# Services That Replace Acceptance Testing with AI Synthetic Respondents: How Agents Are Built, Counted, Shown Stimuli, and Sold

- Author: Shuichiro Ogawa (Design Researcher / Design Consultant), https://jamjet-iqbal.com/
- Published: 2026-11-14; updated: 2026-11-14
- Canonical (HTML): https://jamjet-iqbal.com/en/notes/synthetic-respondent-services/
- Japanese version: https://jamjet-iqbal.com/notes/synthetic-respondent-services/ (Markdown: https://jamjet-iqbal.com/notes/synthetic-respondent-services.md)
- Language: en
- Keywords: synthetic-respondents, synthetic-personas, digital-twins, silicon-sampling, acceptance-testing, concept-testing, market-research, pricing, business-model, research-ethics, esomar, aapor, ai-agents

> This note examines 38 services that build many AI agents with different traits and run them through the same acceptance testing that would otherwise be done with people (showing concepts, ads, or prices and checking whether they are accepted). The evidence comes from official pricing pages and documentation, developer statements, industry codes, and third-party comparisons, as of November 14, 2026. Agent construction does not split into choosing from a vendor's library versus building from the buyer's own data; most products offer both. The products differ instead in the data that grounds the answers: the vendor's panel and purchase data, public statistics, the buyer's research data, or long interviews with real individuals. Stated limits mix three different quantities: population size (100,000 people, or a scale of about 100 million), responses per study (from 5 to 10,000), and plan quotas. None of them counts independent people. Increasing the number of agents from 500 to 5,000 reduced the error only from 6.8 to 6.7 points (Verasight 2026). Text stimuli work in every product, while support for images and video differs by product, and Qualtrics synthetic panels do not process images, videos, files, or links. Pricing falls into monthly self-serve plans (from USD 20 per month), annual contracts with credits or tokens, flat licenses with unlimited queries, and quoted contracts, and most plans are consumed per response or per interview. All agreement figures with human answers are measured by the vendors themselves. In third-party comparisons, topline estimates still missed by 2 to 12 points, and subgroup results, price logic, and change over time missed further. The codes and reports of ESOMAR, MRS, the Insights Association, and AAPOR require that buyers be told when synthetic data are used and which outputs are synthetic and which come from people.

## Overview

Acceptance testing shows people a product concept or an ad and checks whether they accept it.
Between 2023 and 2026, a series of services appeared that replace this testing with many AI agents built to think differently and answer in place of people.
The industry calls these agents synthetic respondents, synthetic personas, or digital twins.

This note answers four questions about 38 such services.
What are the agents built from?
How many agents can be set up?
What is shown to the agents to test acceptance?
Where and how do the services charge?

The answers are as follows.
Agent construction does not divide into a type that chooses from what the vendor provides and a type that builds from the data of the user (hereafter, the buyer).
Most products offer both, and the products differ in the data that grounds the answers.
Stated limits mix population size, responses per study, and plan quotas, and none of these counts independent people.
Text stimuli can be used in every product, whereas support for images and video differs by product.
Most plans are consumed per response or per interview, which follows the "price per respondent" logic of human research.

## Scope and Method

The scope covers commercial services that generate answers from several AI personas in response to stimuli such as concepts, ads, prices, and messages.
Tools that write persona documents without generating answers, tools that only summarize real human answers, and purely academic prototypes were excluded.
The resulting set contains 38 services: 17 specialized startups outside Japan, 11 large research companies and research platforms, and 10 Japanese companies (some Kantar and Ipsos products have only been announced).

Collection on November 14, 2026 ran two modes at the same time.
The AI-frontier mode treated official vendor materials (pricing pages, documentation, terms of service, and announcements) as the primary evidence.
The industry mode collected codes from industry bodies and public agencies, comparisons by research firms and consultancies, developer statements, and the prices and labor market of human research.
The ledgers are in `source/review/synthetic-respondent-services/`.

Accuracy figures that vendors publish for their own products (such as "90% agreement") were treated as marketing claims unless the method and sample were disclosed.
For prices, "quote required," "not stated in the materials checked," and "not retrievable" were kept distinct, and undisclosed prices were not estimated.

## What Are the Agents Built From?

The first question asks whether agents are chosen from what a vendor has prepared or built from data the buyer provides.
When the products are compared, these two options do not separate the products.
Ask Rally offers a general-population panel of 200 personas in its free tier, and from its USD 20 plan it adds a function that "clones" personas from the buyer's own data.
Synthetic Users lists on its pricing page a function that lets buyers upload their own data and retrieve it to enrich answers (RAG).
DNP's Persona Insight offers two kinds of personas: 100 common personas that reproduce the statistical distribution of Japan, and custom personas built from the buyer's data or target profile, the latter as a paid option.

What separates the products is the data that grounds the answers.
The products examined fall into four groups.

**The vendor's panel and purchase data**: Toluna builds one million synthetic personas from anonymized data of its own panel of more than 81 million people and states that they are "not generic AI-generated profiles or segment averages."
NIQ's BASES AI Screener returns synthetic feedback rooted in NIQ's purchasing behavior data and refreshes that data every few weeks.
Dentsu announced that it fine-tuned a language model on its own survey of about 150,000 people to reproduce a population at a scale of about 100 million.
INTAGE's PerSearch builds AI personas from the purchase records of a national consumer panel of 70,000 people (SCI).

**Public statistics**: TOPPAN uses the synthetic population method on the census, the family income and expenditure survey, population estimates, and other sources to build 100,000 personas that reproduce the population structure of Japan, and assigns values and preferences to each persona.
FishDog builds about 340,000 synthetic Americans from US census data and states that, in aggregate, they match the census within about 1%.

**The buyer's research data**: Electric Twin builds its synthetic audiences from the buyer's customer data.
quantilope's Category Twins use the buyer's brand tracking data, and Ipsos PersonaBot uses the buyer's segmentation study.

**Real individuals**: Brox recruits and verifies real people, runs roughly five hours of AI-moderated video interviews with each person, and builds a one-to-one behavioral twin.
Kantar defines a twin as a replica of a real research respondent and describes its consumer twins as "upcoming."
YouGov Parallax lets AI twins answer from YouGov's attitudinal and profile data, after which live panelists validate the results in 30 minutes.

The grounding data show up in how far the answers miss.
Pew Research Center (2026) built a twin for each real respondent in its panel and compared the twins' answers with human answers on the questions of three surveys.
Adding an extended profile of 71 variables to the basic attributes reduced the average absolute survey error to 13.7 points, and adding an expert reflection reduced it to 13.1 points.
More data narrowed the error, but did not remove it.

The same distinction appears in an industry classification.
In November 2026, Synthetic Data Standard, a body that presents itself as a certifier, published a field guide that sorts synthetic research products into six types by two questions: who the product speaks for, and what data drives its answers (as reported by DRNO 2026).

## How Many Agents Can Be Set Up?

The numbers that products present mix three different quantities.
The first is the size of the population, as in TOPPAN's 100,000 personas, FishDog's roughly 340,000, Toluna's one million, and Dentsu's scale of about 100 million.
The second is the number of responses in one study, which plans often cap.
The third is the quota included in a plan for a given period.

| Product | Stated number | Kind of number |
|---|---|---|
| Ask Rally | Maximum sample size of 5 (free), 15, 100, and 500 (plans at USD 20, 100, and 500 per month) | One study |
| Qualtrics synthetic panels | 50 to 10,000 responses per panel | One study |
| Yabble | Up to 8 personas per project, 50 responses per question | One study |
| C3reve Persona AI | 3 personas in the free trial, 100 in the Standard plan | One study |
| Delve AI | 100 synthetic users in the USD 317 monthly plan | Plan quota |
| Artificial Societies | 12 to 3,500 personas per society | One study |
| Fujifilm Business Innovation Persona AI | Several hundred to several thousand responses per run, up to 15 runs per month | One study and plan quota |
| TOPPAN, Toluna, Dentsu | 100,000 people, one million personas, a scale of about 100 million | Population |

None of these numbers counts independent people.
Answers drawn from the same model share the same bias.
The AAPOR task force (2026) states that reporting the number of AI instances is "neither required nor recommended," because the quantity is ambiguous, non-independent, and highly sensitive to implementation details such as batching, retries, or agent architecture.
What the task force asks researchers to report instead is the number of human respondents.

Comparisons show that adding agents does not reduce the error.
In Verasight (2026), increasing the number of agents from 500 to 5,000 changed the error against human polls only from 6.8 to 6.7 points.
The company concludes that the problem lies in the model itself and that a larger sample does not fix it.
Banawa (2026) at Nielsen Norman Group gave an open-ended question previously answered by 604 people to 120 synthetic respondents generated on Synthetic Users.
Both datasets led to the same overall conclusion, but the human participants used more than 50 distinct labels, whereas the synthetic answers were far less varied.

Developers also set the numbers with this property in mind.
In a 2024 podcast, Hugo Alves, a cofounder of Synthetic Users, said that the platform can go up to 30 synthetic users per study and that most people generate around six (That Business of Meaning 2024).
In a 2025 podcast, he said that a study runs at most 20 interviews, which is not a good number for segmentation, and that customers had asked from early on to compare several concepts with the same synthetic users (Getting to Alpha 2025).

## What Is Shown to Test Acceptance?

Text stimuli can be used in every product.
Buyers enter concept descriptions, headlines, value propositions, or price questions, and the personas answer them.
Rosala and Moran (2024) at Nielsen Norman Group wrote that the synthetic-user tools of that time were limited to replicating text-based attitudinal methods, specifically interviews and surveys.

Support for images and video differs by product.
Synthetic Users accepts images, video, and live websites, and Ask Rally offers image analysis from its USD 20 plan and video analysis from its USD 100 plan.
Yabble accepts up to 10 images per chat, and Toluna's creative pretest accepts online video, TV, social video, and high-quality animatics, but not storyboards.
DNP accepts product images, ads, package designs, and service UI screens, and Lakmoos tests social media ads, TV spots, and banners.
Qualtrics synthetic panels are the opposite case: they do not understand embedded images, videos, files, or hyperlinks.
Surveys that include such elements still complete, but the stimuli are not reflected in the answers.

Few materials state whether the model looks at an image directly or answers from a text description of it.
Only Yabble describes its processing: its own algorithms extract attributes from an image, such as colors, objects, text, style, mood, lighting, and scenes, and the personas react to those attributes.
In studies where acceptance depends on visual details, such as the color of a package or the composition of an ad, this processing may affect the results, yet buyers have no means of checking it.

Study formats fall into four groups.
These are surveys (Qualtrics and QuestionPro), conversational interviews (Synthetic Users, DNP, and Ipsos PersonaBot), groups with several personas (Ask Rally and Artificial Societies), and choice methods.
Among choice methods, Subconscious.ai lists discrete choice with price and MaxDiff, and Artificial Societies lists conjoint analysis.
Zappi takes 30 to 300 ideas in text or text-and-image form and returns a relative purchase likelihood tier, typically within 15 minutes for up to 100 ideas and within 45 minutes for up to 300.
Aaru turns an audience into a population of agents (10,000 in the example on its page) and puts simulation forward, in which agents treat each hypothetical as a condition.
Simile attaches to each prediction a confidence level from "High" to "Low" based on its predicted error.

Price acceptance is a weak area for synthetic answers.
In the Strat7 comparison (Research Live 2026) against a nationally representative sample of 3,000 people, prices given by synthetic respondents were generally 16% higher.
In a price-ordering exercise, purely synthetic respondents broke the logical order 68% of the time, and the rate dropped to 32.8% when they were blended with real data.

## Where and How Do the Services Charge?

Pricing falls into four forms.

| Form | Examples (as stated on official pricing pages) | What is consumed |
|---|---|---|
| Monthly self-serve | Ask Rally: USD 20, 100, and 500 per month (billed monthly in advance in US dollars, with no refunds for unused messages or features). SYMAR: EUR 99 per month for 400 credits, and from EUR 199 per month for 1,000 or more credits. Delve AI: from USD 317 to 2,499 per month (an extra 20% off with annual billing) | The maximum sample size per study. For SYMAR, one question or response costs 1 credit, a model with more reasoning costs 3, and videos and large PDFs consume more |
| Annual contract with tokens or credits | Synthetic Users: from USD 12,500 per year; a 10-question text interview uses 1,250 tokens and a 20-question survey response uses 160. Yabble: USD 8,900, 20,000, 40,000, and 80,000 per year | The number of interviews and responses. For Qualtrics, one synthetic response consumes 0.5 to 1.5 credits (up to 5 for B2B audiences) |
| Flat license with unlimited queries | Electric Twin: once an audience is built, there is no charge per query, study, or response; the price depends on the audiences built and the number of seats | Audiences and seats |
| Quoted contract | Aaru, Simile, Brox, YouGov Parallax, and NIQ publish no prices. Evidenza returns results within 72 hours after a proposal, and its self-service version is coming soon | Each contract |

In addition, some tools include synthetic responses in existing research licenses (QuestionPro offers them in its higher-tier licenses).

The tendency to consume plans per response or per interview can be read as a sign that buyers compare synthetic research with budgets for human research.
The pricing page of Synthetic Users places a typical 10-question interview at USD 80 to 120 through a recruitment agency, USD 30 to 60 through DIY recruiting, and USD 40 to 63 with synthetic users.
In this comparison, the price of a synthetic interview falls within the same range as recruiting people oneself.
Human online surveys are also inexpensive, with public prices starting at about one dollar per response (from USD 0.95 at Pollfish and from USD 1 for SurveyMonkey's panel).
The AAPOR task force (2026) writes that online panels remain relatively inexpensive, which complicates the economics of substituting synthetic responses for human participants.
Rothschild et al. (2025) look back on how the spread of non-probability online panels in the 2000s and 2010s lowered the cost of surveying people and thereby weakened the economic case for synthetic modeling.

On the vendor side, large research companies use their own panels as material for synthetic answers.
Toluna, YouGov, NIQ, Kantar, INTAGE, Dentsu, and Hakuhodo all build personas or twins from respondent or consumer data that they have collected.
This can be read as the asset built by collecting and selling human answers becoming the material for a second product, namely synthetic answers.
Among startups, Joon Sung Park, a cofounder of Simile, describes the offering only as "basically a SaaS product" and does not publish prices (Sequoia Capital 2026).

## How Closely Do Synthetic Answers Match Human Answers?

Every agreement figure that vendors publish is measured by the vendor itself.
Electric Twin reports 96% agreement with real people's answers, measured as 1 minus the mean absolute error.
This benchmark was pre-registered and the human survey data were collected independently by another company, but the professor who led it is Electric Twin's Chief Science Adviser.
Artificial Societies reports, in its own January 2026 evaluation, 86% distribution accuracy across 1,000 real-world surveys, against a human self-replication ceiling of 91%.
C3reve reports a mean absolute difference of 0.3 points on factual questions, but limits the claim itself by stating that this figure does not indicate accuracy for opinions, purchase intent, or sales.
Few products disclose their methods, and no product with independent third-party validation was found.

In third-party comparisons, topline errors ranged from 2 to 12 points depending on the comparison, and subgroup results and change over time missed further.
Pew Research Center (2026) found that the synthetic estimates differed from human estimates by an average of 12 percentage points.
On questions with a "not sure" option, human panelists were around four times as likely as the AI model to choose it.
In Verasight (2026), the topline error was 6.6 points, but it widened to 10.4 to 13.8 points when results were split by party, and the model pulled every group toward a 50/50 split.
In the Strat7 comparison, synthetic surveys were on average two to three percentage points away from the human headline results, yet they tracked changes between the 2025 and 2026 studies correctly only 47% of the time (Research Live 2026).

Academic studies report the same pattern.
Argyle et al. (2023) showed that a language model conditioned on attributes can reproduce the distribution of answers in US political surveys, and called the resulting samples "silicon samples."
Bisbee et al. (2024) reported that synthetic responses vary less than real survey responses, that regression coefficients often differ significantly from those of the human survey, that the distribution changes with minor changes in prompt wording, and that the same prompt gives different results over three months.
Wang et al. (2025) showed that language models that replace human participants misportray and flatten identity groups.
Acceptance testing often needs to know which segments reject a product and how strongly, more than it needs the average favorability.
If so, the parts that synthetic answers tend to miss, subgroups and variation, overlap with what acceptance testing needs to know.

## What Do Industry Bodies Ask of Vendors?

No government regulation that directly addresses synthetic respondents alone was found.
What was found is that the codes and reports of industry bodies require vendors to tell buyers when synthetic methods are used and to distinguish synthetic outputs from human answers.

The 2025 ICC/ESOMAR Code requires that clients be informed when AI is used in compiling datasets, analysis, or reporting, including the use of synthetic data and synthetic personas, and that the extent of human oversight be stated (Article 7(e)).
The 2025 MRS guidance requires practitioners to record and communicate to clients which research outputs are generated by AI and related technologies, including synthetic data, and which are due to human insight (requirement 33).
The 2026 Code of the Insights Association in the United States requires that AI-generated data be clearly distinguished from data directly derived from research subjects.
The AAPOR task force (2026) asks researchers who use AI as a respondent to disclose the number of human respondents, to describe synthetic data as such, and not to use terms such as "poll" or "survey" for data created by AI.
The same report calls vendors that market "samples" generated entirely from synthetic responses particularly risky and points out that the method is not a sampling design.

In Japan, an explanatory article by the Japan Marketing Research Association (2025) introduces the issues of synthetic data, but no guideline that sets disclosure rules for synthetic answers was found.
The summary of the AI Guidelines for Business, version 1.2, from the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry does not mention synthetic respondents directly.

From these codes, the points that buyers should check with vendors come down to three.
The first is the data that ground the answers (the data sources, and the type and version of the model).
The second is how the answers were compared with human answers (who compared them, and with which population, questions, and measures).
The third is whether the deliverables can show separately which outputs are synthetic and which come from people.

## Scope and Limitations

The 38 services are products found in English and Japanese, and the set is not exhaustive.
Vendor prices and specifications were checked on official pages on November 14, 2026, and may change.
For Synthetic Users, a current pricing page based on annual plans and tokens exists alongside a "historical" pricing document updated in April 2026, and this note follows the current pricing page.
For products that do not show prices on their official pricing pages, such as Artificial Societies, Simile, and Brox, prices reported by third-party articles were not included in the table.
The adoption figure (69% of researchers in a Qualtrics survey reported using synthetic responses in the past year) comes from a survey by a company that sells synthetic response products and does not measure accuracy.
Neither ESOMAR nor the Japan Marketing Research Association statistics include a separate category for the market size of synthetic respondent services.

## Unverified Items

- The correlation coefficients for Dentsu's People Model (0.47 with survey data only and 0.8 after fine-tuning) appear only in a secondary media article and not in the official announcement, so they are not used in the text.
- The full text of YouGov's white paper (November 2026) could not be retrieved, and its agreement rates are known only from company articles, so they are not used in the text.
- Gartner's forecast (that by 2028, 60% of product marketing teams will test messages with synthetic customer personas) could not be traced to the original source and appears only in the citing article, so it is not used in the text.
- How models process image and video stimuli could not be confirmed for products other than Yabble and Qualtrics, because the materials do not describe it.
- The prices and contract terms of TOPPAN, C3reve, Lakmoos, and DNP were not stated on the pages cited, so amounts seen on other pages or articles during collection are not used in the text.

## References

All sources accessed on 2026-11-14. The ledgers are in `source/review/synthetic-respondent-services/`.

### Vendor primary materials (pricing pages, documentation, terms, announcements)

- Synthetic Users. Pricing. https://www.syntheticusers.com/pricing
- Synthetic Users. Pricing (Historical). https://legal.syntheticusers.com/product-documentation/pricing-old
- Ask Rally. Home. https://askrally.com/
- Ask Rally. Terms of Service. https://askrally.com/terms
- Aaru. Simulation. https://aaru.com/simulation
- Simile. Confidence. https://www.simile.com/blog/confidence
- Electric Twin. FAQ. https://www.electrictwin.com/faq
- Electric Twin. Accuracy. https://www.electrictwin.com/accuracy
- Evidenza. Home. https://www.evidenza.ai/
- Artificial Societies. Pricing. https://societies.ai/pricing
- Artificial Societies. How accurate are AI personas. https://societies.ai/how-accurate-are-ai-personas
- Brox. FAQ. https://www.brox.ai/faq
- SYMAR. Pricing. https://www.symar.ai/pricing
- Delve AI. Pricing. https://www.delve.ai/pricing
- Lakmoos. Solutions. https://www.lakmoos.com/solutions
- Subconscious.ai. Home. https://www.subconscious.ai/
- FishDog. Home. https://fish.dog/
- Qualtrics. Synthetic Panels. https://www.qualtrics.com/support/survey-platform/distributions-module/synthetic-panels/
- Qualtrics. Edge Terms. https://www.qualtrics.com/edge-terms/
- Toluna. Synthetic Personas. https://tolunacorporate.com/ai-and-innovation/toluna-synthetic-personas/
- Toluna. Creative Pre-Test Instant. https://tolunacorporate.com/our-solutions/creative-and-campaigns/creative-pretest-instant/
- Yabble. Virtual Audiences. https://www.yabble.com/virtual-audiences
- Yabble. Pricing. https://www.yabble.com/pricing
- YouGov. Parallax. https://yougov.com/en-gb/business/products/parallax
- Ipsos. Discover Ipsos PersonaBot. https://www.ipsos.com/en/discover-ipsos-personabot
- NIQ. BASES AI Screener. https://nielseniq.com/global/en/products/bases-ai-screener/
- Zappi. Synthetic Idea Pre-Screener. https://learn.zappi.io/marketing-agents/synthetic-idea-pre-screener
- quantilope. 2026. quantilope launches synthetic Category Twins for early-stage research. https://www.quantilope.com/resources/quantilope-launches-synthetic-category-twins-for-early-stage-research
- Kantar. 2026. Digital twins for CMOs. https://www.kantar.com/north-america/inspiration/brands/digital-twins-for-cmos
- QuestionPro. Synthetic Responses. https://www.questionpro.com/help/audience/synthetic-responses.html
- Dentsu. 2025. Announcement of "People Model" (in Japanese). https://www.dentsu.co.jp/news/release/2025/0519-010884.html
- Hakuhodo DY Holdings. 2025. Announcement of "Virtual Seikatsusha" (in Japanese). https://www.hakuhodody-holdings.co.jp/news/corporate/2025/11/5993.html
- INTAGE Holdings. 2026. Launch of the PerSearch beta (in Japanese). https://prtimes.jp/main/html/rd/p/000000615.000001551.html
- DNP. Persona Insight (in Japanese). https://www.dnp.co.jp/biz/products/detail/20176748_4986.html
- TOPPAN. 2026. Launch of a CX support service using AI personas (in Japanese). https://www.holdings.toppan.com/ja/news/2026/10/newsrelease261013_2.html
- Fujifilm Business Innovation. 2026. Beta release of Persona AI (in Japanese). https://www.fujifilm.com/fb/company/news/release/2026/83913
- C3reve. 2026. Official release of Persona AI (in Japanese). https://prtimes.jp/main/html/rd/p/000000174.000034517.html

### Developer statements

- That Business of Meaning. 2024. Hugo Alves on real and synthetic users. https://thatbusinessofmeaning.substack.com/p/hugo-alves-on-real-and-synthetic
- Getting to Alpha. 2025. Episode with Hugo Alves. https://www.buzzsprout.com/222312/episodes/17548993
- Sequoia Capital. 2026. Simulating humans at scale: Simile's Joon Sung Park. https://sequoiacap.com/podcast/simulating-humans-at-scale-similes-joon-sung-park

### Industry bodies and public agencies

- ICC/ESOMAR. 2025. ICC/ESOMAR International Code on Market, Opinion and Social Research and Data Analytics. https://standards.esomar.org/assets/documents/icc-esomar-code-2025.pdf
- Market Research Society. 2025. Guidance on Using AI and Related Technologies. https://www.mrs.org.uk/pdf/AI+RelatedTechnologies_MRSGuidance_April2025.pdf
- Insights Association. 2026. Code of Standards and Ethics. https://www.insightsassociation.org/About-Us/Code-of-Standards
- AAPOR Task Force on Responsible AI Integration in Survey Research. 2026. Responsible AI Integration in Survey Research. https://aapor.org/wp-content/uploads/2026/05/Responsible-AI-Integration-In-Survey-Research.pdf
- Japan Marketing Research Association. 2025. The current state and future issues of synthetic data (in Japanese). https://www.jmra-net.or.jp/activities/trend/international/20250121.html
- Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry. 2026. AI Guidelines for Business, version 1.2, summary (in Japanese). https://www.soumu.go.jp/main_content/001064299.pdf
- DRNO. 2026. Definitions, standards and audits for synthetic data. https://www.mrweb.com/drno/news40566.htm

### Research firms and third-party comparisons

- Pew Research Center. 2026. Can AI Stand In for Human Survey-Takers? Not Really. https://www.pewresearch.org/data-labs/2026/09/30/can-ai-stand-in-for-human-survey-takers-not-really/
- Pew Research Center. 2026. Methodology: Silicon samples. https://www.pewresearch.org/data-labs/2026/09/30/methodology-silicon-samples/
- Verasight. 2026. Can AI digital twins replace human respondents? https://www.verasight.io/reports/can-ai-digital-twins-replace-human-respondents
- Research Live. 2026. Synthetic data survey responses significantly different to humans. https://www.research-live.com/article/news/synthetic-data-survey-responses-significantly-different-to-humans/id/5151157
- Banawa, Rachel. 2026. Synthetic Users vs. Human Respondents. Nielsen Norman Group. https://www.nngroup.com/articles/synthetic-users-vs-human-respondents/
- Rosala, Maria, and Kate Moran. 2024. Synthetic Users: If, When, and How to Use AI-Generated "Research". Nielsen Norman Group. https://www.nngroup.com/articles/synthetic-users/
- Research Live. 2024. Researchers predict synthetic responses will dominate in coming years. https://www.research-live.com/article/news/researchers-predict-synthetic-responses-will-dominate-in-coming-years/id/5131923

### Prices of human research

- Pollfish. Pricing. https://www.pollfish.com/pricing/
- SurveyMonkey. Audience. https://www.surveymonkey.com/mp/audience/

### Academic literature

- Argyle, Lisa P., Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. 2023. "Out of One, Many: Using Language Models to Simulate Human Samples." *Political Analysis* 31 (3): 337–351. https://doi.org/10.1017/pan.2023.2
- Bisbee, James, Joshua D. Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M. Larson. 2024. "Synthetic Replacements for Human Survey Data? The Perils of Large Language Models." *Political Analysis* 32 (4): 401–416. https://doi.org/10.1017/pan.2024.5
- Rothschild, David M., Trent D. Buskirk, Stephanie Eckman, D. Sunshine Hillygus, Frauke Kreuter, and David Lazer. 2025. "Successfully Navigating the Disruption AI Will Bring to Survey Research." *The Survey Statistician* 92: 30–44. https://researchdmr.com/files/SurveyResearchAI.pdf
- Wang, Angelina, Jamie Morgenstern, and John P. Dickerson. 2025. "Large Language Models That Replace Human Participants Can Harmfully Misportray and Flatten Identity Groups." *Nature Machine Intelligence* 7 (3): 400–411. https://doi.org/10.1038/s42256-025-00986-z

## Related Notes

- [What Counts as Evidence in New-Business Acceptance Research: From Words to Behavior, and Then Generative AI](https://jamjet-iqbal.com/en/notes/new-business-acceptance-research/): what new-business acceptance testing treats as evidence, why synthetic answers sit below verbal evidence, and research on inflated willingness to pay
- [Conducting User Interviews: What Academia and Industry Share and Where They Diverge](https://jamjet-iqbal.com/en/notes/design-user-interview-practice/): the practice of user interviews, and why synthetic users do not replace people