Researchers interviewed 51 residents of Vancouver's Downtown Eastside, a low-income neighbourhood, about their lives and their views on artificial intelligence. Four large language models were then asked to answer the same interview while role-playing each of those 51 people, given only a short demographic profile of each person. This paper compares the wording of the real answers with the wording of the model-generated answers. Real speech and model speech differ on every measure of speech texture that we applied. Real residents produce hesitation and filler at 49.4 markers per 1,000 words; the four models produce between 0.25 and 6.9. Real residents hedge more, repeat words more, and use first-person language more than the models do. The two 2026 frontier models did not reduce this gap relative to the 2025 models, and on the central measure one frontier model widened it. The result is descriptive: when a model is asked to speak as a person from this community, it generates fluent composed prose rather than a record of how that person talks.
Geraskin et al. (Sun,) studied this question.