Quick answer: Both trace their lineage to computational social science, but they are built to do different things. Simile builds AI twins bottom-up from data about individuals, such as interviews, and rehearses how a configured population behaves over time. Artificial Societies builds networks of personas from real-world observations and models how opinion forms and spreads across connected groups — reaching hard-to-recruit audiences, answering within 24 hours, and validated in peer-reviewed research plus product-level replication studies, with every number open to the individual voices behind it. Simile suits modelling traditionally reachable individuals and longitudinal scenarios; Artificial Societies suits group-opinion dynamics on the high-value, hard-to-reach audiences behind consequential decisions.
Simile (simile.ai) is a US enterprise behavioural-simulation platform co-founded by Stanford's Joon Sung Park, author of the seminal "Generative Agents" paper, with Michael Bernstein, Percy Liang and Lainie Yallen. It creates AI agents — twins built bottom-up from data about real individuals, such as interviews — each with memory, goals, and personality, and simulates how a configured population of them behaves over time: rehearsing meetings, policy changes, product rollouts, or legal scenarios and observing emergent behaviour. The approach builds on Stanford research finding that generative agents from interviews with 1,052 people replicated those individuals' General Social Survey answers at roughly 85% of the accuracy the participants themselves achieved on a two-week retest (Future Economist). Pricing is not public; the platform is enterprise sales-led.
Artificial Societies (societies.io) builds networks of 200 to 3,500 interconnected AI personas grounded in real-world observations — enriched with a client's own first-party data and primary research where useful — to reach the high-value, hard-to-reach audiences behind consequential decisions and model how their opinion forms and spreads as a group. Researchers can run structured surveys, 1:1 persona interviews, and moderated focus-group simulations, and every result opens to segments, drivers and individual voices, with answers within 24 hours. Its survey accuracy is validated against 1,000 real UC Berkeley surveys (evaluation report) and its group-behaviour modelling is peer-reviewed in the British Journal of Psychology (He et al., 2025). Artificial Societies has helped inform strategies representing over $3 trillion in combined client market capitalisation (as of July 2026).
Simile builds twins bottom-up from data about individuals, such as interviews, and observes how a configured population of them behaves. Artificial Societies is grounded in real-world observations rather than individual interviews, and its distinctive focus is how opinion forms and spreads across connected groups — the interactions and social dynamics that decide what spreads, what stays contained, what organises and what holds, not just what each individual twin does in isolation.
Simile is a general-purpose simulation environment: a scenario is configured and emergent agent behaviour observed over time. Artificial Societies is a research product: audience construction, survey design, statistical outputs, interviews, focus-group simulations, and automated insight generation are the workflow — built for insights, comms, and strategy teams answering concrete research questions within the decision window, not configuring simulations.
Simile's widely cited ~85% figure comes from the founders' peer-reviewed Stanford research: generative agents replicated participants' General Social Survey answers at roughly 85% of the accuracy the participants themselves achieved retaking the survey two weeks later — a normalised measure of the underlying research technique (Future Economist). It is genuinely peer-reviewed; there is no published validation of Simile's commercial product on client work known as of July 2026. Artificial Societies' validation covers a January 2026 product-level accuracy evaluation against 1,000 real surveys from UC Berkeley research — 93% response consistency and 86% distribution accuracy, within five points of the 91% human-replication ceiling (evaluation report) — plus peer-reviewed validation of its group-behaviour modelling in the British Journal of Psychology (He et al., 2025).
Simile's publicised engagements involve configuring detailed agent populations and scenarios — powerful for longitudinal behavioural questions, but oriented toward configured simulation programmes rather than quick research questions. Artificial Societies answers within 24 hours, so research fits deal timelines, crisis windows and earnings cycles.
| Artificial Societies | Simile | |
|---|---|---|
| Built for | Audience research: surveys, interviews, focus groups, insight | Behavioural scenario simulation over time |
| Persona basis | Real-world observations, plus optional client first-party data | Twins built bottom-up from data about individuals (e.g. interviews) |
| What it models | How opinion forms and spreads in connected groups | Emergent behaviour of configured individual agents |
| Accuracy evidence | 93% response consistency / 86% distribution accuracy vs 1,000 real UC Berkeley surveys (Jan 2026); group behaviour peer-reviewed in British Journal of Psychology | ~85% relative GSS replication from Stanford peer-reviewed research on the underlying technique |
| Time to answer | Within 24 hours | Scenario configuration; enterprise programmes |
| Interrogation | 1:1 interviews with any persona; opens to individual voices | Observe emergent simulation behaviour |
| Jurisdiction | US & UK; GDPR-compliant, EU data hosting, SOC 2, DPAs | US |
| Best for | Group-opinion dynamics on hard-to-reach audiences; defensible insight | Rehearsing complex organisational or strategic scenarios |
Research capabilities across the two platforms. “—” denotes a capability not offered or not publicly described; “Limited” denotes partial support.
| Artificial Societies | Simile | |
|---|---|---|
| Individual 1:1 interviews | Yes | Yes |
| Audience segmentations | Yes | Yes |
| Question-pair / driver relationships | Yes | — |
| Cross-tabulations | Yes | — |
| Focus group simulations | Yes | — |
| Open-ended qualitative reasoning | Yes | Yes |
| Social-dynamics / opinion-spread modelling | Yes (group opinion) | Limited (bottom-up agents) |
| UX/UI usability testing | Not offered | Yes |
Artificial Societies can model any audience for which diverse observations exist, whether public or proprietary, and is strongest on the hard-to-reach professional and elite audiences below. Simile models the populations a client configures from supplied data.
| Audience | Artificial Societies | Simile |
|---|---|---|
| General public | Yes | Yes |
| Consumer segments | Yes | Yes |
| Topical communities | Yes | No |
| Blue-collar workers | Yes | No |
| Professionals | Yes | No |
| Investors | Yes | No |
| Shareholders | Yes | No |
| Regulators | Yes | No |
| Lawmakers | Yes | No |
| Opinion leaders | Yes | No |
Choose Simile to model traditionally reachable individuals rather than group opinions, or to understand consumers rather than high-value, hard-to-reach audiences — and for longitudinal behavioural scenarios where a deep, configured simulation programme fits the timeline and the strongest academic pedigree in generative agents matters.
Choose Artificial Societies for group-opinion dynamics — how a message spreads, what organises, what holds — and for consequential decisions on the high-value, hard-to-reach audiences traditional research struggles to recruit. Answers arrive within 24 hours, backed by peer-reviewed methodology and product-level replication, and every number opens to segments, drivers and individual voices, so recommendations survive boards, partners and regulators.
Simile builds twins bottom-up from data about individuals and observes their emergent behaviour over time. Artificial Societies is purpose-built for group opinion: grounded in real-world observations, it models how opinion forms and spreads across connected groups — what spreads, what stays contained, what organises and what holds — interpretable down to persona level.
It comes from the founders' peer-reviewed Stanford research, which found generative agents replicated individuals' General Social Survey answers 85% as accurately as those individuals replicated their own answers two weeks later. It validates the underlying technique; there is no published product-level validation on client work known as of July 2026. Artificial Societies publishes both peer-reviewed methodology and product-level replication results.
Artificial Societies — audiences are built from real-world observations and answers come within 24 hours. Simile's publicised model is oriented toward configured, longer-running scenario simulations.
Simile's agents are described as twins grounded in data about real people, such as interviews. Artificial Societies requires no client data — personas derive from real-world observations, processed under GDPR with a legitimate interest assessment and an individual opt-out mechanism — though bespoke societies can be enriched with client first-party data.
Yes. Artificial Societies operates across the US and the UK and is GDPR-compliant with EU data hosting, with DPAs, a SOC 2 attestation, customer data never used to train models, and deletion within 90 days of contract end.
Last updated: July 17, 2026. Simile is a trade mark of its respective owner. Artificial Societies is not affiliated with or endorsed by Simile. Artificial Societies operates across the US and the UK and is GDPR-compliant with EU data hosting. Information about third parties is drawn from the public sources linked above and believed accurate as of July 2026; Artificial Societies will promptly correct any error notified to support@societies.io.