“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 repeat users of generative AI for emotional and mental health support one open question: "What would ChatGPT be like in your imagination?" From 4,387 Prolific respondents screened in April 2024, 270 people across 29 countries answered, 38% of them South African, 27% with at least one mental health diagnosis. Thematic analysis produced two meta-themes: an ontological identity, the "what" GenAI is imagined to be, with 11 subthemes across 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 to "a small god." A post-hoc count found 33% imagining human characters, 26% anthropomorphic attributes and 15% an omniscient being, which the authors read as comfort and as a pathway to overtrust, dependence and bias wherever Learners turn to chatbots for support.

Key Findings

1. A cross-cultural sample of existing users. Of 4,387 people pre-screened on Prolific in April 2024, 270 repeat ChatGPT users for emotional support across 29 countries were analyzed, 38% South African.
2. The sample carried real mental health load. Mean age was 30.06; 57.8% identified as female, 41% as Black, 52.2% as Christian; 27% reported a diagnosis and 58.2% had been in therapy.
3. Meta-theme one: ontological identity, 11 forms in four quadrants. Crossing human and non-human with abstract and concrete axes yielded spiritual beings, mind-like presences, holograms, robots and fictional characters.
4. Meta-theme two: functional characterization, 14 attributes in five groups. Looking, feeling, knowing, doing and relating covered appearance, emotion, supernatural cognition, competence, and the relational qualities of trust, integrity, independence and affability.
5. The spiritual register gave the paper its title. One participant imagined "a small god that answers and gives direction to almost every question."
6. Post-hoc counts quantify the anthropomorphism. 90 responses (33%) imagined a human character, 71 (26%) used anthropomorphic attributes and 40 (15%) described an omniscient being.

How the data were collected and coded

It is a convergent parallel mixed-method survey reporting only its qualitative strand, so read it as a vocabulary of user expectations rather than a measurement: there is no outcome measure, usage log or follow-up. Two master's-level counseling trainees coded independently under a licensed psychologist, and codes naming entity types and traits separately became an identity taxonomy plus an attribute framework for large language models.

What GenAI is imagined to be: the ontological identity

Users' mental models travel far from the product category, so plain talk about a tool meets imaginings of persons: a virtual assistant that "makes up for my lack of knowledge," an "old wise woman," "a bunch of nerds behind the screen," or, on the non-human side, server racks, a video-game hologram and something "like energy, not an entity." The abstract-human quadrant matters most for Well-Being, where spiritual beings become spirit guardians or gods.

What GenAI is imagined to do: the functional characterization

This framework tells instructors what users expect the system to be good for: appearance, cognition split between the ordinary and the supernatural, agency as competence and speed, and emotional ability hedged ("programmed to deal with human emotions"). Relating was the densest theme and supplies the vocabulary practitioners need: integrity and stability, trust and safety, independence, affability. Whether someone engages a chatbot as companion or as educator therefore changes what they count as it working.

Culture, the small God, and the risk side

The theme structure held across 29 countries, but imagery borrowed whatever cultural material was available, religion for the small god, holograms from games. Gendered images such as "an intelligent male who wears a suit and is highly educated" echo a 2019 UNESCO report on gender bias, so comparative sampling matters more than generalization. The clinical risk is overdependence and misplaced Trust, sharpest where users describe a mind "in my head," paired with Privacy and affective substitution for human support.

What this implies for educational use, and the limitations

Schools and counseling services are where Learners meet chatbots, imaginatively, before any pedagogical frame arrives: a student who imagines an omniscient friend reads unlimited availability as reliability and fluent confidence as knowledge, the habit Trust Calibration exists to correct. Because a reassuring reply to a disclosure of distress is a relational Hallucination Risk, Guardrails must cover escalation to humans, and stereotype-laden imaginings make this a design and equity question Bias Mitigation cannot fix later. The sample's spread across the global South argues for Pedagogical Safety by design and for Social-Emotional Learning and culturally responsive practice.

What this means for practice

• Instructors. Have students name what they imagine the tool to be before discussing what it does: 90 of the 270 respondents (33%) imagined a human character.
• Instructors. Build limits into the deployment itself, visible uncertainty, refusals and explicit boundaries, since unlimited availability and fluent confidence are read as reliability by users imagining an omniscient entity, the habit Trust Calibration exists to correct.
• Administrators. Require human escalation paths in student-facing support deployments, not content filtering alone: a reassuring reply to a distress disclosure is a relational harm.
• Designers. Audit persona defaults, naming, voices and gendered imagery before launch: participants reproduced status-laden images that interface choices can repeat or interrupt.

Limitations

• All 270 participants were recruited through Prolific and reported repeated ChatGPT use; 38% came from South Africa and 52.2% identified as Christian, so the deified register may reflect this sample.
• Imaginations were elicited by one open-ended item in a survey whose qualitative strand alone is reported, so nothing links an imagined role to dependence, Help-Seeking, Self-Efficacy or Well-Being.
• Platforms were self-selected rather than assigned, so interface and persona design are confounded with the reported imaginations.
• Of 334 respondents who completed the questionnaires, 64 failed a validity check, and two master's-level trainees coded the attribute framework, so the taxonomy may reflect the prompt as much as participants' imagery.

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.