How accurate is Cast?
Five peer-reviewed studies, more than 120,000 real participants between them, say the same thing: a language model given a well-defined person answers the way that person's group answers. Cast is built on that finding, and on the engineering that keeps it honest.
Research from
What the research shows
Each card states the result the paper reports and links to the source. The long reads are in Cast Diaries.
Give a model a person's background and it answers the way that person's group answers, in detail and not at random. The authors call this algorithmic fidelity. It is the property Cast's whole method rests on.
Argyle et al., 2023 (opens in a new tab)
Read more…Asked to predict 70 real, preregistered experiments before the results were known, the model's predictions tracked the actual effects about as well as pooled human forecasters did. The thing Cast asks AI users to do, anticipate how people react, is what the model is measured to be good at.
Ashokkumar et al., 2026 (opens in a new tab)
Read more…Agents built from a two-hour interview with a real person reproduced that person's survey answers at 83% of the person's own two-week consistency, 86% when combined with survey data, against 74% for demographics alone. The richer the person, the truer the answers. Cast interviews; it does not poll.
Park et al., 2024 (opens in a new tab)
Read more…Classic behavioral economics experiments, re-run on a language model, gave results qualitatively like the human originals: fairness, status quo bias, framing. The model reacts to price and trade-offs the way people do, which is why Cast can ask about willingness to pay.
Horton, Filippas and Manning, 2023 (opens in a new tab)
Read more…The authors replicated the Ultimatum Game, garden-path sentences and the Milgram experiment with simulated participants, and found the same patterns as the human studies. They also found a hyper-accuracy distortion in some models, which is one of the pitfalls Cast designs around below.
Aher, Arriaga and Kalai, 2023 (opens in a new tab)
Read more…What Cast does with it
Every mechanism below runs in code on every study. None of it is a prompt you have to remember to write.
Each AI user is interviewed on its own, in a conversation that follows up on what was just said. Park et al. found interview-built agents the most faithful; a survey tick-box throws that fidelity away. No answer leaks from one interview into another.
The research's first warning is that synthetic answers come out too uniform. Fuzziness spreads each cast around the Ideal Customer Profile (ICP): on Medium, about a third of the cast sits off the anchor on every dimension, so no two AI users are the same person and a theme has to earn its count.
Every cast books at least one in four AI users who start unconvinced: a quarter of the cast, rounded up, is held for the people who have been burned before, see no problem, or would rather not change. If everyone loves your idea, the cast was built wrong, and Cast's rules do not allow it.
Before anyone is cast, an ICP Researcher runs web searches on your market and writes a fact sheet with sources. That sheet shapes who is in the cast: the spend bands, the tools in use, the real pains. AI users themselves never see the web. They know only what their own record says, like a real customer.
An Auditor checks every AI user before the interviews: facts that contradict each other, knowledge the person would not have, a stance that does not match the record. Then a Critic, a separate model family from the AI user, reads every finished interview against that record and flags answers that are invented, out of character or too eager to please. Nobody grades their own homework.
Every quote in a report is checked word for word against the transcript: one unbroken stretch of that AI user's own answer, or it is dropped. Every count, every "6 of 8", is computed by code from the interview records. The model writes the prose; it never gets to write a number.
When the Casting Agent adds a group you did not name, because the brief warrants it, the ICP card says "Added by Cast" with the reason. You can remove it in one click. Nothing in a cast is a surprise.
A study goes from brief to findings in about a quarter of an hour. In the load test, five studies run at once all reached findings in 12 to 16 minutes, every interview finished.
Markdown, PDF, PowerPoint and an Infographic, a podcast in one voice or as a two-voice debate, and a chat with the findings that answers from the transcripts. One study, every format, in English or Arabic.
Every model that touches your brief runs on a zero-data-retention host and nothing you paste is used for training. The data rule is enforced in code, not in a policy.
Read Trust and DataWhere the research warns
The same literature that measures how well models simulate people also names where they slip. Those papers shaped the design of Cast as much as the five above.
The research says
Synthetic answers can be too uniform.
Cast does
Fuzziness varies every AI user, and the skeptic floor guarantees people who disagree. The report shows the spread.
The research says
Answers can shift with the wording of the prompt.
Cast does
Every AI user is interviewed separately with the same guide, and the Critic flags an answer that reads like the model rather than the person.
The research says
Models can be too accurate, or misjudge a price.
Cast does
An AI user knows only what its record says, price goals run at Low Fuzziness so the price reads stay comparable, and the report tells you which findings to confirm with real people before you bet the company.
Sources: Bisbee et al., 2024 (opens in a new tab) · Brand, Israeli and Ngwe (opens in a new tab) · The long read in Cast Diaries
Describe your customers. Meet them in minutes, with the receipts.
