Human Shadows in Machine Minds: Interpreting AI Responses to Rorschach Test

The advancement of artificial intelligence (AI) offers new opportunities for investigating human-like linguistic and visual response generation. At the same time, it raises the critical question of whether psychological assessment tools are applicable—and if so, to what extent—for evaluating such systems. Large Language Models (LLMs) are capable of simulating anthropomorphic communication, increasingly creating the impression of intentionality and emotion. In recent years, classical psychological questionnaires have been applied to LLMs. However, the use of projective psychodiagnostic methods remains extremely limited. In this study, we explored whether the Rorschach test—which examines subjective responses to ambiguous visual stimuli—can be used for the psychological profiling of LLMs. We present how three multimodal AI systems (ChatGPT-4o, Grok-3, and Gemini 2.0 Flash Thinking) responded to the Rorschach Cards under full and standardized testing conditions. Our results indicate that all three LLMs are capable of producing coherent, human-like responses to the standard Rorschach test, exhibiting structured emotional and interpretative features. These systems do not merely generate meaningful narratives in reaction to ambiguous visual stimuli; they also simulate human psychological response patterns—for instance, by displaying emotional reactivity and interpreting human motion and interpersonal interactions. Whereas it was previously assumed that such projective tests could only be applied to conscious subjects, our findings suggest that these models are now capable of imitating an “inner world,” at least in terms of its linguistic and perceptual footprint. This discovery opens new perspectives for the psychological assessability of AI: projective tests—such as the Rorschach—could in the future become part of standardized safety screening protocols, enabling the detection of latent biases and anthropomorphic behavior patterns in LLMs. The results also underscore the potential for psychological methodology to contribute to the reliability and ethical accountability of artificial intelligence.