Synthetic respondents, tested against real ones: three case studies — Radius and MilkPEP

Mike Patterson and Kiki Rell · 2024 · 43:28 · ▶ Watch

Summary

Central question. Can a large language model (LLM), told to play a real survey respondent, give the answers that respondent actually gave? Radius, a market research firm, ran three paired tests. In each, real people answered first. Then the firm built an AI twin for each person from their demographics, attitudes and behaviours, asked the twins the same questions, and compared.

Main thesis. Synthetic respondents are not ready to replace real ones. The webinar’s three cases give a split verdict:

• Aggregate numbers can look good, individual answers do not. In a concept test, about a third of individual ratings matched, but top-two-box scores landed within a point or so on four of six concepts.
• The model misses what is surprising or persuasive to real people. In a messaging test for MilkPEP (the US milk marketing board), the claim that tested best with real mothers, a hydration claim, was rated “surprising” by none of the AI twins.
• Qualitative answers are flat. The AI women gave “functional” goals and concerns and all sounded alike. The real women gave emotional, personal reasons, such as setting an example for a daughter.

Close. Patterson’s rule: validate synthetic against real respondents for your own topic, metric and study type before trusting them. The buyer, Kiki Rell of MilkPEP, is blunter (“too good to be true”). Both see AI helping research in other ways (coding open-ended text, transcription, data quality, better choice models), and both can imagine synthetic respondents supplementing real ones, not replacing them.

Stance. Sceptical, evidence-first, but open. Patterson calls it “early in the game”. Rell is in “wait and see” mode. The session is also a vendor and client speaking together, and the client’s marathon is promoted at the end. Read it as practitioner testimony, with small samples, not as a peer-reviewed study.

Map

#	Point
1	Opening: stakeholder pressure on insights teams, “is there a better way?“
2	What a synthetic respondent is, and the many names for it
3	Radius method: real respondents fed into an LLM as personas
4	Case 1, concept test (gaming controllers, six concepts, ~200 each): individual match ~1/3, aggregate close on 4 of 6
5	Case 2 background: MilkPEP messaging test with moms (believable, surprising, increases consumption)
6	Case 2 results: believability similar, surprise and consumption diverge, hydration claim missed
7	Case 3 background: the Every Woman’s Marathon, a 36-woman online board
8	Case 3 results: six AI personas are functional and uniform; real women are emotional and varied
9	What the real research changed: name, design, shoes, an influencer series
10	Conclusions: real beats synthetic; aggregate OK, individual weak; validate
11	Buyer’s view; other uses of AI in research
12	Q&A: consumer vs B2B; what would change her mind; replace or supplement

Putting it together

Mental model. A synthetic respondent is a prediction of a person, built from a few labels about them plus everything the model absorbed in training. Three things can then happen. If the question is about common knowledge, the model’s prior and the person’s real answer coincide and it looks like the twin “works”. If the question is about what is new, surprising or emotionally loaded for that group, the model answers from the generic text of the internet and misses. And because every twin is drawn from the same prior, the twins sound more alike than real people do. This explains the three observed patterns: aggregates match, individuals do not, and qualitative answers go flat.

Open tensions.

• The aggregate match may reflect the model’s prior and the base rates in the labels, not understanding. Without a simple baseline, such as predicting the average, a match is hard to read.
• The tests are small and single-model: six concepts, six statements, six personas. The talk does not name the model.
• “Validate against real respondents” is sound, but it needs real data, which is the cost the tool was meant to save.

Reading path.

• Start: ideas 1 and 2 (method, then aggregate vs individual).
• Then: idea 3 (the milk claims) and idea 4 (qualitative flatness).
• Deeper: idea 5 (validation), plus the sibling talks on this theme: Prediction-Powered Inference (a way to combine a small real sample with many synthetic answers while keeping statistical guarantees), Joon Sung Park on Generative Agents (2023) and on Simile’s “8 billion digital twins” (2026), and Ishan Anand on Persona Engineering.

Part 1: Before you read

Key terms

• Synthetic respondent. The speakers’ own list of names: “virtual respondents digital personas silicone sample … digital twins”. Caption “silicone” is silicon sample, the term from the academic literature.
• Persona conditioning. Telling the LLM to take on the traits of a particular person, then asking it questions.
• Top-two-box score. The share of respondents choosing the top two points of a rating scale (here 5 and 4). The standard way to report concept tests.
• Individual-level vs aggregate-level match. Individual: did this twin give this person’s rating? Aggregate: do the group percentages agree? A model can pass the second without passing the first.
• Chance level. On a five-point scale, random answers match about 20% of the time.
• Believability, surprise, increase consumption. The three measures in the MilkPEP test. A claim that scores on all three is the winner.
• Concept test. Showing consumers a product idea and measuring interest. Online board. A multi-day qualitative panel where recruited participants answer questions in writing.

People named

Person	Role in the talk
Mike Patterson	Chief research officer, Radius; host and moderator
Kiki Rell	MilkPEP (joined 1998), consumer insights for the “Got Milk” campaign; the research buyer
Amanda Gorman	Poet who launched the marathon, mentioned in passing

The speakers’ own frame

They are not academics. Patterson sells research and wants to know whether claims by synthetic-sample vendors hold up. He says “various companies will suggest that they can” replace real respondents, and “there’s also been various publications that have purported” a match. Rell is the sceptical client who has worked with Radius for 20 years. Both are testing a commercial claim, not a theory.

The current debate

The webinar was June 2024. Since then the evidence has grown, and it points the way the speakers did, with qualifications.

• The optimistic case. Argyle and colleagues showed in 2022 that GPT-3, conditioned on real people’s backstories from large US surveys, could reproduce subgroup response patterns, a property they call “algorithmic fidelity” (Out of One, Many, arXiv; published in Political Analysis, 2023).
• The sceptical case. Bisbee and colleagues tested ChatGPT personas on feelings toward 11 sociopolitical groups. Averages tracked the American National Election Study, but responses varied less than real ones, regression coefficients often differed, and the same prompt gave different results three months later (Synthetic Replacements for Human Survey Data?, Political Analysis). Dominguez-Olmedo, Hardt and Mendler-Dünner found that once answer-order bias is removed, 43 models drift toward uniformly random answers, so apparent alignment with some subgroups can be an artifact (NeurIPS 2024).
• Aggregate vs individual, now measured. Reviews and benchmarks report that personas capture distributions in stable domains like values and demographics but fail at individual prediction (summary of a persona-survey review; a 2025 Chilean benchmark finds the sample approximates a probability sample overall but with large item-level variation). A 2026 preprint argues that many studies lack a naive baseline and proposes comparing LLMs against a demographic-conditional predictor (When Synthetic Users Fail; preprint, not peer-reviewed at time of writing).
• Industry testing. Verasight’s 2025 report found subgroup errors above 10 percentage points against real crosstabs, and that adding more persona detail could make accuracy worse (Synthetic Sampling Report II).
• Where it stands. The pattern in the webinar (aggregate fine, individual weak, subgroups unreliable, prompt-sensitive) is now the mainstream reading. The live argument is about use: as a hypothesis generator, as a way to extend a small real sample, or as a replacement. The first two have defenders. The third has few outside vendors.

The ideas

Five arguments run through the talk. Here are the traditions behind them, including some the speakers never name:

 SURVEY METHODOLOGY ────────┐             ┌──── SIMULATING HUMANS WITH LLMs
 total survey error         │             │     silicon samples · algorithmic fidelity
 response process models    ▼             ▼
                   ┌─────────────────────────────┐
 MEASUREMENT  ───► │  RADIUS: do AI twins give   │ ◄──── HOW LLMs ARE TRAINED
 VALIDITY          │  the answers real people    │       priors · RLHF · mode-seeking
 Cronbach·Campbell │  gave?                      │
                   └─────────────────────────────┘
                        ▲                  ▲
       QUALITATIVE RESEARCH          THE MARKET-RESEARCH BUSINESS
       thick description · emotion   concept testing · vendor claims

Where each is covered:

• Simulating humans with LLMs: idea 1
• Survey methodology and the aggregate/individual gap: idea 2
• How LLMs are trained, and why they miss the unfamiliar: idea 3
• Qualitative research and emotion: idea 4
• Measurement validity and the market for synthetic samples: idea 5

1. The method: build a twin of a real person, then compare

In short. Radius does not invent personas. It takes real survey respondents whose demographics, attitudes and behaviours it already knows, feeds those into an LLM, and asks the model to become each one. Because the real person’s own answer exists, every twin can be scored against its source. This paired design is the strongest thing in the talk.

🎙 In the talk.

• The term can mean many things: “virtual respondents digital personas silicone sample … digital twins”.
• Their method: “we take real survey respondents where we know their demographics their attitudes and their behaviors”, then “we ask the AI to take on the Persona”.
• Why test at all: “various companies will suggest that they can” replace respondents.

📚 Deeper theory.

• Paired comparison is rarer than it sounds. Many vendor claims compare a synthetic group with a real group. Pairing each twin to its source allows the individual-level check in idea 2.
• The twin is built from labels. A handful of labels (age, income, attitude scores) is a thin description of a person. Whatever else the twin “knows” comes from the model’s training.
• The sibling talks. Park’s Generative Agents (2023) built agents with memories and routines. Park’s Simile talk (8 billion digital twins, 2026) concerns twins built at far larger scale. Anand’s Persona Engineering is about how the prompt shapes the persona. Radius’s design sits at the thin-description end.

📚 Roots: Simulating human samples with language models

The problem. Surveys are slow and costly. If a language model has read much of what people write, perhaps it can stand in for them.

The point. Radius tests this idea directly, and the results (ideas 2 to 4) are a field test of the claim made below.

The chain of thinkers

1. Argyle et al., “Out of One, Many” (2022; Political Analysis 2023). Conditioning GPT-3 on real survey backstories reproduced subgroup patterns. They coined algorithmic fidelity, the idea that the model’s biases are fine-grained and track demographics (arXiv). Radius’s design is a close relative: persona from real survey data, then compare.
2. Aher, Arriaga and Kalai (2023), “Using Large Language Models to Simulate Multiple Humans.” Replicated classic experiments such as the Milgram shock study and the ultimatum game with model “participants”.†
3. Horton (2023), “Homo silicus.” Economists proposed treating LLMs as simulated economic agents for cheap pilot experiments.†
4. Bisbee et al. (2024). The first widely cited warning: good averages, too little variance, results unstable across time and wording (Political Analysis).
5. The later view. Respondent simulation is a tool with a limited validity range that must be checked per use (idea 5).

Where others push back. Defenders of simulation say the Radius result depends on the model, the prompt and the persona detail. A richer persona or a better model might do better. That is a fair objection, and the talk’s single, unnamed setup cannot answer it.

What they leave out. Which model, what prompt, how many runs per twin, and whether temperature was varied. All matter. Answers from a model vary from run to run.

Check yourself

• Why is a paired design stronger than comparing two group averages?
• What information does a twin have that did not come from the real person’s labels?
• What would you need to know about the prompt before repeating this test?

2. Aggregate match, individual miss (Case 1: the concept test)

In short. In a test of six gaming-controller concepts with about 200 real consumers each, the AI twins matched their real counterparts about one time in three. That is above the 20% chance level but, as Patterson puts it, “not great”. Yet the group-level top-two-box scores were close on four of six concepts. So aggregate reporting hid a weak individual-level match.

🎙 In the talk.

• Method: each real respondent rated “How likely they would be to seek information about the concept” on a five-point scale, and each twin was exposed to the same concepts.
• Individual level: “by chance we would expect about 20% of the time for those to match” and “about a third of the time our real respondents match” the twins.
• His verdict: “it’s above chance but it’s not great”, and he would like to see the number “about twice as high”.
• Aggregate level: “really nice congruence”, with 76% of real respondents and 75% of twins giving a top-two-box score on Concept A.
• “for two of the measures we do see some differences”.

📚 Deeper theory.

• Why aggregates can match when individuals do not. Errors can cancel. If a twin overrates half the time and underrates half the time, the group average stays right while individuals are wrong. A model that predicts “the typical answer for this kind of person” will often get the group mean right and the individual spread wrong.
• Was the aggregate match meaningful? Concept tests have a known property: most concepts score in a narrow band, and the order of concepts matters more than the level. Matching four of six on level is helpful, but a naive baseline would show what that is worth. A 2026 preprint makes exactly this point: compare against a simple demographic-conditional predictor, because a single similarity number can flatter a model (arXiv). The webinar gives no baseline.
• What matters for the buyer. A client picking among concepts needs the ranking and the gap, and often the segment differences. Segment cuts are individual-level data in aggregate clothing, and the related industry finding is that subgroup errors are larger than overall ones (Verasight).
• Six concepts is a tiny test. With six points, a matching rank order is likely by luck alone for some correlations.†

📚 Roots: Survey methodology and the response process

The problem. A survey answer is not a readout of a stored opinion. It is made at the moment of asking.

The point. If real answers are partly constructed in the moment, then a twin that “knows” the person’s labels is predicting a noisy event, and individual-level agreement has a ceiling below 100%. Radius treats 100% as perfect alignment, but real people are not perfectly repeatable either.

The chain of thinkers

1. Philip Converse (1964), “The nature of belief systems in mass publics.” Many respondents give unstable answers on issues they have not thought about, which he called “non-attitudes”.†
2. Tourangeau, Rips and Rasinski (2000), The Psychology of Survey Response. A four-stage model: understand the question, retrieve, judge, report. Context and wording affect each.†
3. Jon Krosnick (1991), satisficing. Tired or unmotivated respondents take shortcuts. This adds individual noise.
4. Robert Groves, total survey error. Measurement, coverage, sampling and non-response errors all add up. A synthetic sample adds a new one: the model’s own error.†

Where others push back. If real individuals are noisy, a one-third match could be close to the achievable ceiling. The right comparison is a test-retest match, asking the same real people twice. The talk does not report one.

What they leave out. Test-retest agreement among real respondents, and a naive baseline.

Check yourself

• Why can a group average be right when most individuals are wrong?
• What does a 33% match mean if real people only match their own earlier answers 50% of the time?
• Which business question does a concept test answer, and does aggregate agreement answer it?

3. The milk claims: the model misses what is new, surprising and persuasive (Case 2)

In short. MilkPEP tested six statements about milk with moms on three measures: believability, surprise, and likelihood to increase consumption. The twins ranked believability like the real moms (same top and bottom statement). On surprise and consumption they diverged. The standout was a hydration claim. Real moms found it surprising and compelling. No twin found it surprising, and few picked it as a reason to drink more. The best statement overall would have been missed.

🎙 In the talk.

• Rell on what mattered: believability, how surprising the claim was, and the aim “to drive the consumption of fluid milk”.
• Believability: both groups rated the same statement most believable and the same one least believable, “some similarities but also definitely some differences” in the middle.
• Rell’s note on the least-believable claim: the twins “go even lower than the real respondents”, which was “a little bit of a flag” because “you would think with AI that they had access … to all the information that’s out there”.
• Surprise: “none literally none of the synthetic respondents” found the hydration statement surprising.
• Consumption: the real respondents found the hydration claim compelling, but “very few” twins selected it, “a pretty big Miss”.
• Consequence: the hydration message “when you looked at the three factors combined was the top performer”, so going the synthetic route would have missed it.
• Her explanation for surprise: “that’s a human emotion right that’s not an artificial reaction”.

📚 Deeper theory.

• Why surprise fails. Surprise is relative to what a group knows. A model answers from what it knows. It has read the milk-and-hydration literature, so the claim is not new to it, whereas moms have not been told it. The model confuses “available in my training data” with “known to this person”. Rell’s point that the model has “access” to the literature is the cause of the error, not a sign of accuracy.
• Why consumption fails. Whether a message changes behaviour depends on a person’s habits, doubts and context. Stated intent is already a weak predictor of behaviour in human research; the twin predicts stated intent from labels, a prediction of a prediction.
• Believability worked best. It is the measure closest to general knowledge (“is this claim plausible?”), where model and public priors overlap. The pattern across the three measures follows how close each is to common text.
• The stakes. The cost was a missed winner, not a bad average. Research is often used to find the outlier idea, the one that does better than expected. A tool that regresses toward the typical is least useful exactly there.

📚 Roots: How LLMs are trained, and why they regress to the typical

The problem. An LLM predicts likely text. Typical answers are more likely than unusual ones, and later training stages push further toward the typical.

The point. The hydration miss is what you would expect from a system that learned what is commonly said about milk, not what a particular audience finds fresh.

The chain of thinkers

1. Pre-training. The model learns the distribution of text. Common views dominate. Minority and local views are thin.
2. Santurkar et al. (2023), “Whose Opinions Do Language Models Reflect?” Compared LLM answers with US opinion polls and found that models lean toward some demographic groups, even when told to answer as another.†
3. RLHF and alignment training (Ouyang et al. 2022; Kirk et al. 2023). Tuning to human preferences makes outputs more agreeable and less varied.†
4. Bisbee et al. (2024). Show reduced variance and unstable results in persona survey responses (source).
5. Dominguez-Olmedo et al. (2024). Show that after controlling for response-order bias, models look close to random, so apparent subgroup alignment is partly artifact (source).

Where others push back. Some say prompt design could draw out minority views, or that fine-tuning on the client’s own survey data would fix this. That would be a different product from the one tested, and the fix needs the real data the tool was meant to replace.

What they leave out. Whether the hydration claim was in the model’s training text, and whether another model would have found it surprising.

Check yourself

• Why might more knowledge in the model make the twin worse at predicting surprise?
• Why did believability line up better than surprise?
• What does a “missed winner” cost compared with a “wrong average”?

4. The flat voice: qualitative research and emotion (Case 3)

In short. MilkPEP ran a 36-woman online board to design a women-only marathon. Radius then used attributes from six of those women to build twins and ask the same questions. The twins gave “functional” goals (endurance, strength, flexibility) and concerns (stay motivated, avoid injury), and gave almost the same answers across very different women. The real women spoke about emotion, self-doubt and family, and the research changed how the race was designed.

🎙 In the talk.

• Sample: 36 recruited women “at different stages of running ability”, from seasoned to “never say never”. The twins used “six of the women”.
• Twin answers: “very very functional in nature”, and “very little difference” between different groups, with little nuance.
• Rell on the real answers: “it’s all about emotional benefits it’s very holistic it’s very personal”. And on a reason like setting an example: “synthetic respondents not have synthetic children”.
• On concerns: real women expressed doubts like “am I good enough” and “I don’t want to be crowded in”; the twins’ answers were “this could be anybody”.
• Names: twins “were all over the place”, while real women converged on “every woman’s Marathon”.
• Outcomes cited by Rell: over 6,000 registered with a 7,000 cap, “41% of the register or first timers”, a time limit generous enough that “you can almost walk it”, a shoe designed for women’s feet, and a series with a self-described non-runner influencer. These are the client’s claims and are not independently checked.
• Her conclusion: “There’s a lot of things we would have missed”.

📚 Deeper theory.

• Function before feeling. Asked “what are your goals?”, an LLM produces the textbook answer. The real answer to “why do you run?” is a story. Qualitative research exists to elicit that story and its context.
• Variation inside the group is the point. An online board is built to show how a seasoned runner and a doubtful beginner differ. Twins drawn from one prior flatten that.
• The sample is small. Six twins, compared by eye, against 36 women. The finding that they sound alike is consistent with the homogeneity literature (idea 3), but six is thin evidence on its own.
• A different job. Hypothesis generation from a model is a legitimate use. Rell says “I think we can use it … to generate hypotheses”. The test showed what the model could not do: be surprised by the real audience.
• Embodied experience. The “daughter” example is vivid because the twin has no child. The same limit would apply to any lived experience not captured in the labels: caregiving, fear in public space, shoe fit. The shoe finding (women’s shoes are “just men’s running shoes in a smaller size”, in her words) came from talking to women.

📚 Roots: Qualitative research and what it captures

The problem. Some knowledge is about meaning: what a thing is for the person doing it. It cannot be read off demographics.

The point. Radius’s qualitative result is a practical case of this: a model gives a generic account of exercising, while the board gives local meaning.

The chain of thinkers

1. Clifford Geertz, “Thick description” (1973). Understanding an action means understanding the layers of meaning around it, not just recording it.†
2. Glaser and Strauss (1967), grounded theory. Categories should emerge from participants’ own words, not be imposed.†
3. Thomas Nagel (1974), “What is it like to be a bat?” Subjective experience is not captured by third-person description. Used loosely here, as the philosophical root of “synthetic respondents not have synthetic children”.†
4. Qualitative-sampling logic. The aim is range of meaning, not frequency. A tool that returns the modal view fails the aim by design.

Where others push back. Practitioners of AI-assisted qualitative work argue that LLMs are useful for analysis of real transcripts. The webinar agrees on that, and lists open-text analysis and transcription as clear gains. The objection is only to simulating the participants.

What they leave out. What a stronger prompt, such as full interview histories, would have produced.

Check yourself

• Why does the “daughter” example show a limit that more demographic labels would not remove?
• Is uniformity across twins a model problem, a prompt problem, or both?
• Which parts of qualitative work did the speakers say AI can help with?

5. Validate first, supplement at most: what the speakers conclude

In short. The conclusion is practical. Real respondents do better. Aggregate agreement is sometimes fine, individual agreement is poor, emotion and nuance are lost, and you must check any synthetic approach against real respondents in your own industry, topic, study type and metrics. Both speakers see a future for AI in research, not as a replacement for people.

🎙 In the talk.

• “pretty much universally we’re finding that real respondents just perform better” than synthetic.
• On idea generation: whether synthetic respondents can generate “new ideas and New Concepts” he doesn’t “really know if they’re there”.
• “individual level data is not great”.
• Method: “you have to validate the results”, by characterising real respondents and comparing the answers, for “the industry that you’re in the topic area the type study” and metrics.
• Rell: “if it sounds too good to be true it probably is”, and “prefer my respondents like I prefer my milk”.
• Elsewhere AI helps: coding open ends, transcription, analysing large data sets, better screening, and use in “choice models our conjoint studies”.
• Consumer vs B2B: models are “more consumer based”, so he expects more success there, though that is his opinion.
• Replacement: “way down the road potentially”, but more likely to “supplement” real respondents; “we need to base AI on humans”.

📚 Deeper theory.

• Validation is construct validity. The question is whether the synthetic measure tracks the thing you care about, in the setting where you use it. That is a property of use, not of the tool. Passing for gaming controllers says little about milk claims.
• Supplement is a statistical question. If you have a small real sample and many cheap synthetic answers, you can use the synthetic ones to sharpen estimates only if you measure and correct their error against the real data. That is what Prediction-Powered Inference formalises, and it is the stronger form of Patterson’s “validate”. The sibling talk on it explains the method. In this view the synthetic answer is a helper whose bias is estimated, not a respondent.
• Where the field has moved. Since 2024 the distinction has sharpened between simulating populations (more defensible at aggregate level in stable domains) and simulating individuals (weak). The webinar’s pattern fits that split.
• A market warning. Rell’s “too good to be true” reflects a buyer’s view of vendor claims. Look for vendors who publish paired individual-level tests with baselines, and who say where the method fails.

📚 Roots: Measurement validity

The problem. A new measuring method must show it measures what it claims to.

The point. Patterson’s advice to compare twins with real respondents is, in effect, convergent validity testing. What is still missing is discriminant validity: does the twin differ from other people the way the real person does?

The chain of thinkers

1. Cronbach and Meehl (1955), “Construct validity in psychological tests.” A measure’s meaning rests on a network of expected relationships, tested over time.†
2. Campbell and Fiske (1959), multitrait-multimethod matrix. Show that a new method agrees with established methods (convergent) and does not agree where it should not (discriminant).†
3. Bisbee et al. and later benchmarks. Show that agreement on means is not enough; relationships between variables must also hold (source).
4. Angelopoulos et al. (2023), Prediction-Powered Inference. Gives a way to use machine predictions while keeping valid error bars, by estimating the predictor’s error on a real labelled sample. See the sibling talk.

Where others push back. Vendors say validation per use is too costly and that their model is validated in general. The reply is that validity does not transfer between domains, as the Radius cases show.

What they leave out. Costs. How large must the real validation sample be, and when is it cheaper to run the real study?

Check yourself

• Why does passing a concept-test check say little about a messaging test?
• How is “supplement real respondents” different from “replace” in statistical terms?
• What would a baseline, such as predicting the average, add to Radius’s tests?

---

About the source and how this page was made

• Speaker: Mike Patterson and Kiki Rell
• Event: Radius webinar, “Testing the potential of AI-generated synthetic respondents”
• Date: 2024-06-26
• Length: 43:28
• Channel: Radius
• Video: Webinar: Exploring AI and Synthetic Respondents — 3 Case Studies with Radius and MilkPEP

Source notes

Speakers

Label	Who	Role	Basis
Mike Patterson (“Michael”)	Chief research officer, Radius	Host, presents Cases 1 and 2 analysis, conclusions	Self-introduction; channel is Radius
Kiki Rell	MilkPEP (joined 1998); consumer insights for “Got Milk”	Client and co-presenter; Cases 2 and 3 background	Introduced by Patterson. Spelling from captions (“Ryell”, “Rell”); may differ

Identification and confidence. High for both speakers’ roles, which they state. Medium for the spelling of Kiki’s surname, since captions vary.

Dating. Video upload date 26 June 2024. The marathon is described as upcoming (“November 16th in Savannah”), consistent with a mid-2024 date.

Caption corrections. “silicone sample” is silicon sample; “sythetic” is synthetic; “Mel” is likely a garbled name or aside and was not used; “never won a marathon” is likely never run one; “milk TB” is MilkPEP; “got milk” campaign is as spoken; “Sy” prefixes are caption splits of synthetic; “more consumer-based … than consumer based” is a slip for B2B; “pany hose” is unclear.

Slips and cautions.

• At Patterson says the hydration message would have been missed “if we gone down the synthetic route”. The case supports this, but only for this study and these models.
• The marathon figures and outcomes are the client’s own and are not independently checked.
• The talk does not name the LLM, the prompt, the number of runs, or the six-statement text, so nothing here can be reproduced.
• The test design is persuasive but small: six concepts, six statements, six personas.

🎙 marks what the speaker says. 📚 marks added background and the wider debate, not the speaker’s claim. † marks background from general knowledge that was not checked against a source.