Simile AI vs Artificial Societies: methods compared

Artificial Societies’ 86% distribution accuracy is 95% of the 91% human self-replication ceiling, measured against 1,000 real-world surveys in its Survey Evaluation Report, January 2026. Distribution accuracy measures overlap between simulated and human opinion distributions. For a buyer testing a message, it tells you how closely the simulated shares choosing each answer match the human shares. It leaves a separate question about which people hold which combinations of opinions.

We publish measures of the responses beneath those totals. The January 2026 report records under 2% self-contradiction, compared with about 9% for human panels. Our evaluation page calls this the hallucination rate: it measures personas contradicting themselves, rather than factual errors about the world. Internal coherence is 89%, within the human-panel range of 60–95%, according to the same report. That measure, Cronbach’s alpha, tests whether answers to questions about the same underlying attitude agree; it does not establish whether findings generalise to another population.

Artificial Societies’ peer-reviewed paper on group behaviour (He, Wallis, Gvirtz and Rathje, British Journal of Psychology (opens in a new tab), December 2024) supports the network modelling, not the survey accuracy figures; the science page explains what it found.

Artificial Societies helps organisations understand people before making consequential decisions. We start with observations of real people, reconstruct how they reason and connect personas through relationships and patterns of influence. Our method page describes analysis at market, segment and individual persona level, so you can examine both the audience response and the reasoning beneath it.

The third stage connects those people through relationships and influence. A society contains 12 to 3,500 personas, according to Artificial Societies’ method and evaluation page. The audience defines the scope: a specialist group around a stock can call for a different society from a consumer population. Before a persona enters a society, we validate it against audience benchmarks and prior behaviour.

Our input material includes published research, online commentary and reviews, alongside anonymised social data and public records of voting and investment. That grounding sets a condition on the work: we model audiences for which diverse observations exist, whether public or client-provided. If those observations do not exist, there is nothing to ground the personas in. A benchmark measured on an observed audience cannot establish accuracy for an audience nobody has observed.

Simile AI says it simulates how customers, employees and populations respond to change. Its homepage (opens in a new tab) describes daily studies combined with proprietary human-behaviour datasets to train its models. It also says agents encounter news in real time and that the company trains continuously on new behavioural, pricing and policy data.

The confidence model is a specific part of that offer. Simile says it validates against humans weekly through over 7,000 evaluations covering subpopulations and enterprise use cases, then uses those validations to predict an accuracy level for each result. That is Simile’s account of its process, published on its homepage (opens in a new tab), rather than a test we have conducted.

Simile names earnings-call rehearsal, litigation outcomes and policy changes as uses for its simulations in The Simulation Company (opens in a new tab). Earnings-call rehearsal also appears in our hyperscaler case study. For procurement, the useful distinction is what you need to inspect: how a reaction moves through a network, or the predicted reliability attached to a particular result. Simile’s August 2026 confidence-model post (opens in a new tab) gives you a published account of how it assigns those labels.

Artificial Societies and Simile AI both describe simulations grounded in real people. The difference to examine is the role of relationships between personas and the role of a predicted accuracy label. Our table applies the same criteria to each, with a final column naming the reason to consider the other approach.

Sources: Artificial Societies, method and evaluation and group-behaviour paper (opens in a new tab); Simile, homepage (opens in a new tab) and confidence-model post, August 2026 (opens in a new tab); Park and colleagues, revised paper, June 2026 (opens in a new tab).

The research score associated with Simile needs its version attached. The original Generative Agent Simulations of 1,000 People, posted on 15 November 2024 (opens in a new tab), reported that agents replicated General Social Survey responses 85% as accurately as participants repeated their own answers two weeks later. That percentage expresses replication relative to the same people’s consistency on a later test.

The 28 June 2026 revision (opens in a new tab) is titled LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals. On held-out General Social Survey questions, it reports 83% for interview-only agents, 82% for survey-only agents, 86% for combined agents and 74% for a demographics-only baseline. Those scores use participants’ own two-week consistency as the reference.

The revised paper therefore offers a comparison between ways of grounding an individual in self-reports. It is not a commercial Simile product accuracy score. Nor does its combined-agent figure measure our distribution overlap. Reading the shared percentage as a tie would erase the distinction between reproducing a person’s answers and reproducing the proportions choosing each answer in a population.

Simile AI is the option we would examine first when the object of study is an individual represented through their own interviews and survey answers. Its researchers’ June 2026 paper (opens in a new tab) directly tests that form of representation against the participants’ later answers. If fidelity to those particular people is the research question, that is relevant evidence on the question you are asking.

We would also send you to Simile when a per-result accuracy label is a requirement. Its August 2026 confidence-model post (opens in a new tab), by Andrew Wesel, Sarah Chen and Percy Liang, reports that results tagged “High” were 95% likely to meet its decision-quality threshold, compared with 38% for “Low”. The threshold is 0.16 total variation distance, a measure of the difference between answer distributions. Here, its purpose is to decide which differences are small enough for Simile to call a result decision-grade.

That calibration gives a buyer something concrete to discuss with Simile. The authors also expect the threshold to change as they add ratings, particularly from external raters, and use more diverse evaluation sets. Ask for the calibration relevant to your intended audience and question type. A label becomes useful to your decision when you understand the error it predicts and the threshold your team is prepared to accept.

We recommend Artificial Societies when the decision involves reactions across connected stakeholder groups, where message spread and social influence are part of the question. Our earnings-call work shows what that means in a study.

For a leading hyperscaler, Artificial Societies built four societies around the stock, including 14 sell-side analyst personas representing the analysts who had asked questions on the company’s last eight calls, according to the earnings-call case study. The work examined likely analyst questions and investor reactions to guidance and scripts, followed by how the story would travel through the corporate ecosystem and financial press. The client could test unreleased figures without showing them to real investors.

The company expected its margin reset to dominate the call. Simulated analysts instead concentrated on demand durability and financing commitments; the case study reports that the live questions followed that pattern. The client received a pre-call briefing and access to rerun scenarios as the numbers changed. This is a concrete reason to consider our approach for investor communications, with the observed agreement kept separate from the survey benchmark.

For a broader narrative decision, Artificial Societies’ Teneo case study describes a late-2025 engagement testing six technology narratives across three societies. The late-2025 Teneo study records 189,756 unique responses and identifies the strongest overall narrative, with messaging adjusted for each audience segment. Teneo received both a report and access to explore approval scores, emotional sentiment and verbatim reactions. Those deliverables let the team inspect why a narrative worked differently with policymakers, industry peers and consumers.

For a Simile vs Aaru comparison, start with the population and question format behind each result. EY’s 7 October 2025 account (opens in a new tab) reports a 90% median correlation for Aaru across 53 single-select questions, against research involving 3,600 affluent investors in more than 30 markets. EY produced and published the account within a commercial partnership. It gives a wealth-research buyer a specific study to examine, rather than a general guarantee for other audiences.

Simile’s June 2026 primary paper (opens in a new tab) tests individual response replication against human retest consistency. Aaru’s partner study measures correlation, or how the results vary together. Our January 2026 Survey Evaluation Report measures distribution overlap. We would not choose a supplier by ordering those figures: they answer different questions on different test sets.

Other Simile alternatives fit a different research task. Synthetic Users’ science post (opens in a new tab) describes a late-February-2024 comparison of eight human and eight synthetic interviews with UK school teachers about classroom technology. It reports qualitative similarity, scored through thematic overlap, depth and coverage, alongside qualitative alignment. Examine that evidence for an interview-led study; examine our network method when relationships between stakeholders are part of what you need to model.

The right Simile AI alternative depends on the research task, so start with the decision and the evidence you need to inspect. Our Simile alternatives page compares the main options by what each simulation represents and how it tells you when to trust a result. Our guide to AI market research tools sets Simile AI beside 10 other tools, applies the same criteria to each and lists the vendors alphabetically. Each page below sets one vendor’s published materials beside those of Artificial Societies:

• •

  Aaru AI vs Artificial Societies

• •

  Electric Twin AI vs Artificial Societies

• •

  Evidenza AI vs Artificial Societies

• •

  Artificial Societies vs Synthetic Users

Two further pages cover several platforms at once, Simile among them: Artificial Societies alternatives and Synthetic Users alternatives. Whichever vendors you shortlist, how to evaluate a synthetic research vendor sets out eight validity tests to put to each of them.

Simile AI does not publish pricing on its site (opens in a new tab). We price on usage and scope each engagement with the team; Book a call to talk through the audience and decision you have in mind. A price comparison needs to specify the work you expect to receive, including access to results and the ability to rerun a study.

Artificial Societies’ earnings-call case study records a 48-hour engagement with a briefing and platform access. The Teneo study records an engagement completed in under three weeks, including building the societies. Use these as examples of delivered scope when discussing a proposal; neither duration establishes a delivery promise for your project or a speed comparison with Simile.

We publish this comparison. We used the sources listed at the end of this page: company descriptions, our evaluation report and case studies, the primary research papers, partner studies and third-party comparisons. Source dates are listed below; the comparison uses material recorded on 17 September 2026. We have not run a head-to-head product trial for this page, and the table compares published methods rather than assigning a common accuracy ranking.

Information is believed accurate as of those source dates. Third-party trade marks belong to their owners; naming them implies no affiliation or endorsement. Send corrections to support@societies.ai. We will review corrections against the cited material and update the comparison where needed.