist, enabling exploration of abstract concepts and aspirations central to understanding dementia.

However, these visual methods show limitations when working with PLWD. Photography-based methods require technical skills, fine motor skills and adequate vision, which may be compromised in this population, while the delay between taking photographs and discussing them introduces memory retention challenges (Shell, 2014). Ethical concerns also arise regarding the use of identifiable images of participants (Lood et al., 2023). Collaborative illustration, on the other hand, is time-intensive and requires access to skilled artists. The delay between verbal descriptions and receiving visual output can create similar memory-related challenges.

Considering these opportunities and challenges, there is a need to explore visual methodologies that retain the participatory and expressive strengths of these approaches while addressing their practical and ethical constraints in dementia research. GenAI text-to-image generation offers a potential solution by providing real-time visual generation with immediate iterative refinement capabilities. These tools could allow healthcare providers or professionals to apply similar approaches in their practice. Furthermore, GenAI enables serendipity, meaning that elements in the output can elicit meaningful interpretations, enriching the co-creative dialogue and opening up new avenues for reflection. Our approach draws on the core principles of these visual approaches, while adapting them using GenAI to address mentioned methodological limitations.

Co-Creation With People Living With Dementia

Creating visuals of how PLWD perceive themselves and experience life with a dementia diagnosis requires a creative and reflective process with active participant involvement. This demands a person-centered approach, thorough preparation, flexible methodology, and a multidisciplinary research team (Campbell et al., 2023; Webb et al., 2020). There is broad consensus that PLWD can meaningfully contribute to academic studies, marking a shift from research on ‘end users’ to co-creation as participatory design (Sanders & Stappers, 2008). In dementia-related research on tools and technologies, this involves dynamic interplay between co-creation and co-design (Tsekleves et al., 2020), where PLWD are regarded as experts who contribute throughout the design process based on lived experience (Leorin et al., 2019). GenAI extends this further: rather than serving as a passive tool, it enables iterative collaboration where visual representations emerge through active dialogue between co-creators, researchers, and AI systems, making the process an act of collective creativity (Rezwana & Maher, 2023; Wang et al., 2019).

In our approach, researchers serve as artist-facilitators (Tsekleves et al., 2020), translating PLWD’s input into prompts suitable for GenAI – a role requiring commitment to social change and sensitivity to the ethics of image-making, including awareness of biases, personal preferences, and power dynamics (Wang & Burris, 1997).

Since each person with dementia is unique, the goal is not to produce generic visual representations of living with dementia, but to co-create individualized images that reflect each co-creator’s subjective experience. GenAI is used to generate tangible visual representations in real time during the co-creation process and to iteratively refine them to closely match the individual’s experiences. This process embraces the unpredictability of the generative model, allowing for serendipitous visual elements that enrich the co-creation and deepen the participants’ reflection. By enabling co-creators to express their experiences without requiring technical skills or identifiable exposure, GenAI-based image generation may expand the expressive toolkit available to researchers and co-creators. This approach not only preserves the core values of participatory and visual research but also opens new avenues for capturing the lived realities of PLWD.

In summary, our study defines GenAI-facilitated co-creation as a collaborative process where researchers enable individuals affected by stereotypical social representations, working together toward a shared outcome through intentional use of GenAI. This study investigates how this technology can be meaningfully and ethically integrated into qualitative research to deepen our understanding of dementia from the inside out. Our research question reads as: How can GenAI-based image generation be integrated into a participatory research design to gain insight into the lived experience of people living with dementia?

Method: AI-Facilitated Co-Creative Sessions

Facilitated by GenAI

Each co-creation session was a one-on-one meeting with a person living with dementia, centering their identity, lived experience, and personal narrative in the creative process. The primary outcome was an GenAI-generated image, serving not as a literal photograph but as a visual representation of how the co-creators perceive themselves, their experiences, and their relationship with dementia. In cases where a co-creator demonstrated limited insight into their condition, this was not treated as a limitation but rather understood as an integral aspect of their lived experience, which could be meaningfully reflected in the final image.

Unlike traditional visual approaches, this method employed LLM-based text-to-visual GenAI to democratize the creative process. Three advantages informed this choice. First, it allowed creative expression without requiring artistic expertise, enabling PLWD and researchers’ equality. Second, rapid, iterative images generation kept co-creators actively involved in shaping their portraits, supporting agency and engagement throughout the process. Third, although it can be a hassle to generate exactly the image co-creators have in mind, GenAI applications offer great flexibility to start over or combine different life aspects, such as past and present, and unexpected elements that GenAI introduces can stimulate creativity.

We generated images with 30 co-creators (see Table 1, names are pseudonyms to ensure co-creators’ anonymity.) using DALL·E 3 via ChatGPT between October 2024 and December 2025, as this setup best supported complex prompts and iterative feedback from semi-structured interviews. From March 2025, we adopted GPT-4o for image generation within ChatGPT, introducing new possibilities and challenges – a transition we experienced as a liminal shift. In September 2025, we therefore shifted back to DALL.E3. In November 2025 we used Google’s model Nano Banana Pro for the generation of two images (see Table 1). Each visual representation was created in a dedicated thread. In accordance with ethical research guidelines and to safeguard co-creators’ privacy, the associated prompts were excluded from use in the broader training of OpenAI’s and Gemini’s language and image models. We disabled data sharing for model improvement, and we applied data minimization by using anonymized input and avoiding unnecessary personal details. When reference images were used, we removed embedded metadata prior to upload to reduce the risk of unintended disclosure.

Setting

Co-creators were recruited through visits to day centres and residential care facilities, where researchers introduced the project and identified individuals who expressed interest. To ensure the inclusion of less socially outgoing participants, additional one-on-one conversations were conducted. Each co-creation session involved one person living with dementia at a time and proceeded only after verbal – and where possible, written – informed consent was obtained. Informed consent forms were reviewed together with the researcher to ensure that all aspects of participation were understood and approved. Informal caregivers were informed about the study and consulted after. In cases where a co-creator was unable to provide written consent, it was obtained from an informal caregiver to allow the use of the co-created materials in research. In our study, there was no one who did not have a close informal caregiver or family member.

The co-creation sessions took place in day care and residential care settings rather than in a research environment, allowing co-creators to remain in a familiar environment. To minimize distractions, sessions were held in quiet areas – gardens, terraces, or co-creators’ own rooms – chosen based on their preferences. These familiar and informal locations were chosen to avoid the emotional distance or discomfort associated with more formal settings (Beuscher & Grando, 2009). The sessions were conducted in a circle to encourage natural conversation and equality between co-creators and researchers (see Figure 2). The multidisciplinary research team brought expertise in clinical psychology, communication, and the arts.

Figure 2.

Research setting from the third researcher’s perspective, drawn by Baldwin Van Gorp. Note. Left to right: interviewer, co-creator, live prompting researcher

Most sessions involved three researchers, each fulfilling a distinct role. One researcher facilitated participative live prompting using LLM-based text-to-visual GenAI. A second conducted semi-structured interviews while monitoring co-creators’ non-verbal responses. The third observed the overall process to safeguard methodological consistency. Live prompting during interviews required intense concentration to filter information and transform it into effective prompts, making it unfeasible for one person to conduct the interview and create prompts simultaneously. We also encountered technical challenges that risked disrupting the co-creation process. The three-researcher setup proved essential for managing these challenging moments without significantly disrupting the conversation flow. This approach enabled insight into co-creators’ perspectives while monitoring their emotional and physical well-being (e.g., signs of fatigue, confusion, or distress), aiming to ensure they felt safe and at ease (Stanyon et al., 2016).

A GoPro camera recorded the interviews and interactions from a distance that captured co-creators’ facial expressions without disrupting the conversation. These recordings, together with screen recordings and webcam footage of the co-creators’ facial expressions, allowed for later examination of co-creators’ engagement and facilitators’ influence. Sketches and photographs documented the setting, while prompts, responses, and generated images were archived, to providing a holistic overview of our research approach and output. All materials were stored in a secure environment.

In-Depth Interviewing

The interviews began with a brief introduction, during which we presented ourselves and explained the study’s purpose. To demonstrate GenAI’s creative potential in an accessible way, we generated unrelated images based on co-creators’ input, such as a skating rabbit in graffiti style or a photorealistic unicorn on a tractor. These examples were deliberately general, carried a surreal tone, and were unrelated to dementia so as not to influence the co-creators’ perspectives. The conversation then began with an open-ended question such as: “Now that you know who we are, and what we are going to do, could you tell us a something about yourself?” Depending on which aspects co-creators spontaneously shared, the interview explored these topics further, focusing on themes such as personal preferences, daily activities, emotions, experiences with dementia, and their presence in day center or residential care facilities.

Once an initial visual representation had been generated based on co-creators’ narratives, questions shifted to image-related inquiries like: “Do you feel this image captures your experiences well?”, “Do you recognize yourself in it?”, and “What would you change?”. For co-creators who found answer open-ended questions challenging, researchers selected parts from the images to discuss one by one. Both verbal and non-verbal responses, such as facial expressions, posture, and orientation towards the computer, were taken into account. This iterative process continued until a visual was created that co-creators felt accurately represented them. For co-creators with limited or no verbal communication skills, researchers instead relied on interpreting facial expressions, body language, non-verbal sounds, and gestures when discussing questions and visual elements.

Participative Live Prompting

In this study, we employed a participatory and iterative live prompting framework to translate PLWD’s lived experiences into co-created visual representations (Figure 3). While all phases and components were present in most sessions, they were applied flexibly to maintain the natural flow of the conversation. This workflow not only provided the necessary flexibility in our methodology but also ensured structure and detail in our prompt design (Ahmad & Ruslan, 2024).

Figure 3.

Prompting framework for GenAI-Inspired co-creation sessions with PLWD

The challenge was staying true to the conversation while creating structured and detailed prompt aligned with the co-creator’s expectations. The generated images were shown directly to the co-creators, allowing them to follow the process closely and respond immediately to what was happening on screen.

With the design of this framework, we created a human-centered and participatory methodology, one that aligns closely with the needs, voices, and agency of PLWD. In the initial sessions, co-creators were asked to press the enter key to confirm each prompt submission, as a gesture of ongoing consent. However, this practice was soon discontinued, as it disrupted the conversational flow and caused confusion. Recognizing this, we adapted our approach. Even without the confirmation step, co-creators remained engaged and actively guided through the process, indicating when adjustments were needed or when a different visual representation better reflected their experiences. This highlighted the limitations of strict structural approaches and underscored the value of a more flexible method, one that can accommodate the unpredictable nature of GenAI facilitated co-creative sessions with PLWD.

Initial Scene Request

The first stage of prompting was informed directly by the initial open-ended questions of our semi-structured interview, our observations, and ongoing dialogue. Through this process, researchers and co-creators collaboratively identified the key elements to be included. The stage began with a scene description used to generate a visual representation of the co-creator situated within a specific scene. It was followed by a detailed persona description, specifying age, gender, hairstyle, and eyewear. We only included the dementia diagnosis in the prompt description if the co-creator mentioned it. For co-creators who did not identify as PLWD, due to anosognosia or denial, we omitted the term from the prompts.

We incorporated biographical and thematic elements into the prompt design by including occupations, hobbies, family connections, and/or symbolic motifs meaningful to the co-creator. We defined the setting and geographical data, situating each scene in a recognizable interior or exterior location. Styling instructions were added to specify emotional tone (e.g., warmth, hope, distress) and visual style (e.g., photorealistic versus pop-art or surrealism), added with lighting or camera-angle cues to set a specific atmosphere. The initial scene serves as a starting point for the iterative feedback cycle. An example of a prompt used to establish the initial visual scenario for a co-creator, named Walter (Figure 4A) is shown below:

Figure 4.

(A) GenAI-Generated image in co-creation with Walter and (B) Composition image of the iterative cycle of 25 alternations. Note. Images are generated by DALL·E3 on March 7^th, 2025, one month before the Walters’ euthanasia. He wanted to generate an image of himself on how he wanted to be remembered

Can you create an image of a fifty-year-old man? He has short dark hair and black round glasses. He is wearing black pants, a T-shirt, and a light jacket. On his t-shirt is a self-portrait by Rembrandt, an etching. The man has always taught and studied art history. Art is truly his passion. He loves painting and especially enjoys action painting, using his whole body. It makes him very happy. Can you create an image of this man in a museum full of art, where he is simultaneously making an action painting? Paint splatters are flying everywhere.

The Iterative Feedback Cycle

The second stage enabled the co-creators to guide refinements in response to each generated image (Figure 4B). Each prompt tweaked a specific element to keep the process transparent and manageable. We also discussed technical limitations of GenAI, including constraints, artifacts, and model hallucinations. For example, in one session, we had to wait eleven minutes for an image, and at times the model failed to make requested changes, such as adding a butterfly (Figure 5B). Another important component of this phase was active bias detection and mitigation. In our study, the GenAI-model frequently produced images showing individuals far older than the co-creators, particularly those with early-onset dementia. To counter this, we reinforced their approximate age and active posture within the prompts. Feedback was phrased directly in response to the visual outcome, to give further creative directions like: “Very good! Can you also depict the man with bare feet, so without shoes?” or “Can you add a Labrador Retriever to the image and have it lying between the man’s legs?” (Figure 4A and 4B).

Figure 5.

(A) and (B) GenAI-Generated images in co-creation with Alice. Note. Image is generated by DALL·E3 on March 3rd, 2025

The Narrative Alignment Phase

The final stage aimed to reach consensus with the co-creator on the generated image and the narrative. At times, we restarted the process to explore new visual directions that better aligned with the co-creator’s narrative or offered greater depth to their lived experiences, while retaining key elements such as biographical details and persona descriptions. For example, during the co-creation session with Alice (Figure 5A and 5B), we used the following prompt to explore a new creative direction and capture another aspect of her experience: “Can you generate a new image of this woman with a gigantic, heavy backpack that’s bursting with clothes, and her husband helping her carry the backpack?”. Once co-creators were satisfied with the result and wanted no further refinements, the session concluded by printing and framing the final visual and handing it over to them.

Case Examples and Challenges With AI-Facilitated Co-Creative Sessions

Visualizing Authentic Lived Experiences

Many images were created with co-creators who were aware of their dementia diagnosis and wanted their personal experiences to be represented accurately. As a result, most articulated clearly what they wanted the images to represent. For some co-creators, this meant including concrete symbols or objects. Some of them, for instance, asked for a clock in their images to symbolize cognitive decline – something once intuitive and used daily had become difficult to interpret. The clock also resonates with familiar neurological assessments such as the Clock Drawing Test (CDT) or the Mini-Mental State Examination (MMSE). Others focused on situational themes. David, for example, chose to include his kitchen to underscore the theme of vulnerability, as this once-familiar space had become unpredictable and unsafe due to his decreasing abilities to assess the function and potential dangers of everyday objects (Figure 6).

Figure 6.

GenAI-Generated image in co-creation with David. Note. Image is generated by DALL·E3 on September 25^th, 2024

Others were particularly attentive to the emotional tone and visual style of the images. For instance, Eileen, who has a background in psychology, valued imagery that captured the subjective reality of living with dementia. She revised her images meticulously to achieve a sense of accuracy and resonance (Figure 7). She experiences hallucinations and the disorienting feeling of being in a bedroom without knowing whether it is hers or a neighbor’s. She therefore felt that images rendered in a surrealistic style and setting better reflected the nature of her condition.

Figure 7.

GenAI-Generated image in co-creation with Eileen. Note. Image is generated by DALL·E3 on November 19^th, 2024

Finding Strong Opening Narratives

We found that a strong initial scene request was crucial for generating an image that immediately resonated with co-creators’ lived experiences, allowing researchers to swiftly capture the visual narratives’ essence. Once a solid foundation was present, it became easier to fine-tune the image and incorporate additional key elements, producing visuals that held genuine value for the co-creator. In many cases, this initial resonance stemmed from a clear visual resemblance to the persona and recognizable scenery. Interviews lacking a strong opening story or scene request often required more time and effort to establish that same connection through image generation.

A thoroughly developed initial scene also strengthened co-creators’ trust and engagement in the process. Co-creators expressed appreciation and appeared genuinely pleased by the sense of being understood. For example, after we showed Alice our first generated image (Figure 5A), she was surprised and said with a smile: “Oh, that’s right! But yes, that’s exactly how it was! Those clothes were really lying at the bottom.”

A Demand for Visual Resemblance and Nuance

Some co-creators emphasized the importance of being accurately represented, not necessarily in a photorealistic way, but in a manner that allowed for a better identification with their GenAI-generated personas. This became evident from their initial responses to the images, with remarks such as: “Is this supposed to be me?” and “They [my peers] look so old?”. The most frequently requested adjustments related to age category, gender, hair color, facial hair, and accessories such as glasses, hats, and jewelry. Yet, modification often went beyond these basic features. Alice, for instance, highlighted the significance of her permanent makeup, which she described as a reflection of her desire to maintain an elegant appearance despite no longer being able to manage her former daily beauty routine (Figure 5A). She was also highly specific about the AI-generated persona of her husband, prompting several adjustments, such as fuller, black hair, a moustache, and a more youthful appearance (Figure 5B). Agatha, by contrast, focused less on facial features and more on the clothing worn by her persona, particularly the color, shape, and collar of the dress. As a former seamstress known for this specific style of clothing (Figure 8), she wanted her representation to reflect her craftsmanship. These preferences also showed notable interpersonal differences: some co-creators were satisfied with a general resemblance, while others desired almost photographic accuracy.

Figure 8.

GenAI-Generated image in co-creation with Agatha. Note. Image generated by ChatGPT 4o on April 8^th, 2025. Due to the photorealistic resemblance in this image, it is pseudonymized to protect the anonymity of the co-creator

One of the main issues we noticed with more photorealistic images was that co-creators identified less with their persona if it did not fully align with their self-image. Detailed prompts alone often failed to achieve the desired resemblance, an issue less pronounced in the more stylized, graphic images produced by DALL·E 3. To address this challenge within ChatGPT 4o, we photographed the co-creator, removed all metadata and uploaded it with consent into the thread to generate a virtual persona based on the image. This approach, used, among others for Figure 8, proved highly effective and promising in several cases.

A consequence of the demand for physical similarity was that the generated images became less stereotypical and more diverse. These images moved beyond the stereotypical portrayal of dementia as affecting only older, mostly white women, offering more inclusive representations instead (Figures 4 and 8). Furthermore, our person-centered approach contributed to a more humanizing portrayal of dementia by shifting the focus away from deficit-based or diagnosis-driven representations and highlighting the individuals and their experiences.

Need for Recognizable Scenery

Co-creators expressed a clear preference for images depicting a specific, unambiguous situation related to their lived experience with dementia. Although they often shared multiple examples of what daily life with this health condition entails, combining multiple experiences into a single image was generally perceived as less effective in conveying their lived experience. This was because co-creators recalled distinct situations, and merging these made it difficult for them to identify with the combined scenario.

Because many co-creators repeated certain aspects of their lived experience – due to memory difficulties – we focused the visual representations on the elements that were mentioned frequently, assuming these were the most meaningful or defining. Alice, for instance, repeatedly mentioned her struggle with taking initiative and staying organized since her diagnosis, contrasting this with how tidy she used to be. She expressed frustration about her wardrobe overflowing with clothes she could no longer manage to put away (Figure 5A) and deep gratitude for the friends and caregivers who helped her reorganize it. When we visually represented the chaos in her bedroom, the image resonated deeply with her.

The Inclusion of Co-Creators With Less Verbal Abilities

Several co-creators had limited verbal abilities. Peter, for example, could primarily mimic sounds, especially from bird species, and repeated only a few words (Figure 9). To create a compelling opening image, we first gathered background information from his caregivers, who explained that he enjoyed spending time in nature. Consequently, we conducted the co-creation session in the garden, while walking. Throughout the interview, he frequently imitated the calls of birds and frogs, which revealed a fascination with these animals. When they were added to the image, he responded with enthusiastic approval, characterized and noticeable in the video footage by widened eyes, nodding, and a higher pitched voice. His reaction indicated that he may have felt seen, as if the image acknowledged his way of interacting. It also became clear that representing his former military profession was important to him, as he mimicked the sound of a helicopter. Including visual elements related to his military background contributed to a more meaningful and personalized portrayal. A smile appeared when he saw the image containing all these elements. When one of the researchers asked whether the image resembles him, he replied with genuine excitement: “Gosh. Yes [emotional vocalization], that i[s] it”.

Figure 9.

GenAI-Generated image in co-creation with Peter. Note. Image generated by DALL·E3 on October 23rd, 2024

Technical Difficulties in Co-Creating With LLM-Based Text-To-Visual Generators

During several co-creation sessions, we encountered technical difficulties when working with GenAI. At times, we had to prompt ChatGPT multiple times before a specific object appeared in the image, even though the chatbot indicated that it had already been included. For example, in the creation of Figure 5B, depicting a butterfly proved difficult – it only appeared after seven prompting attempts.

Because certain elements were symbolically important to co-creators, we explored alternative technological solutions whenever ChatGPT was unable to generate them. For instance, we used Adobe Photoshop in combination with Adobe Firefly to manually add missing elements. This approach was necessary for Figure 7, where an additional bouquet of flowers had a symbolic meaning for the co-creator, as they reminded her of the support she had received from her friends.

We also encountered uncertainties about where the boundaries lie of what GenAI is allowed to generate. For instance, in Figure 4A, Walter requested “an image of a man doing action painting in the style of Jackson Pollock”. The chatbot refused, citing regulatory and copyright restrictions. After “in the style of Jackson Pollock” was omitted, the image was successfully generated. However, generating an image of Rembrandt or Picasso painting on the persona’s T-shirt did not pose any difficulties.

Liminal Shifts in LLM-Based Text-To-Visual Generators

On March 20, 2025, during the transition from DALL·E 3 to the more integrated and multimodal ChatGPT-4o model (OpenAI, 2025; Wittig et al., 2025), we encountered challenges in image generation. ChatGPT-4o began producing more photorealistic, less surreal visuals. This liminal shift required us to adjust our prompting strategies and methodology, paying closer attention to prompt components such as persona description, style and atmosphere to generate images that resonated with the lived experiences of the co-creators.

Based on our GenAI-facilitated co-creation sessions, DALL·E 3 appeared to be better suited for conveying complex visual and often surreal narratives within a single image, while ChatGPT-4o proved more effective in generating lifelike, photorealistic representations. Furthermore, DALL·E 3 supported a quicker co-creation of initial scene requests and narrative alignment that resonated with the co-creators. Although the photorealistic representations enhanced co-creators’ sense of recognition and identification with the images, they also introduced new challenges regarding the use and publication of such visuals due to the high degree of personal identifiability. That’s why we returned to DALL·E 3, as its outputs best aligned with the co-creators’ stories and lived worlds. Google’s Nano Banana Pro also appeared to be a suitable tool for this purpose, as it seemed to combine a surreal tone with realistic representation. However, we draw this conclusion based on two images, as this tool only became available toward the end of our study, in November 2025.

Discussion

This study aimed to explore the potential of integrating GenAI-based image generation into a research design to gain insight into the lived experience of PLWD by conducting GenAI-facilitated co-creation sessions. These sessions demonstrated that it is possible to effectively convey lived experience, provided the persona resembles the co-creator and the scenery evokes emotion through recognizable elements such as color, style, or familiar situations. Overall, we conclude that co-creating with GenAI and PLWD offers a complementary and effective addition to other participatory visual methods, such as photovoice and photo elicitation.

Strengths

A key advantage of working with GenAI lies in its ability to generate images in a targeted, efficient, and creative manner, while maintaining transparency and active participation with PLWD as co-creators. Although feedback in the co-creation process was mainly provided verbally within an interview setting, we observed strong engagement from all co-creators. Notably, it was possible to generate powerful images that closely aligned with the lived experiences of less verbal individuals, relying primarily on body language and vocalizations. The real-time generation of images also allowed the visuals to serve as a guiding framework for the interview itself. Co-creators were visually reminded of the central theme of the conversation, which helped reduce unrelated storytelling. When co-creators began to drift into thematically unrelated or incoherent narratives, the evolving on-screen images often naturally redirected their attention. They would become curious about the system’s actions and spontaneously return to discussing the visuals, suggesting changes or adding experiences they wanted to include. Thus, this tool helped sustain attention, structure thoughts and anchor the conversation. Based on documented reactions of the co-creators, we noticed a sense of feeling recognized during the development of the image.

Most co-creators and their caregivers were unfamiliar with GenAI and expressed curiosity and fascination when introduced to the technology. The novelty often evoked a sense of wonder and pride among co-creators, which helped build rapport and made the sessions more engaging. Furthermore, this dynamic shifted the tone of the interaction toward collaboration, increasing both the quality of engagement and the richness of the data collected (Maund et al., 2022). The approach demonstrates that participation in research, even on sensitive health topics, can be a creative and enjoyable experience.

GenAI can produce surreal or unexpected results, often described as ‘hallucinations’ in this context (Huang et al., 2024). While these unintended outputs necessitate a critical approach to GenAI usage, they sometimes resonated with co-creators’ experiences, particularly those who reported altered perceptions or temporal disorientation. In several instances, such images were embraced by the co-creators and prompted deeper reflection. In this setting, GenAI functions not only as a tool, but also as an agent that can redirect interpretation and prompt new narrative alignments within the session without implying autonomy. Therefore, the co-design process benefits from recognizing GenAI’s agency as part of the entangled co-creative arrangement. Following Nunes (2024), we use the term “Agential Art”' to describe artworks co-created with GenAI, where agency is understood as distributed and emergent within the socio-technical arrangement, rather than located in a single human or machine author.

Limitations

This study’s sample consisted of co-creators residing in high-quality care facilities and co-creators with high socioeconomic statuses and intellectual backgrounds. Therefore, the findings of this study may be mainly applicable to this specific group. Besides, these findings may not apply to individuals in more advanced stages of dementia since this research method requires relatively abstract and complex reasoning. This reasoning involves interpreting an image, registering it, analyzing it, comparing it to one’s lived experience, and formulating feedback. Completing this cognitively demanding sequence of tasks requires abilities that PLWD gradually lose.

Finally, working with AI itself also introduces several limitations, arising from the fact that LLM-based text-to-visual generation tools depend on various external factors, such as access to electricity and a stable data network. Moreover, these tools are continuously being updated. Such liminal shifts can affect both the workflow and the visual outcomes, making the co-creation process less predictable.

Ethical Concerns and Responsible GenAI

Despite the possibilities this method may offer, several ethical concerns should also be considered. First, the co-created images are based on personal experiences, thoughts, emotions, and insights. Because these images are grounded in such personal – and sometimes even intimate – information, they may contain sensitive content that people in the co-creator’s immediate social context are not necessarily aware of. The fact that these potentially private moments can be quickly visualized through GenAI may pose a threat to the co-creators’ personal privacy and sense of intimacy (Al-kfairy et al., 2024). It is therefore necessary to implement strict pseudonymization and data minimization measures, and to ensure that confidential data are neither used for model training nor retained for any subsequent training or system improvement.

Second, co-creators often aimed for personas that closely visually resembled them. When these personas are rendered in a photorealistic style, the distinction between an AI-generated image and a real photograph becomes increasingly difficult to make. As a result, there is a risk that co-creators or others may no longer view these images as (symbolic) representations of feelings or lived experiences, but rather as accurate, documentary-like depictions (Al-kfairy et al., 2024). In cases where co-creators forget that they themselves created the image with GenAI, such visuals could potentially lead to the formation of false memories (Pataranutaporn et al., 2025). This risk is less pronounced with more graphic or illustrative images, as their stylized nature makes it more evident that they are creative constructs rather than real depictions (Hausken, 2024). In GenAI-facilitated co-creation sessions, researchers are thus expected to take an active role in upholding ethical principles and addressing ethical challenges, functioning as a kind of ethical guardian throughout the process (Nicholas et al., 2019).

Third, this method involves multiple researchers, each bringing their own values, perspectives, and unconscious biases, which can influence the design process (Bormans & Van Gorp, 2025). This may lead to misrepresentation or overlooking certain aspects of the co-creators’ lived experiences (Nicholas et al., 2019). It is therefore important to remain aware of the biases that may emerge when translating narratives into prompts, as these may influence the way the images are shaped (Al-kfairy et al., 2024). Thus, in this context, the role of the researcher is to help make co-creators’ feelings and interpretations visible, acting as a facilitator of the co-creation process. This involves being attentive to potential power dynamics and making them open for discussion (Greenhalgh et al., 2016). For this reason, we recommend co-creating in a (research) team, allowing collective reflection and critically monitoring one another’s assumptions and biases.

Fourth, we must consider infringement of copyrights belonging to artists and designers. Their images and visual styles are often used as training data for GenAI models without recognition or compensation (Centivany, 2024; Nunes, 2024). A striking example of this occurred with the release of ChatGPT-4o, which introduced significantly more lenient copyright rules. This made it possible to generate works in the style of specific artists and studios, such as Studio Ghibli, without their permission or remuneration (Knodel & Hingle, 2025).

Lastly, it is important to consider the environmental impact of generating visual representations, especially when iterative feedback loops allow for multiple variations of the same image (Li et al., 2023; Sasha; Luccioni et al., 2024). We provide an overview of the water and energy consumption associated with generating images in our study to raise awareness about this issue (Table 2). Given this environmental impact (Brevini, 2020; Chen, 2025), it is essential to reflect on efficient ways of generating visuals with thoughtful prompts, aiming to minimize the number of discarded images.

Conclusions

This study offers insight into how the integration of GenAI-based image generation within a research design can support the visualization of the lived experience of PLWD. GenAI-facilitated co-creative sessions show strong potential for visualizing feelings or experiences that are difficult to express verbally. Moreover, the co-creative aspect helps counteract biases and stereotypical representations that are still present in GenAI images.

Using GenAI-generated visuals to express themes that are difficult to express verbally may not only be highly viable for PLWD, but also prove valuable in other settings, like schools working with immigrant or refugee children, or in therapy with young people affected by war (Sullivan & Simonson, 2016). Besides helping people express their feelings, the creation and discussions about images can support emotional processing, while keeping a safe distance from difficult or potentially traumatic experiences (Kevers et al., 2022).

Given that GenAI has become deeply integrated into our society and its use can no longer be excluded, the focus should shift toward promoting responsible use (Bhavsar et al., 2025). This study aligns with this perspective by prioritizing responsible AI use, particularly through human oversight and regulation (Kaur et al., 2025). However, it is important to critically reflect on the ethical concerns on this methodology, and we advocate the development and further application of responsible AI (Gu, 2024). We therefore call for the development of clear guidelines that can help address the ecological and ethical challenges of GenAI. We believe it is important to emphasize that generating images that depict specific communities, or social groups should always take place in interaction, ideally through dialogue or co-creation with the people represented. We should not place blind trust in GenAI or LLMs when it comes to visual representation and remains vigilant about potential bias. The development of prompt guides and image creation practices tailored to specific communities, audiences, and purposes is therefore worth further investment.

In addition, we suggest that education includes sufficient attention for developing both digital and visual literacy, to support critical and informed engagement with GenAI-generated media. Therefore, we consider it important to communicate openly about the prompts used to generate images and other media. Initiatives such as the https://betterimagesofai.org/project, developed by Dihal and Duarte (2023), which communicate transparently and clearly about prompt usage to generate better images of AI, are therefore highly valuable. Furthermore, we think it is essential to communicate more transparently about the energy and water usage of AI systems, so that we can use this technology more consciously and better assess its material and ecological impact. Initiatives such as an AI-energy score on the Hugging Face platform (Luccioni et al., 2024), which aim to promote greater transparency through an open framework and a star-rating system for energy consumption, could be certainly meaningful in this context.

Possible responses for the infringement of copyrights of GenAI and LLM’s could include implementing a consent-based system, where creators can explicitly indicate whether their work may be used for training purposes. Such a framework could provide a valuable model for these issues, emphasizing transparency and giving creators control over how their work is used, or credited (Centivany, 2024; Nunes, 2024). At the same time, we recognize that consent-based approaches do not resolve past infringements already incorporated into trained models. This underscores the need for broader sector-wide solutions—legal, economic, and infrastructural—that can support meaningful redress and more equitable redistribution of benefits to creators.

By amplifying the voices of PLWD through responsible GenAI-facilitated co-created imagery, we not only challenge the biases embedded in these visuals but also contribute to a more inclusive future that centers and empowers PLWD. At the same time, we argue that a participatory live prompting methodology holds broader potential in other complex social contexts and communities, enabling the visual articulation of difficult, often inexpressible, lived experiences.

Supplemental Material

Supplemental Material -Co-Creating GenAI-Generated Visualizations of People Living With Dementia: A Participatory Text-To-Image Method and Case-Based Evaluation

Supplemental Material for Co-Creating GenAI-Generated Visualizations of People Living With Dementia: A Participatory Text-To-Image Method and Case-Based Evaluation by Kristof Vrancken, Lisa Bormans, Baldwin Van Gorp in International Journal of Qualitative Methods.