Instead of describing a character in words, could you just show an AI how they talk — and lose nothing?

This explores whether example dialogue can replace a written profile of a persona without dropping the demographics or the way of speaking. The corpus never runs that head-to-head test, so it can't give a clean yes or no. What it does have is evidence about why written descriptions are weak, what persona information is fragile whatever the format, and how you would check a swap.

The case against written descriptions is real. Short character blurbs don't reliably hold up behavior over a conversation, which is why Can layered persona architecture sustain coherent character behavior? adds structure around them. Worse, a high persona-consistency score can be gamed. Models often score well by copying phrases from the description while ignoring what the other person just asked, as shown in Do persona consistency metrics actually measure dialogue quality?. A description invites parroting, and dialogue samples might avoid that by showing the persona in motion. But nothing here tests it.

The risk is that demographics are the fragile part in any format. Persona prompts mostly change surface output. Between-group sentiment gaps stay put, which suggests the prompt steers style rather than deeper knowledge (Can persona prompts actually reduce bias in language models?). Run-to-run variance for a single persona can match or exceed the variance between different personas (Why do LLM persona prompts produce inconsistent outputs across runs?). My inference is that dialogue samples would carry register, word choice and rhythm well, since those are on the page. Age, occupation or cultural background would only come through if they happen to show up in those lines, and a model would have to infer them rather than read them.

A lateral reason samples underdetermine a person is that a language model doesn't commit to one character. It holds a spread of characters consistent with the context and samples from it, so regenerating gives different answers that each fit (Do large language models actually commit to a single character?). A description acts like a constraint, stating what must stay fixed. A few dialogue samples only narrow the spread, and many different people could have said those lines. Descriptions may be clumsy, but they are the more explicit way to pin down the facts you care about.

The corpus does show how to test a swap, and it points to a hybrid. Realistic synthetic conversations seem to need several layers at once: topic, personality and situation (Can synthetic dialogues become realistic through layered diversity?). Explicit profile and intent variables also let researchers check realism with human discriminators and classifier ensembles (Can controlled latent variables make LLM user simulators realistic?). The drift-reduction work already scores a simulator against its prompt, against its own earlier lines, and on factual Q&A (Can training user simulators reduce persona drift in dialogue?). Run those same three checks on a dialogue-sample persona versus a written one and you would find out what was lost. Another middle route is to derive personas from real source documents instead of writing them, as in Can personas extracted from documents generalize across evaluation tasks?. The safest reading of the evidence is to use dialogue samples for style and keep a short explicit statement of the demographics.