Abstract
Does the way we picture our ideal life reveal what we value, who we are, and how we thrive? In a preregistered study (N = 273), we tested whether free-form descriptions (median 109 words) and seven-keyword lists of one's ideal life predict a broad battery of psychological constructs: the Big Five, the ten Schwartz values, social dominance, political orientation, meaning in life, and subjective well-being. Using regularized linear probes on embeddings from transformer large language models (LLMs), inspired by model-interpretability research, we find that ideal-life narratives predict most constructs above chance in both conditions (mean r = 0.24 for free-form, mean r = 0.12 for keywords). This free-form advantage was concentrated on constructs people rarely name outright, such as achievement, benevolence, conformity, and emotional stability, implying that part of the signal lies in the manner of the telling rather than in the referents alone. Nineteen of the 24 constructs were recovered from roughly 120 words of text, a far broader battery than earlier language-based work, which has relied on much longer samples and focused on personality and well-being. Our approach offers an open, privacy-preserving alternative to zero-shot commercial LLMs, with comparable accuracy, no API cost, and freely released data and code.