Glossary

A B C D E F G H I J L M N O P Q R S T U V W

Agent-based modeling: Simulating a system by modeling many individual agents and their interactions, then watching what emerges at the level of the population.
AI persona: A representation of a type of person or a group, usually a written description given to a language model. ESOMAR calls these synthetic personas.
AI-moderated interview: An interview with a real person run by an AI interviewer instead of a human moderator. It follows a discussion guide, listens, and asks follow-up questions, by voice or text. AI interviewer →
Algorithmic fidelity: How closely a language model conditioned on real people's backgrounds reproduces how those people actually answer. The term comes from Argyle et al. (2023).
Augmented synthetic data: Extra rows generated to boost or fill in a small subgroup within a real dataset, also called boosting or imputation.
Behavioral traces: Timestamped records of actions people actually took, in context: purchases, clicks, visits, choices. The training data a foundation model of human behavior learns from.
Brand tracking: Measuring awareness, perception and preference for a brand repeatedly over time, so changes can be seen wave to wave.
Churn interview: An interview with a customer who stopped buying or using a product, to learn what happened and what would bring them back.
Concept testing: Showing people a product, feature or idea before it is built and measuring their reaction, to decide which concepts deserve investment.
Conjoint analysis: A survey method that shows people bundles of features and prices and asks them to choose, revealing how much each feature is worth to them.
Counterfactual: What would have happened under a different choice or condition. Simulation is useful partly because it can play out counterfactuals that cannot be observed.
CSAT: Customer satisfaction score: the share of customers who rate an interaction or product as satisfying, usually on a short scale.
Diary study: Research where participants record their behavior and thoughts over days or weeks, capturing what happens in context rather than from memory.
Digital twin: A model of one specific, real person that answers as they would, built from that person's own interviews, answers or behavior. Digital twins →
Discussion guide: The planned questions and topics for an interview. A good interviewer, human or AI, follows it while probing on what the person says.
Distributional alignment: Whether the spread of answers from synthetic respondents matches the spread from real people, not just the average. Measured with distances such as Wasserstein or total variation.
Ethnography: Observing people in their own settings over time to understand behavior as it actually happens.
Fine-tuning: Further training a pretrained model on a specific dataset so it performs better on a narrower task or population.
Focus group: A moderated discussion with a small group of people, used to explore reactions and language. Prone to group influence.
Foundation model: A model trained on broad data, generally with self-supervision at scale, that can be adapted to many downstream tasks. Coined by Stanford researchers in 2021.
Foundation model of human behavior: A model trained on records of what people actually did, under a general self-supervised objective, that learns a representation of each person and serves many tasks: prediction, simulation, ranking and profiling. What it is →
Freshness: How recent the data behind a digital twin is. People change, so twins need dated grounding data and a schedule for re-interviewing.
Generative agent: A language model agent with memory and planning that behaves like a person. Introduced by Park et al. (2023) as characters in a simulated town, later built from interviews with real people.
Grounding: What a synthetic respondent is built from. A twin grounded in a real person's interview is different from a persona grounded in a demographic description.
Hallucination: When a model states something fluent and confident that is not true. In research, an invented reason or an answer with no grounding.
Holdout validation: Testing synthetic respondents on real answers they were not built from, such as a past survey wave, to see how well they predict.
Human ceiling: The best accuracy any simulation can reach: how consistently real people give the same answer when asked twice. Accuracy is best reported as a share of it.
In-depth interview (IDI): A one-to-one interview, usually 30 to 90 minutes, that explores a person's experience and reasoning in detail.
Incidence rate: The share of a population that qualifies for a study. Low incidence segments are expensive and slow to recruit.
Instruction tuning: Training a model to follow instructions and be helpful. It makes assistants useful and tends to narrow the range of answers they give as respondents.
Jensen-Shannon divergence: A measure of how different two probability distributions are, used to compare synthetic and real answer distributions. Zero means identical.
Jobs to be done: A framework that studies the progress a customer is trying to make in a situation, rather than their demographics, to explain why they choose a product.
Likert scale: An ordered answer scale such as strongly disagree to strongly agree. Comparing synthetic and real answers on it needs distance measures that respect the order.
LLM-as-judge: Using a language model to grade other models' outputs. Common in benchmarks, and less reliable than scoring against real human answers.
Margin of error: The range within which a survey estimate is likely to fall from the true population value, driven mainly by sample size.
MaxDiff: Maximum difference scaling: people repeatedly pick the most and least important item from small sets, producing a ranked priority list.
Message testing: Comparing how different headlines, claims or copy land with an audience before a campaign runs.
Mixed methods: Research that combines qualitative depth and quantitative measurement in one study, often called qual plus quant.
Mode collapse: When a model gives nearly the same answer every time instead of the range of answers real people give. A common failure of assistant models used as respondents.
Net Promoter Score (NPS): The share of customers who would recommend a product minus the share who would not, from a 0 to 10 question.
Next-action prediction: Predicting what a person will do next from their history. The leading candidate objective for foundation models of human behavior.
Open-ended question: A question answered in the person's own words rather than chosen from options. The raw material of qualitative research and of digital twins.
Panel: A standing group of people who have agreed to take part in research, recruited and profiled in advance so studies can field quickly.
Persona prompting: Asking a general language model to answer as a described person. Cheap and fast, but less accurate than twins grounded in real interviews.
Pricing research: Studies that estimate what people will pay and how demand changes with price, using methods such as Van Westendorp, Gabor-Granger or conjoint.
Qualitative research: Research that explores why people think and act as they do, usually through interviews or open answers, in their own words.
Quantitative research: Research that measures how many people think or do something, usually through structured surveys large enough to count.
Quota sampling: Recruiting until each group, such as age or region, reaches a target count so the sample mirrors the population.
Representativeness: How well a sample, real or synthetic, reflects the population it stands in for, overall and within each group.
Retrieval-augmented generation (RAG): Having a model look up relevant records, such as a person's interview, before answering, so the answer is grounded in that data.
RLHF: Reinforcement learning from human feedback: training a model on human preferences between answers. It improves helpfulness and can make models more agreeable.
Sample size: How many people answered. Larger samples narrow the margin of error; synthetic samples can be large without being more accurate.
Screener: The short set of questions that decides whether a person qualifies for a study.
Segmentation: Dividing a market into groups that differ in needs or behavior, so each can be understood and targeted on its own.
Self-supervised learning: Training a model on the data itself, such as predicting the next word or next action, without hand-made labels. How foundation models are trained.
Sentiment analysis: Classifying text as positive, negative or neutral. Useful for scale, too coarse on its own to explain why people feel as they do.
Silicon sample: A language model conditioned on the backstories of real survey respondents and used as a stand-in sample. From Argyle et al. (2023).
Simulated population: A set of synthetic respondents meant to represent a market, audience or population, queried in place of fieldwork.
Statistical significance: Whether a difference is large enough relative to sample noise to be unlikely by chance. Synthetic data can make tiny differences look significant.
Sycophancy: A model's tendency to agree with or please the person asking. In research it shows up as answers that are too positive.
Synthetic data: Information created to replicate the characteristics of real-world data. In market research it ranges from boosted samples to fully synthetic respondents.
Synthetic panel: A standing population of synthetic respondents that can be queried repeatedly, the simulated counterpart of a research panel.
Synthetic respondent: Any AI that answers a survey or interview in place of a person. The broadest of the terms; it covers both personas and digital twins. The terms compared →
Synthetic user: A UX research term for an AI-generated profile that stands in for a user group, producing findings without research on real users.
Test-retest reliability: How consistently the same people answer the same question at two points in time. The benchmark that synthetic respondents should be measured against. How to evaluate →
Thematic coding: Grouping open answers or interview transcripts into recurring themes so they can be counted and compared.
Total variation distance: The largest difference in probability two distributions assign to any outcome. SimBench uses it to score how well models simulate people.
TURF analysis: Total unduplicated reach and frequency: finding the combination of products or options that reaches the most people.
Usability testing: Watching people try to complete tasks with a product to find where it confuses or fails them.
User model: A system's standing representation of one person, used to personalize or assist. One of the fields converging on foundation models of human behavior.
Van Westendorp price sensitivity meter: Four questions about when a price feels too cheap, a bargain, expensive and too expensive. The crossing points bound an acceptable price range. How to run it →
Voice of the customer: A program for collecting and acting on what customers say across surveys, interviews, support and reviews.
Wasserstein distance: A measure of how far one distribution must move to match another, suited to ordered answer scales. Used by OpinionQA to compare models with survey populations.
Weighting: Adjusting survey results so groups count in proportion to their size in the population, correcting for over- or under-sampling.
WEIRD bias: The tendency of models to answer like people from Western, educated, industrialized, rich and democratic societies, and less like people elsewhere.
Willingness to pay: The highest price a person would pay for a product. Estimated with pricing research, and one of the better-studied uses of synthetic respondents.