How the models are built

A general-purpose model is told who to be: “You are a 34-year-old nurse in Leeds.”

NeedsA description of the audience. No data about real people is collected.

ExamplePolitical scientists gave a language model the demographic backstories of real US survey respondents and compared its answers with theirs. Argyle et al., 2023

NamesAI personas, silicon samples

LimitGets some groups’ answers right and others wrong, and treats each person as a stereotype of their group.Argyle et al., 2023; Santurkar et al., 2023; Wang et al., 2025

Critiques of this method

A general-purpose model reads one person’s interview or survey answers before it replies as them.

NeedsA long record for every person simulated, such as a two-hour interview or answers to hundreds of questions.

ExampleEach of 1,052 people gave a two-hour interview, and each person’s simulated twin read their own transcript before answering. Park et al. (Simile), 2024 (v1)

NamesGenerative agents of real people, and the digital twins in Toubia et al. and Peng et al.

LimitTwins’ answers are less varied than real people’s, and drift towards the average on questions they haven’t seen.Peng et al., 2025

Critiques of this method

A model is trained on what real people said and did, usually by fine-tuning a language model.

NeedsAnswers from many people: either a large set across many studies, or a set for each audience.

ExampleA separate model was fine-tuned for each of 420 clusters of similar people, using their own records. On small businesses that take card payments in person, it scored a median of 0.775 out of 1 on ordinal alignment, on the newest 20% of each person’s records, which were held back from training. Dahiya (Semilattice), 2026

LimitA model trained across many studies may not fit a particular audience. Training one model per audience needs data for each.Kolluri et al. (Simile), 2025; Dahiya (Semilattice), 2026

What the critiques tested