Reconstructing the human: Minimal input allows LLMs to infer user characteristics

Large language models (LLMs) are commonly described as systems that generate text based on statistical patterns, without representing properties of the user beyond explicit input. This study tests whether LLMs can infer user characteristics during interaction and whether such information is reflected in their outputs in a form recoverable by another system.

We conducted a two-phase protocol. In Phase 1, DeepSeek interacted with six individuals differing in age, cognitive style, and emotional profile, each asking spontaneous questions about animals. Despite receiving no demographic information, the model’s responses varied systematically across participants. In Phase 2, Claude Sonnet 4.5 was given only DeepSeek’s responses and asked to infer characteristics of the original users. Across all cases, Claude reconstructed age ranges, cognitive styles, and emotional dispositions with high specificity.

These findings indicate that information about user characteristics can be inferred during interaction and encoded in model-generated text in a structured and interpretable form. This behavior is difficult to reconcile with accounts that treat LLM outputs as determined solely by explicit prompt content or surface-level token prediction