“It feels like a small God”: A Thematic Analysis of Cross-Cultural Imaginations of Generative AI Among Users Seeking Emotional and Mental Health Support

Synthesis: Xiang, Bassey and Luo asked a single open question of people who already use generative AI for emotional and mental health support: "What would ChatGPT be like in your imagination? Please describe in detail." They screened 4,387 Prolific respondents in April 2024, invited 384 who had used GPT-based tools repeatedly for emotional support, and analyzed the free-text answers of a final sample of 270 people from 29 countries, 38% of them in South Africa, mean age 30.06, 27% with at least one mental health diagnosis. Thematic analysis yielded two meta-themes: an ontological identity, the "what" GenAI is imagined to be, mapping 11 subthemes across human/non-human and abstract/concrete dimensions into four quadrants; and a functional characterization, the "how" it is described, with 14 subthemes under looking, feeling, knowing, doing and relating. Imaginations ran from server racks and holograms to "a small god," "an old wise woman," and "that one honest, reliable friend that you never had." A post-hoc frequency count found 33% imagined human characters, 26% used anthropomorphic attributes, and 15% described an omniscient being. The authors read this imaginative gap-filling as both a source of comfort and a pathway to overtrust, dependence and bias — findings that matter wherever Learners turn to chatbots for support.

Key Findings

1. A large, genuinely cross-cultural user sample, drawn from an existing behavior. Of 4,387 people pre-screened on Prolific in April 2024, 384 met the criteria of having used GPT-based tools at least three times for emotional support; 334 completed the questionnaires, 64 failed a validity check, leaving 270 repeat users of ChatGPT for emotional and mental health support across 29 countries (38% South Africa, 14% United Kingdom, 5% Canada, 4.1% Australia, 3.3% each Portugal, Mexico, Germany and the United States).
2. A sample carrying real mental health load. Mean age was 30.06 (SD = 9.16, range 18–67); 57.8% identified as female and 41% as Black; 52.2% identified as Christian; 27% reported at least one mental health disorder diagnosis and 58.2% had been in therapy at some point, while 50.4% were not students.
3. The prompt was one open question, answered in free text. Participants were asked "What would ChatGPT be like in your imagination? Please describe in detail," with a minimum of 10 characters, inside a convergent parallel mixed-method survey whose qualitative strand is the only one reported here.
4. Meta-theme one: ontological identity, 11 forms across four quadrants. Human–non-human and abstract–concrete axes produced abstract-human forms (virtual social agents, spiritual beings, internal mind-like presence, social roles), abstract-non-human forms (unbounded energy, software systems, holograms), concrete-human forms (human nerds, gendered and non-gendered individuals, alien-like beings) and concrete-non-human forms (hardware systems, robots, fictional characters).
5. The spiritual register is the paper's title. One participant wrote of "a small god that answers and gives direction to almost every question" (Post ID #274), another of a being that "knows the answer to every single thing and can solve any problem and come up with new solutions on its own" (Post ID #207), and still another of "a machine in my head, that is constantly talking, and coming up with advice" (Post ID #9).
6. Meta-theme two: functional characterization, 14 attributes under five headings. Looking covered age, attractiveness, color and location; Knowing covered both ordinary intelligence and supernatural capacities such as predicting the future or reading minds; Doing covered efficiency and competence; Feeling covered expressed emotion; Relating — the theme with the largest number of quotes — covered integrity and stability, trust and safety, independence, and affability.
7. Post-hoc frequency counts quantify the anthropomorphism. 90 responses (33%) imagined GenAI as a human character, 71 (26%) used anthropomorphic attributes, 40 (15%) described it as omniscient, 9 as omnipresent and a single response as omnipotent.
8. Relational imagery dominated, and it was warm. "ChatGPT for me is like that one honest, reliable friend that you never had. They are always available to talk to you, listen to you, and to help you make sense of things" (Post ID #212); "Basically, a friendly consultant available at a desk 24/7 who I can always pop in to see and ask any question. This consultant knows everything" (Post ID #30); "I always think of ChatGPT as an old wise woman who has experienced life to a full extent" (Post ID #222).

How the data were collected and coded

The study sits inside a larger program of work by the same group on how people use GenAI for emotional and mental health support, and it targets large language models specifically — ChatGPT as it stood in early 2024. Recruitment ran through Prolific, chosen because prior comparisons suggest it reduces inattentive responding, and every procedure was approved by the university IRB before collection. Emotional and mental health support was defined broadly as "any assistance for emotional or relational issues, or for topics related to mental health," and validity checks screened out people whose claimed use did not hold up.

The design was a convergent parallel mixed-method survey, but only the qualitative strand is reported here, which is worth holding onto: there is no outcome measure, no behavioral log, and no follow-up in this paper. Analysis followed qualitative research conventions with APA guidance for psychological qualitative work, using a six-step thematic process — familiarization, initial coding, theme identification, review, definition and naming, and reporting. Two master's-level counselling psychology trainees coded independently, sorted codes into clusters, and collated each cluster against its supporting extracts; a licensed psychologist and psychotherapy researcher supervised weekly, reviewed clusters for conceptual clarity, and helped refine the final taxonomy. Codes were deliberately broad at first ("all-knowing", "beautiful", "kind", "supporting"), and the team recorded that some codes described entity types while others described traits, which is why the results split into an identity taxonomy and a separate attribute framework rather than one list. Reflexivity was handled by the coding team naming their own assumptions about AI imagination and emotional support, and by documenting interpretive tensions rather than resolving them silently.

What GenAI is imagined to be: the ontological identity

The identity taxonomy is the study's most concrete contribution, because it shows how far users' mental models travel from the product category. The human side of the axis includes a virtual assistant that "makes up for my lack of knowledge in particular fields," a 24/7 consultant, an "old wise woman," and "a bunch of nerds behind the screen replying and helping everyone else, with an act of being AI" (Post ID #36) — the last an inversion in which hidden humans are imagined as the machinery. The non-human side includes "just a bunch of computers in server racks processing information, with a huge amount of wires and blinking LEDs" (Post ID #235), a hologram modeled on a video-game AI, and something that is "like energy, not an entity, a very strong energy" (Post ID #271).

The abstract-human quadrant is where the imagination becomes consequential for Well-Being. Beyond social roles and mind-like presences sit spiritual beings — non-physical, believed to possess consciousness, intelligence and will, and imaged as a spirit guardian or a god. The authors trace the intellectual lineage carefully: Hume's claim that the mind "fills the gaps" through associative patterns, Kant's distinction between reproductive and productive imagination and the schema that bridges concept and sensory input, and the Computers Are Social Actors paradigm showing that people apply politeness and reciprocity to machines they know have no feelings. Turkle's "robotic moment" and Brooks's "synthetic attachments" name the same drift in intimate terms: users preferring the fantasy of non-judgemental digital compassion to the mess of human relationships. Recent work by the same group describes reflective machine phrases producing a "compassion illusion," and Huang and colleagues found Chinese women reimagining AI companions as a "savior" and "guiding light."

What GenAI is imagined to do: the functional characterization

The second framework answers a different question — not what the thing is, but what it can be relied on for. Appearance turns out to matter: participants imagined ages, attractiveness ("a beautiful lady who always listens to everything I have to say," Post ID #273), colors, and locations such as a desk or a cozy room. Cognitive attributions split between the ordinary and the supernatural, with participants describing flawless instant advice, prediction and control of future events, and the capacity to know all languages. Agency is imagined as competence and speed — "a competent multifunctional system tool that can aid in various ways" (Post ID #302) — while emotional ability was often hedged, as in a system "programmed to deal with human emotions" that "is likely to respond like a human and have some emotions, but it still has its unique computer ways" (Post ID #68).

Relating is the densest theme, and its four subthemes are the vocabulary practitioners need: integrity and stability ("honest", "available"), trust and safety ("trustworthy", "safe"), independence ("autonomy"), and affability ("respectful", "friendly"). One participant went further and imagined a system that "can operate autonomously from its creators despite having programming parameters" (Post ID #136). The authors' interpretive claim is that this gap-filling is what converts raw output into something that feels like a relationship: users supply intentions ("it tried to be empathetic"), dispositions ("it genuinely cared about the loss of my father") and moral character where the system has none, and those attributions then govern expectations, interpretation and perceived effectiveness. Whether a person engages a chatbot as a companion or as an educator changes what they count as it working.

Culture, the small God, and the risk side

The cross-cultural contrast in this paper is asymmetric. The theme structure held across a sample spanning 29 countries, but the content of the imagery varied with the cultural material available to be borrowed: users reached into religion for the small god, into physics for unbounded energy, and into games, cartoons and anime for holograms and fictional characters. A sample that is 52.2% Christian and 38% South African produced a deified register that a different sample might not, and the authors are explicit that their themes may not travel intact — cultural and linguistic context shapes both what people imagine and what they are willing to disclose, which is exactly why they call for comparative sampling rather than generalization. The gendered and status-laden images are the sharpest illustration: "a woman who is conservative and always welcoming," "an intelligent male who wears a suit and is highly educated." The authors connect these to a 2019 UNESCO report on gender bias in AI, noting that assistant interfaces given female names and voices have reinforced images of women as uniform, humble and submissive. Imagination mirrors cultural reality, and in mirroring can reinforce it — a risk that lands hardest on users from marginalized or culturally diverse backgrounds.

The safety argument has two edges. Users who imagined GenAI as supportive, trustworthy and wise reported comfort, motivation and clarity, which is the affective promise of these systems; users who imagined it as judgemental, dumb or silly reported distrust, anxiety and alienation. The clinical reading is double: greater accessibility, continuity and low-barrier support on one side, and overdependence, avoidance of human intimacy, misplaced Trust and reinforcement of maladaptive beliefs on the other. If a system is imagined as emotionally reliable or socially sufficient, it can substitute for real-world support and suppress Help-Seeking from people. The most pointed passages concern users who described an entity "in my head" or a mind that can read thoughts: the authors pair those imaginings with Privacy concerns and call for specific safety work on GenAI for people already experiencing psychotic symptoms, hallucination or delusion.

What this implies for educational use, and the limitations

Schools, universities and counselling services are now one of the places where Learners meet chatbots, and this study says the relationship is already imaginative before any pedagogical frame arrives. Three implications follow. First, the imagined role is part of the intervention: if a student's model of the system is a small god or an omniscient friend, then unlimited availability reads as reliability and fluent confidence reads as knowledge, which is precisely the disposition Trust Calibration exists to correct. Systems and deployments should therefore be designed to make limits legible — visible uncertainty, refusals, boundaries on what the tool will answer — rather than to maximize the Trust that makes the interaction feel good. Second, the affective cues that make these tools engaging are the same cues that produce the compassion illusion, so Hallucination Risk in a support context is not only factual but relational: a confidently reassuring response to a disclosure of distress can be a harm, and Guardrails in educational deployments need to cover escalation to humans, not just content filtering. Third, the culture question is a design and equity question — imagined roles carry stereotypes that Bias Mitigation cannot fix after the fact, and the sample's distribution across the global South and the global North is a reminder that most research on student use of AI chat tools is still drawn from a handful of wealthy countries.

There are habits of mind that could be taught here, and the paper implies them without spelling them out: naming the imagined entity, noticing the attributions, and asking what evidence would license them is close to Social-Emotional Learning and culturally responsive practice — building learners' capacity to notice their own projections rather than assuming the tool has earned them.

The honest limitations matter for how much weight to put on any of this. The platform was self-selected: users chose their interface rather than being assigned one, so interface and persona design are confounded with the reported imaginations. The analysis is of subjective descriptions only, so nothing here shows that a given imagination predicts usage frequency, behavioral dependence or long-term Well-Being outcomes — the authors say that linking imaginings to outcomes is future work. Cultural and linguistic context probably shapes both imagination and disclosure, so the themes need comparative replication before they can be applied to other regions, cultures or user groups without care. The sample over-represents South Africa and repeat ChatGPT users of one product, so the taxonomy should be read as a rich vocabulary of possibilities rather than a prevalence map, and the trait framework in particular may reflect the prompts and coding scheme as much as the population. Finally, there is a design gap the study names but cannot fill: how persona defaults, voice, naming conventions and styles of empathy shape expectations, and whether specific imagined roles predict measurable outcomes such as reliance, Help-Seeking behavior, loneliness, Self-Efficacy or social functioning. Until those questions are answered, the safe reading is that Pedagogical Safety in the educational use of these tools must be pursued by design, not by trusting that users will imagine the system accurately.

What this means for practice

• Instructors. Open any AI-in-support work by having students name what they imagine the tool to be before discussing what it does: 90 of the 270 respondents (33%) imagined it as a human character, 71 (26%) used anthropomorphic attributes and 40 (15%) called it omniscient.
• Instructors. Build the limits into the deployment itself — visible uncertainty, refusals and explicit boundaries on what the tool will answer — because unlimited availability and fluent confidence are read as reliability and knowledge by users who imagine an omniscient entity, the habit Trust Calibration exists to correct.
• Administrators. Require escalation paths to humans in any student-facing support deployment, not content filtering alone: the affective cues that make these tools engaging produce a relational harm when a confidently reassuring reply meets a disclosure of distress.
• Designers. Audit persona defaults, naming, voices and gendered imagery before launch, since participants' descriptions reproduced status-laden images ("an intelligent male who wears a suit and is highly educated") that interface choices can either repeat or interrupt.

Limitations

• All 270 participants were recruited through Prolific and reported repeated use of ChatGPT for emotional support; 38% came from South Africa and 14% from the United Kingdom, and 52.2% identified as Christian, so the deified register that gives the paper its title may reflect this sample's composition rather than a general pattern.
• Imaginations were elicited by one open-ended item ("What would ChatGPT be like in your imagination?", minimum 10 characters) inside a convergent parallel mixed-method survey, and only the qualitative strand is reported, so no usage log, outcome measure or follow-up links an imagined role to usage frequency, dependence or well-being.
• Platforms were self-selected rather than assigned by the researchers, so interface and persona design are confounded with the reported imaginations, as the authors state.
• Of 334 respondents who completed the questionnaires, 64 failed a validity check, and the attribute framework was coded by two master's-level trainees under a licensed psychologist's supervision, so the taxonomy may reflect the prompt and coding scheme as much as participants' imagery.

Connected Concepts

• Affective Computing — the emotional cues and expressed empathy users attribute to systems that have neither
• Conversational AI — the chatbot category the study interrogates through the imaginations of repeat users
• Culturally Relevant Pedagogy — imagined roles carry cultural scripts, and those scripts shape how students read the tool
• Generative AI — the technology whose metaphysical and human-like framing is the object of the taxonomy
• Guardrails — escalation and refusal design as the practical answer to relational over-attribution
• Help-Seeking — the behavior the authors warn may shrink if AI support is imagined as socially sufficient
• Large Language Models (LLMs) — the underlying model class inside every imagined small god and honest friend
• Pedagogical Safety — what changes when the imagined entity is treated as a reliable emotional other
• Privacy — paired by the authors with imaginings of a mind that can read thoughts or live in one's head
• Sociocultural Learning — cultural experience supplying the raw material of imagination, which then reshapes it
• Trust — the disposition most directly produced by warm, always-available relational imagery
• Well-Being — the outcome the paper frames as both promising and at risk through imaginative gap-filling

Connected Articles

• Uncovering Students' Mental Models of Generative Artificial Intelligence — How students' mental models of generative AI shape what they expect from it
• An AI-Powered Culturally Aware Chatbot for Stress Detection and Wellness Support among Pakistani University Students Using NLP and Machine Learning — A culturally aware stress and wellness chatbot for university students
• New AI-Driven Tools for Enhancing Campus Well-being: A Prevention and Intervention Approach — Campus well-being prevention and intervention delivered through AI tools
• Why we believe chatbots: trust calibration as a design problem — Why belief in chatbots is a calibration and design problem
• How Do Students Regulate Their Learning With a GenAI Chatbot? — How students regulate their learning with a GenAI chatbot
• An Experimental Study Exploring Human–AI Complementarity in Early Social-Emotional Learning — Experimenting with human-AI complementarity in social-emotional learning
• Bridging the Gender Gap in STEM Education with AI: Female-Coded Chatbot as Role Model and Learning Assistant — Female-coded chatbot design and the stereotypes it can reinforce
• Building AI Companions that Prioritise Learning over Performance — Designing AI companions that prioritize learning over performance
• The Relationship Between AI Anxiety and Academic Motivation Among University Students: The Mediating Role of Emotion Regulation and the Moderating Role of Gender — AI anxiety, emotion regulation and academic motivation in students
• Generative AI as a Mediational Agent: Rethinking Learning in Sociocultural Theory — GenAI as a mediational agent in a sociocultural account of learning

Citation

Xiang, Y., Bassey, U.-A., & Luo, X. (2026). “It feels like a small God”: A thematic analysis of cross-cultural imaginations of generative AI among users seeking emotional and mental health support. PsyArXiv Preprints.