Does an AI persona get more accurate from being grounded in the right real-world people and documents, or from more data?

This explores whether a persona simulates its target accurately because it is grounded in the right domain, or because you feed it more data. The corpus never tests that head-to-head. No note varies data volume while holding alignment fixed. Still, the notes point the same way: what a persona is built from, and how it is structured, seems to matter more than how much material or model size goes in.

The strongest evidence is on the alignment side. LLM agents conditioned on anonymized real behavioral data predicted the direction of A/B test results with 75 to 90 percent accuracy across 40 experiments (Can behavior-based personas predict A/B test outcomes?). Another approach builds personas by clustering stakeholder perspectives out of domain documents, rather than assigning arbitrary roles (Can personas extracted from documents generalize across evaluation tasks?). Even well-grounded personas have limits. The behavioral personas were reliable for large effects and untrustworthy for near-zero ones. AI personas replicated 76 percent of published marketing-experiment effects, and success tracked how strong the original evidence was, with false positives and negatives on marginal effects (Can AI personas reliably replicate human experiment results?). Alignment gets you the direction of big signals, not fine detail.

On the volume side, the closest evidence is about model scale, which is not the same as data volume. Claude 3.5 Sonnet improved only 2.97 percent over GPT 3.5 on persona consistency despite a huge capability gap (Does model capability translate to better persona consistency?). The note's explanation is that standard training rewards per-turn quality, not staying in character across turns. Structure shows a similar pattern. Realistic synthetic dialogue came from layering subtopic, Big Five personality and contextual characteristics, which recovered 90.48 percent of in-domain performance (Can synthetic dialogues become realistic through layered diversity?). Shallow character descriptions fail to hold up, while layered, scripted personas produce more human-like consistency (Can layered persona architecture sustain coherent character behavior?).

There are two wrinkles. First, alignment is not the same as matching the task. A persona population built once was reused across surveys and chatbots through simple interface adapters (Can one persona population evaluate different application types?), and document-grounded personas transferred between summarization and dialogue evaluation. So what needs to match seems to be the people and perspectives, not the application. Second, accuracy scores can mislead. High persona-adherence scores often come from copying the character description while ignoring what was asked (Do persona consistency metrics actually measure dialogue quality?). And persona prompts steer surface output without removing underlying bias (Can persona prompts actually reduce bias in language models?). Part of a persona's apparent accuracy may be a measurement artifact.

In short, the corpus favors grounding and structure over raw quantity, but it has no controlled test of "more data versus better-matched data." That comparison is a gap in the collection.