To make an AI talk like a real person, does it matter more to show it real conversations or to describe the person?

This explores which features of the conversational material make an AI-simulated human reply resemble what a real person would say. The corpus doesn't answer it in one place, but several notes agree on a few properties. The first is that fidelity comes from real dialog, not from instructions about dialog. A study found that pretrained base models conditioned on short samples of actual conversation predict human responses more accurately, and with more variety, than instruction-tuned assistants told to play a persona. The stated mechanism is task mismatch: training a model to be a helpful assistant systematically degrades its ability to sound like the person on the other side of the chat (Do pretrained models simulate humans better than instruction-tuned assistants?). The content does the work here, and a role-play prompt doesn't substitute for it.

The second property is layered specificity. Realistic synthetic dialogues seem to need three things working together: a narrow subtopic, a persona (Big Five personality variation), and a set of 11 contextual characteristics. With all three, the study reports capturing about 90% of in-domain performance (Can synthetic dialogues become realistic through layered diversity?). Any one layer alone gives generic talk. A related approach conditions the simulator on hidden state at two timescales: a session-level user profile and a turn-level intent. Realism was then checked by crowd workers, by discriminator models, and by matching distributions with classifier ensembles (Can controlled latent variables make LLM user simulators realistic?). What the person is trying to do right now matters as much as who they are.

The third property is consistency, which the model doesn't supply by default. One line of work argues that an LLM never commits to a character. It holds a spread of possible characters and samples one at each generation, so regenerating the same reply gives different outputs, each plausible given the context (Do large language models actually commit to a single character?). That helps explain persona drift. Training user simulators with multi-turn reinforcement learning against three consistency checks (prompt-to-line, line-to-line, and question-and-answer) cut drift by over 55% (Can training user simulators reduce persona drift in dialogue?). The three checks catch different failures: drift within a turn, drift across a conversation, and outright factual contradictions. A believable simulated human has to stay the same person from start to finish.

The fourth property is human messiness, and this part is my inference from the notes rather than a direct finding. Preference-tuned models produce 77.5% fewer grounding acts than humans, meaning fewer clarifying questions and understanding checks (Does preference optimization harm conversational understanding?). Standard reward training also pushes models toward passive answers instead of asking what the person means (Why do language models respond passively instead of asking clarifying questions?). A simulator built from assistant-style outputs would therefore inherit the wrong conversational habits. Emotion also matters: one system uses a simulated user's emotion trajectory across turns as its training signal, so the simulator's feelings are part of what makes it feel real (Can emotion rewards make language models genuinely empathic?).

The evidence also shows how far fidelity goes. Persona-based simulations reproduced 84 of 111 published marketing-experiment main effects (76%), and success tracked how strong the original evidence was. Marginal effects came out unreliable, with both false positives and false negatives (Can AI personas reliably replicate human experiment results?). So the properties above seem to buy fidelity for large, robust human behaviors, and subtle effects remain out of reach. The corpus has little on which single property matters most, so the ranking is still open.