Hipnotalamusz/AI_Assisted_Self_Images_With_Prompts_And_Personality_Tests

Digital Mirror of the Soul - AI-Assisted Self-Images with Prompts and Psychological Questionnaires

This dataset originates from a study that examines the intersection of artificial intelligence, psychology, and art. It provides a comprehensive collection of AI-generated images and textual prompts from participants engaging in a task designed to express their self-image. This work is ideal for researchers in the fields of psychology, artificial intelligence, and art therapy, offering a novel dataset for exploring self-representation and the psychological dimensions of AI-assisted art creation.

Dataset Details

Dataset Description

This dataset comprises 18,219 images and 6,519 textual prompts created by 153 participants using the Midjourney v.4 and subsequently upgraded to v.5 AI software. Participants were tasked to create images that they believe are reflective of their personality, with a creation window limited to 45 minutes, meaning each entry is a participant's attempt to visualize aspects of their self-perception. In addition to image creation, participants completed a series of psychological questionnaires, detailed further in the description. They also engaged in a 15-20 minute interview with a psychology student, discussing the creation process, their images, and along with any thoughts, feelings, or memories evoked during the procedure. This dataset, containing the images, prompts, and results of various psychological questionnaires, supports a variety of research objectives, including the development of models to analyze visual and verbal self-expression, their development and temporal changes over the image creating session, and their potential use in inferring relevant psychological constructs ranging from body image, perfectionism, self-esteem, stability of identity, and BIG5 personality traits.

• Curated by: Klaus Kellerwessel (0009-0005-6420-5691) and Lilla Juhász
• Funded by: ÚNKP-23-2 New National Excellence Program of the Ministry for Culture and Innovation from the source of the Hungarian National Research, Development, and Innovation Fund.
• Language(s) (NLP): English

Dataset Sources [Optional]

• Repository: https://huggingface.co/datasets/Hipnotalamusz/AI_Assisted_Self_Images_With_Prompts_And_Personality_Tests
• Paper [Optional]: Under revision - we will add it later

Uses

The dataset is intended for academic researchers and practitioners interested in the cross-disciplinary areas of AI, psychology, and art therapy. It offers a unique dataset for studying the nuances of self-representation, providing a basis for both quantitative and qualitative analyses. Researchers interested in machine learning, gender studies, and the psychological impact of AI on art creation will find this dataset particularly useful. It facilitates a deeper understanding of the role of AI in art therapy practices and the broader implications for psychological research.

Direct Use

• Developing AI-driven psychological assessment tools that interpret visual and textual data.
• Investigating the nuances of identity expression through digital art.
• Enhancing art therapy practices with AI technology.

Out-of-Scope Use

The dataset is designed for scholarly research and is not intended for commercial use or any applications that could compromise the privacy or anonymity of the participants. Ethical guidelines should be strictly followed to ensure respectful and responsible use of the data.

Dataset Structure

The dataset is structured with comprehensive metadata for each participant's AI-generated images and textual prompts. Here's a guide to navigating and utilizing this rich dataset:

Overview

The dataset includes multiple rows for each participant, where each row corresponds to an image generated from a single text prompt. The columns encompass demographic information, psychological assessments, and details related to the image creation process.

Columns Description

• Participant_ID: Each participant created a unique identifier for themselves. This was necessary because participants filled in the questionnaires online one day before the image creation procedure, and we needed to connect the questionnaire results to the images and interviews somehow. We opted not to generate an ID from the participants' names due to privacy concerns, and this self-chosen approach seemed to work. Since the participants were all Hungarian, they sometimes used Hungarian words like "kémény" (chimney), "teknős" (turtle), and some in English (e.g., "soviet cat") or seemingly nonsensical (e.g., "t2ki7m") IDs also appear.
• Image number: A sequential number indicating the order of the image generated by the participant. Images generated from the same prompt at the same session (see Event type) share their Image number.
• Text Prompt: The textual description provided by the participant to generate the image in natural language.
• Event type: Describes the nature of the image generation event: "imagine" for initial creations, "variation" for variations of an initial image, and "upscale" for a bigger and more detailed version of an initial image. The Midjourney program generates 4 images per 1 prompt for the initial image type "imagine" and their "variations", resulting in these images sharing image numbers in groups of 4 (see Image numbers).
• file_name: The name of the file corresponding to the generated image, encapsulating the participant ID, image number, and the first 40 characters of the text prompt used. We also added some random characters to the end to ensure that every image's file_name is unique.
• Gender: Participant's gender (1 indicates female, 0 male - we will add the data of our non-binary participants later).
• Age: Participant's age.
• Highest level of education: Coded value representing the participant's highest level of education achieved.

And various psychological measures, including:

• Rosenberg self-esteem The Rosenberg Self-Esteem Scale is a widely used tool for assessing an individual's self-esteem. It consists of 10 items designed to measure both positive and negative feelings about the self. The scale is scored on a four-point Likert scale, ranging from strongly agree to strongly disagree, with higher scores indicating higher self-esteem. Participants with high self-concept clarity might produce images and prompts that are more consistent and coherent, reflecting a stable and well-defined sense of self. (Rosenberg Self-esteem Scale: Horváth, Zs., Urbán, R., Kökönyei, Gy., Demetrovics, Zs. (2022). Kérdőíves módszerek a klinikai és egészségpszichológiai kutatásban és gyakorlatban I. Budapest: Medicina könyvkiadó.)
• Extraversion/Big5 to Openness/Big5 This shortened version of the Big Five Personality Test measures five key dimensions of personality: Extraversion, Agreeableness, Conscientiousness, Neuroticism, and Openness. Each dimension is assessed with two items, offering a brief yet effective insight into an individual's personality traits. The Big Five traits could offer a nuanced understanding of the themes and motifs chosen in the art creation process. For instance, high Openness might be linked to more creative and diverse prompts, while high Extraversion could relate to more socially engaging or dynamic content. (10-item Personality Inventory (Big5 – shortened version): Chiorri, C., Bracco, F., Piccinno, T., Modafferi, C., & Battini, V. (2015). Psychometric properties of a revised version of the Ten Item Personality Inventory. European Journal of Psychological Assessment.)
• Self-concept clarity The Self-Concept Clarity Scale assesses the extent to which an individual's self-concept is clearly and confidently defined, internally consistent, and stable over time. High scores indicate a clear and confident self-concept. Participants with high self-concept clarity might produce images and prompts that are more consistent and coherent, reflecting a stable and well-defined sense of self. (Self-Concept Clarity Sale: Hargitai, R., Rózsa, S., Hupuczi, E., Birkás, B., Hartung, I., Hartungné Somlai, E., ... & Kállai, J. (2021). Énkép egyértelműség mérése és korrelátumai. Magyar Pszichológiai Szemle, 75(4), 557-580.)
• Beck depression The Beck Depression Inventory is a 21-item self-report inventory, one of the most widely used instruments for measuring the severity of depression. Each item is scored on a scale from 0 to 3, with higher total scores indicating more severe depressive symptoms. Depression scores could influence the emotional tone of the generated images and prompts. Higher scores might be associated with themes of sadness, isolation, or other negative emotional expressions. (Beck Depression Inventory (BDI): 75 Papír-Ceruza teszt. Pszicho-ped Bt. - https://animula.hu/konyv/75-papir-ceruza-teszt )
• Interpersonal.../Ego Identity Status to Ideological achieved identity/Ego Identity Status This assessment tool measures Ego Identity Status across different domains, including interpersonal relations and ideological commitments. It categorizes identity status into diffusion, foreclosure, moratorium, and achievement, providing insight into the individual's identity exploration and commitment processes. Identity status may impact the thematic diversity and depth of participants' creations. Those in the achievement status might exhibit a greater variety of themes, reflecting a well-explored sense of identity. (Extended Objective Measure of EGO Identity Status II. /EOM-EIS II.: Jámbori, Sz., Kőrössy, J. (2019). A szándékos önsza-bályozás jelentősége serdülő és fiatal felnőttkorban a társas támogatás, az identitásállapotok és a reziliencia tükrében. Alkalma-zott Pszichológia 19(3): 33-52.)
• Standards/Perfectionism to Discrepancy/Perfectionism This scale assesses perfectionism by measuring standards and discrepancy aspects. High standards reflect the setting of high personal performance standards, while discrepancy refers to perceived shortcomings in meeting those standards. Perfectionism scores, especially high discrepancy, might relate to how participants critique their own creations or the iterative process of refining their images through variations. (Almost Perfect Scale (perfectionism): Horváth, Zs., Urbán, R., Kökönyei, Gy., Demetrovics, Zs. (2022). Kérdőíves módszerek a klinikai és egészségpszichológiai kutatásban és gyakorlatban I. Budapest: Medicina könyvkiadó.)
• Total/Body Image to Rest/Body image This questionnaire assesses body image across several dimensions, including general satisfaction with one's body, evaluation of body size, knowledge about one's body, and attitudes toward specific body parts or aspects. Body image scores could influence how participants choose to represent themselves or others in their images. Issues with body image might lead to avoiding personal representation or altering aspects of appearance in the generated art. (Personal body attitudes questionnaire: Horváth, Zs., Urbán, R., Kökönyei, Gy., Demetrovics, Zs. (2022). Kérdőíves módszerek a klinikai és egészségpszichológiai kutatásban és gyakorlatban I. Budapest: Medicina könyvkiadó.)

Navigating and Utilizing the Dataset

• Participant Analysis: Isolate data for individual participants using the Participant_ID column for qualitative case studies or aggregate data across participants for broader analyses.
• Image Type and Number: Key to understanding the context of each image. Variations or upscales from an image might reflect a participant's preference for continuing to work on this particular image, assuming that the upscaled and variated images might contain more useful information than the more accidental "imagine" types. The image number enables us to investigate the dynamic processes of image creation, the shifts of focus, and experimentations.
• Text Prompts: Explore thematic (categories, topics used in self-description) or formal (linguistic, stylistic) features of the prompts, their changes in the creation process, or correlations with the psychological measures provided.
• Images: Analyze thematic (content of the image) or formal (colors, brightness, composition, edge density, etc.) features of the images to infer different psychological and demographic data.
• Statistical and Machine Learning Analyses: Use the dataset for both traditional statistical analyses and advanced machine learning models (both supervised and unsupervised) to explore underlying patterns and correlations in how participants express themselves through AI-generated art.

Dataset Creation

Curation Rationale

As passive recipients, we often differentiate sharply between AI and human-made artworks. However, studies show that when individuals personally engage with AI art softwares, they tend to view the creative process as collaborative, feeling a sense of ownership over the finished works. This type of involvement and sense of personal connection can serve as a basis for the use of these tools in art therapy and even psychometrics. This dataset primarily investigates how individuals can express complex psychological states through AI-generated art - and whether it is possible to deduce them from only the images and the prompts used.

Source Data

Data was collected in a controlled study environment, with participants guided through the process of creating AI-assisted art.

Data Collection and Processing

The test session, which lasts between 80 and 90 minutes, involves only the examiner and the participant. The procedure begins with a brief tutorial on the image-generation software (Midjourney®), where an example image is generated using a 'Dogs and flowers' prompt. After obtaining informed consent, participants are instructed to take self-portraits for 45 minutes, using natural language prompts as guided by the following standardized instruction:

"I would like to ask you to try to take pictures of yourself that express who you are and how you feel about yourself. These images do not have to be lifelike, but they can be. They can be based on several basic ideas, embedded in scenes or situations, and can deviate from reality as much as you like. During the image creation process, you can specify artistic styles, play with the format, composition, lighting, and colours. The goal is to produce as many images as possible that you feel capture something of your personality."

During image creation, participants are encouraged to ask technical questions and may use the online DeepL translator to overcome language barriers, given that the participants were Hungarian and the image creation process was conducted in English. However, they receive no further assistance; for example, they cannot use a mirror or a specific image found on their phone as a template. The 45 minutes starts after the first successfully completed image pack is loaded, and the examiner gives a signal five minutes before it runs out. Upon completing the 45-minute period, participants are allowed to finish the prompt they have already started but are not permitted to initiate new ones. The image-making phase was followed by a 15-20 minute semi-structured interview, guided by the following questions:

• How are you feeling now?
• What was it like to go through the task?
• Was the 45-minute duration enough, or did you feel it was too much or too little?
• Did you experience a state of flow? Were you able to settle in?
• What goals did you set for yourself?
• Did you have any strategy or did you simply allow yourself to associate freely?
• How difficult was it for you to try to define yourself for 45 minutes?
• How do you relate to the finished images?
• Which ones do you feel are the most expressive of yourself?
• Which ones the least?
• Which one do you think your best friend would find the most expressive of you?
• How does it feel to look through the pictures now?
• Have you had any realisations in terms of self-knowledge?
• Did you encounter any drawbacks or difficulties in using the program or during the test-taking process?
• What would you change if you could start over?

Who are the Source Data Producers?

All of our participants were Hungarian young adults between 18 and 28 years old. They have been anonymized, with their questionnaires, interviews, and images linked solely through the unique ID each participant selected.

Bias, Risks, and Limitations

The dataset's interpretations should be made with caution, considering the socio-cultural context of the participants and the influence of AI technology on artistic expression. Ethical considerations are paramount, especially concerning participant privacy and the interpretation of artistic expressions.

Recommendations

Researchers are encouraged to approach the dataset with a multidisciplinary perspective, integrating insights from psychology, artificial intelligence, and art theory. This dataset offers a unique opportunity to explore the boundaries of AI-mediated human expression and its implications for psychological research and practice. Careful, ethical analysis can lead to significant advancements in our understanding of AI as a tool for self-exploration and expression in both clinical and research settings.

Citation [Optional]

APA: Kellerwessel, K. (2024). AI_Assisted_Self_Images_With_Prompts_And_Personality_Tests (Revision cfbe255) [Adatbázis]. Hugging Face. https://doi.org/10.57967/hf/1942

Dataset Card Authors [Optional]

Klaus Kellerwessel (Eötvös Loránd Tudományegyetem, Budapest; University of Pannonia, Veszprém)

Dataset Card Contact

[email protected]