Interpretations of Tarot Card Spreads by AI: Predictive Versus Introspective Questions

To explore whether AI-generated interpretations of Tarot card spreads are different when predicting external events versus introspection on one's own thoughts. Thirty items made up the balanced test questions for prediction and verification; after querying, card-specific analysis results would be given as a reference. For each of the question-spread pairs, this study obtained five answers from DeepSeek and five from GPT-4, along with a human-divination score classified as "no", "partly yes" or "yes". Convert all responses into numbers (0, 1 or 2) and then calculate the average chatbot's response per question; next, perform a two-factor ANOVA on questions divided by groups and people to answer levels. ANOVA showed a significant main effect of the human answer level (p=0.0017), indicating that AI scores are correlated with human certainty at this point. However, a one-year follow-up of the "partly-yes" group found that there was still a significant difference (P < 0.01). Predictive questions had an average score of 1.58 compared to 0.82 for introspection; the mean difference reached 2.13, indicating it was highly pronounced (Cohen's D). Based on these results, although AI interpretations generally conform to human judgement, question type affects their output under an ambiguous human reference. Predictive Questions may trigger stronger affirmations due to bias in the training data. A new method for measuring sensibility in divination Systems and its impact on the behaviour of AI Advisory System through questions framing is demonstrated here.