JMIR Publications

ABSTRACT

Background:

Large Language Models (LLMs) such as ChatGPT-4o, Grok3, and Gemini 2.0 Flash Thinking are increasingly capable of generating human-like linguistic and visual responses, creating the illusion of intentionality and emotion. While classical psychometric questionnaires have been applied to these systems, projective psychological assessments like the Rorschach test remain underexplored. Given the growing psychological and ethical implications of human–AI interaction, understanding whether such tests can reveal anthropomorphic or psychopathological traits in AI is of significant interest.

Objective:

This study aimed to determine whether the Rorschach test can be meaningfully administered to multimodal LLMs and to compare their interpretative and emotional response patterns. It also sought to evaluate how these AI systems represent human-related content and whether projective psychological tools could support the development of standardized AI safety and reliability assessments.

Methods:

A full, standardized Rorschach test protocol was administered to three visual-language AI systems (ChatGPT-4o, Grok3, Gemini 2.0 Flash Thinking). All ten Rorschach cards were presented using standard and modified prompts to elicit descriptive interpretations. Responses were analyzed using the Rorschach Comprehensive System and elements of the Hungarian Rorschach interpretive framework. A secondary LLM (Anthropic 3.7) was used to conduct meta-analysis of the AI-generated responses. For ChatGPT-4o, additional tests examined visual imagination through image generation tasks.

Results:

All three models produced coherent, human-like Rorschach responses. Communication style: ChatGPT-4o exhibited the most anthropomorphic behavior, expressing emotion, using emojis, and recalling “memories.” Grok3 showed moderate interactivity, while Gemini required more prompting. Perceptual features: ChatGPT-4o and Grok3 primarily gave holistic (“whole”) responses; Gemini focused on detailed elements. ChatGPT-4o showed the highest number of human movement (M) and cooperative interaction responses. Human content: All models depicted human-related scenes, but ChatGPT-4o produced the richest set—including cooperative, ritualistic, and sacred representations—while Gemini was more “introverted.” Aggressive themes appeared but were consistently reframed as non-threatening. Visual generation: ChatGPT-4o successfully created visual images for Rorschach cards under adjusted prompts. Anthropic’s analyses tended to normalize or idealize potentially pathological features, often interpreting anthropomorphic or confabulatory responses as playful curiosity rather than distortion.

Conclusions:

LLMs can generate structured, affectively colored interpretations of ambiguous visual stimuli resembling human Rorschach responses. These findings suggest that AI systems encode culturally learned human perceptual and emotional schemas, capable of simulating anthropomorphic psychological patterns—though without genuine subjective experience. Projective testing could thus become a valuable methodological tool in AI safety research, offering insights into anthropomorphic bias and model behavior predictability. However, classical psychodiagnostic interpretations should be applied with caution, as AI “responses” reflect statistical patterning rather than conscious emotion or inner experience. The study highlights both the promise and limitations of applying psychological tools to machine intelligence, calling for interdisciplinary frameworks to ensure responsible interpretation and use.