How the models are built
There are three ways to build a model: prompt it, ground it in one person’s data, or train it, and a provider can combine them. The introduction sorts the names for the category by which of these they describe. Providers also differ on whether each simulated person is a twin of one real person or a simulated individual, and on whether accuracy is scored at the individual level or the audience level.
A general-purpose model is told who to be: “You are a 34-year-old nurse in Leeds.”
NamesAI personas, silicon samples
LimitCan match some groups’ answer shares but misses others, and flattens individuals into stereotypes of their group.Argyle et al., 2023; Santurkar et al., 2023; Wang et al., 2025
A general-purpose model reads one person’s interview or survey answers before it replies as them.
NamesGenerative agents of real people, and the digital twins in Toubia et al. and Peng et al.
LimitAnswers vary less than real people’s and move towards the average on new questions.Peng et al., 2025
A model is trained on what real people said and did, usually by fine-tuning a language model.
LimitNeeds large sets of real responses. Some systems train one model across many studies; others train one per audience, which needs data for each.Kolluri et al. (Simile), 2025; Dahiya (Semilattice), 2026
Where each provider sits
Most providers now publish something about their methods. Each is placed by what its own pages and papers say.
Not placed: Electric Twin, which does not say whether its models are twins of real people.
Twins of real people
Agents grounded in each person’s own interview reproduced 85% of that person’s retest consistency. Park et al. (Simile), 2024 (v1)
- Artificial Societies Each persona grounded in a real individual
- Listen Labs One model per real interviewee
- Simile One agent per participant
- Aaru Simulated individuals built from population data
- Semilattice Each simulated person represents a cluster of similar real people, with only the traits they agree on, mined from their own answers
- Prompting tools Personas written into a prompt
- Electric Twin
Score each simulated person against the real one
Early studies built simulated people from demographics alone and mostly checked answer shares for groups. Argyle et al., 2023; Santurkar et al., 2023
- Listen Labs Held-out interview answers
- Semilattice Each person’s newest records held out
- Simile Against each person’s own retest
- Aaru Scored on answer shares
- Artificial Societies Scored on answer shares
- Electric Twin Scored on answer shares
- Prompting tools
Train a model on data about people
General-purpose models are tuned on human feedback to be helpful assistants (RLHF), which reduces the variety of their answers. Kirk et al., 2024
- Aaru Its own population model
- Artificial Societies Rothko, trained further on decisions
- Listen Labs Each twin trained on the client’s outcomes; method not given
- Semilattice A fine-tuned adapter per cluster of similar people
- Simile A fine-tune of Qwen
- Electric Twin General-purpose LLMs constrained by survey data
- Prompting tools General-purpose LLMs as they ship
Test beyond survey questions
Synthetic answers can match a survey while regression results estimated on them change size or sign. Bisbee et al., 2024
- Aaru Brand rankings against card-purchase data
- Artificial Societies Question-wording experiments and coherence
- Electric Twin Focus groups and image and video tests
- Simile Personality tests, economic games and experiments
- Semilattice Survey items only; journeys not yet validated
- Listen Labs
- Prompting tools
Report how confident each result is
- Aaru Predicts its own error for each question
- Listen Labs A confidence score, and abstains without signal
- Simile A trained model predicts each result’s error
- Artificial Societies
- Electric Twin
- Semilattice
- Prompting tools
Simulate a new audience without collecting new data
- Aaru Simulates a described audience from census and licensed data
- Simile Returns answers for a described population
- Prompting tools Any persona can be written into a prompt
- Listen Labs Needs an interview with each person
- Semilattice Needs a corpus for each audience
- Artificial Societies
- Electric Twin
From each provider’s own pages and papers, October 2026. “Not described” means they do not say either way.