PaperlayerView exact revision on arXiv

Source-linked AI summary

AI-Moderated Interviews for Market Research and Digital Twins Calibration

Yuting Deng, Jingxuan Liu, Olivier Toubia, Naman Jain

arXiv:2609.29143v1cs.CYcs.AIcs.HCcs.MA

TL;DR

The paper asks whether AI-moderated interviews can match human-interview richness at a fixed budget, against a backdrop of survey methods missing the “why” behind consumer choices. It compares interview approaches and finds richer interview data and greatest customer-need recovery under the same market-cost budget.

  • Problem

    The paper asks whether AI-moderated interviews can replicate the richness of human-moderated interviews while holding budget constant; survey methods can miss the “why” behind consumer choices.

    Source: §1 · pp. 1–2

  • Method

    The paper compares AI-moderated interviews with static and human-moderated interviews on interview-data quality and customer-need recovery under market-cost budgets.

    Source: §5 · pp. 22–23

  • Results

    AI moderation improved interview-data quality over static interviews and recovered the most customer needs under the same market-cost budget, at roughly 20% of human moderation’s per-interview cost.

    Source: §5 · pp. 22–23

  • Takeaways & Limitations

    AI moderation has potential to generate rich qualitative data for consumer research.

    Source: §5 · p. 24

    Emotional engagement is identified as AI moderation’s main limitation.

    Source: §5 · p. 23

Abstract

from arXiv · show

AI-moderated interviews are emerging as a scalable market-research method for generating consumer insights and building consumer "digital twins." Yet it remains unclear whether they match human-moderated interviews or improve on simpler, static data collection methods. In a pre-registered, between-subjects study (N = 317) with three industry partners, we compare AI-moderated (N = 139), human-moderated (N = 24), and static interviews (N = 154). AI moderation matches human moderation in depth, covers more themes, and, holding budget constant, recovers significantly more customer needs than human moderation or static interviews. However, participants sound more emotionally engaged when speaking to a live human. We then create digital twins using interview data and evaluate each twin against the participant's own held-out responses to six real-world marketing stimuli. We find that digital twins created from AI-moderated interviews predict consumer responses better than demographics-only personas. However, the additional richness from AI moderation does not translate into better quantitative predictions compared to static interviews. By analyzing open-ended thoughts generated from humans versus their twins, we find that prediction errors are connected both to differences in (self-reported) thinking styles between twins and humans, and to gaps between training and validation data (i.e., asking questions that are too far out of distribution).

1. Introduction

The paper compares AI-moderated interviews with human and static methods for eliciting consumer insights, then evaluates how interview-based digital twins predict human responses. It also examines prediction errors in relation to training-data relevance and differences in human and twin reasoning styles.

1. Introduction

  • The study compares AI-, human-, and static-interview methods on volume, thematic breadth, elaboration depth, customer needs, and speech-based affect.It also tests whether interview data can support digital-twin predictions about marketing outcomes.Source: §1 · pp. 2–3
  • AI-moderated interviews elicit longer, broader, and deeper responses than static interviews, match human interviews in depth and thematic breadth, and recover the most customer needs under a fixed budget.The reported comparison says AI interviews achieve at least as broad thematic coverage as human-moderated interviews.Source: §1 · p. 4
  • Participants sound more positive, activated, and assertive when speaking live to a human; AI- and static-interview responses are emotionally indistinguishable.Thus, AI matches human moderation on interview content, but not on the emotional experience of talking to a person.Source: §1 · p. 4
  • Digital twins built from AI- and static-interview data were evaluated against participants’ held-out responses to real-world marketing stimuli.The stimuli included direct-mail images, commercials, and claim-ranking tasks; twins were not built for the smaller human-moderated sample.Source: §1 · p. 4
  • Qualitative-data digital twins predict better than demographics-only twins, but AI-moderated interview data does not improve quantitative predictions over static-interview data.Source: §1 · p. 4
  • The authors identify missing validation-task information in training data and limited reasoning fidelity as two sources of twin–human discrepancies.They define fidelity as the extent to which a twin replicates the human’s self-reported thought processes.Source: §1 · p. 4
  • Humans rely more on fast, intuitive System 1 reactions, while twins tend toward deliberative System 2 reasoning, with the gap widest for images and videos.Improving reasoning fidelity is beyond the paper’s scope.Source: §1 · p. 4
  • Adding a training task similar to the holdout task improves digital-twin prediction.The authors present this as a practical way to make twins more useful in market research.Source: §1 · pp. 4–5

2. Methods

The methods compare AI-, human-, and statically moderated interviews using a shared topic guide, then assess interview-process measures and operationalize customer needs and interview quality. They also construct interview-based digital twins, evaluate them against participants’ held-out responses, and examine how data relevance, reasoning similarity, and responses to similar tasks relate to prediction error or twin calibration.

2.1. Design and participants

  • Final samples were 139 AI-moderated, 154 static, and 24 human-moderated; because human interviews were costly, conditions were market-cost-matched, and unexpectedly high no-show rates raised that condition’s marginal cost.Source: §2.1 · pp. 6–7

2.2. Interview structure

  • Across four open-ended blocks, static interviews used fixed questions without probing, AI probed toward objectives, and humans used guides with discretion; validation stimuli were held out for twin evaluation.The validation stimuli were identical across conditions and were excluded from twin construction.Source: §2.2 · pp. 7–8

2.3. Interview process and data-quality measures

  • Measures used interviewee-only speech from the category- and brand-specific open-ended blocks, transcribed with Whisper timestamps; major statistical comparisons used Welch’s t-tests unless otherwise specified.Source: §2.3 · p. 9
  • Interview volume and interactivity were measured using participant speaking time, word count, and conversation turns across the four open-ended blocks.Source: §2.3 · p. 9
  • Interview quality was measured through codebook-based theme breadth and depth, speech emotion scores, and prosodic features of participants’ audio.Five independent coding agents were aggregated by majority vote; prosodic features included voiced fraction, speaking rate, and pause fraction.Source: §2.3 · p. 10

2.4. Customer needs as managerial insight

  • 252 unique needs were identified by extracting needs from four open-ended interview blocks and clustering embedded mentions into need categories.A customer need was defined as an underlying benefit articulated at a decision-relevant level and grounded in the respondent’s own words.Source: §2.4 · pp. 10–11

2.5. Twin construction

  • Twins were prompted with participant training data to predict held-out marketing-stimulus responses; Full personas added interview transcripts and individual-differences scales to demographics-only personas.Gemini was used for multimodal stimuli, with a GPT-5-mini replication on claims and mailer tasks.Source: §2.5 · p. 11

2.6. Twin evaluation and metrics

  • Twins were evaluated on six held-out marketing stimuli using each participant’s own response as ground truth, with response-format-specific metrics pooled to participant level.The study used paired t-tests for Demo-versus-Full differences and Welch two-sample t-tests for Static-versus-AI differences.Source: §2.6 · p. 11

2.7. Understanding prediction error: training data relevance vs. reasoning fidelity

  • Training-data relevance and reasoning fidelity were assessed separately, but may overlap because gaps in reported reasoning can reflect training data that is not fully relevant to the task.An LLM rated whether human reasoning about held-out choices traced to interview statements and whether twin explanations matched the human’s logic and priorities, both on 0–4 scales.Source: §2.7 · p. 12

2.8. Similar-task calibration ladder

  • The calibration test added a participant’s choice and open-ended response to another marketing stimulus of the same type as training data for the target stimulus.For example, responses to one commercial could be used to predict responses to the other commercial.Source: §2.8 · p. 13

3. Results: AI-Moderated Interviews as a Source of Consumer Insight

AI moderation increased interview volume and theme coverage relative to static interviews, while matching human moderation on some qualitative measures but not on the emotional experience of speaking to a person. Under a fixed budget, it recovered the most unique customer needs.

3.1.1. Interview volume

  • Relative to static interviews, AI moderation elicited nearly twice the speaking time (8.1 vs. 4.3 minutes), more words (952 vs. 609), and more turns (30.5 vs. 8.3).Human moderation produced longer speech than AI, but AI and human participants had statistically similar numbers of turns (30.5 vs. 32.4, p = .238). Enjoyment and comfort were statistically similar between AI and static conditions.Source: §3.1.1 · p. 13

3.1.2. Interview Breadth and Depth

  • Compared with static interviews, AI moderation covered more themes in three of four blocks and produced greater mean depth in purchase journey and brand perception and satisfaction.The short category-usage block, limited to three turns, showed no difference in breadth or depth; length-controlled advantages persisted for breadth in three blocks and depth in two blocks.Source: §3.1.2 · p. 14
  • Compared with human moderation, AI covered more themes in three of four blocks; depth was comparable in two, higher for AI in one, and higher for humans in the short block.AI’s theme counts were 7.29 vs. 6.08 for purchase journey, 3.32 vs. 1.79 for brand awareness, and 5.88 vs. 5.12 for brand perception and satisfaction.Source: §3.1.2 · p. 15

3.1.3. Speech-based affect: do participants sound different when speaking to a human vs. an AI?

  • Participants speaking to a human sounded more positive, emotionally activated, and assertive than those speaking to AI, while AI and static interviews were indistinguishable on affect dimensions.AI moderation also produced more pausing than human moderation; voiced fraction and speaking rate did not differ significantly across conditions.Source: §3.1.3 · p. 15

3.2.1. Coverage efficiency

  • At 24 interviews, AI moderation recovered 111.5 unique needs on average, close to human moderation’s 109.0 and above static’s 75.9.AI moderation achieved this matched-sample coverage at roughly 20% of the cost per human interview.Source: §3.2.1 · p. 16
  • 238 unique needs were recovered by AI moderation under a fixed budget, compared with 188 for static and 109 for human moderation.The implied cost per unique need was $19 for AI, $27 for static, and $36 for human moderation.Source: Table 9 · p. 17, §3.2.1 · p. 17

4. Results: AI-Moderated Interviews as Training Data for Digital Twins

This section examines how interview-derived training data and reasoning fidelity relate to digital-twin prediction of held-out responses. It also considers whether adding task-related reasoning or prior responses to training improves prediction.

4.1. Demo vs. Full; Static vs. AI-Moderated

  • Full personas improved on demographics-only personas across all four pooled measures for Static interviews, while AI-moderated Full personas improved on τ-b and Cosine, with marginal Accuracy and non-significant claim-ranking gains; AI did not outperform Static overall.Static was slightly higher on Cosine for both persona types; other AI-versus-Static differences and condition-by-persona interactions were non-significant.Source: §4.1 · p. 18, Table 10 · p. 19

4.2. Understanding prediction error: training data relevance vs. reasoning fidelity

  • Reasoning fidelity positively predicted performance on Claims (b = 0.338, p< .001), Mailers (b = 0.093, p< .001), and Commercials (b = 0.080, p< .001); training-data relevance was significant for Claims and Mailers, but not Commercials.This suggests prediction errors relate both to whether relevant information appears in training data and to whether the twin’s reasoning resembles the human’s.Source: §4.2 · p. 19, Table 11 · p. 20
  • Reasoning fidelity and reasoning-style gap were only weakly correlated (r= −.23, −.15, and −.25 for Claims, Mailers, and Commercials); the gap predicted lower Claims accuracy, but not Mailers or Commercials.Thus, the tendency to reason more analytically accounts for only part of the divergence between twin and human reasoning.Source: §4.2 · pp. 20–21
  • Twins were more analytical than humans for Mailers (2.18 vs. 1.31), Commercials (2.23 vs. 1.60), and Claims (2.75 vs. 2.50), with p< .001 for each comparison.Separate response-focus coding found humans more likely to react to what they directly saw or heard, while twins more often drew on interview context.Source: §4.2 · p. 20, Table 12 · p. 21

4.3. Can adding a task similar to the holdout task to the training data improve predictions?

  • Adding 1Val to Full-persona training improved prediction only modestly, with rating accuracy for mailers and commercials rising from .743 to .787 (p< .001), while rank agreement changed little.Source: §4.3 · pp. 21–22
  • One observed response may calibrate participants’ rating-scale tendencies across similar stimuli, whereas rankings depend on feature tradeoffs that are difficult to infer from a single response.Source: §4.3 · p. 22
  • Even with ground-truth reasoning, multimodal-task performance reached Accuracyall = .815 and τ-ball = .376, leaving individual choices difficult to recover.Adding open-ended human reasoning about the actual validation task improved performance, but gains for mailers and commercials remained limited.Source: §4.3 · p. 22
  • The 1Val condition adds reasoning for another stimulus in the same task family, while 2Val also uses holdout-stimulus reasoning and is an upper bound, not an out-of-sample condition.Source: §4.3 · p. 22

5. Discussion & Conclusion

This section describes the survey and interview procedures used to compare AI, human, and static data collection, including shared moderation objectives and marketing-stimulus validation. It discusses the findings on qualitative research, digital-twin predictions, emotional engagement, and the study’s limits and possible calibration improvements.

5. Discussion & Conclusion

  • At matched interview counts, AI moderation recovered an equivalent number of customer needs to human moderation at roughly 20% of the per-interview cost, and it recovered the most needs under the same market-cost budget.Compared with static interviews, AI moderation also elicited more speech, broader theme coverage, and deeper elaboration without reducing reported enjoyment or comfort; it matched or exceeded human-interview depth in all but the short opening block.Source: §5 · pp. 22–23
  • Participants sounded more positive, emotionally activated, and assertive with a live human, while AI-moderated and static interviews were emotionally indistinguishable.The discussion suggests human moderation may matter most when rapport, affect, or sensitive disclosure is central.Source: §5 · p. 23
  • Twin prediction errors reflect not only training information but also how twins use it: for multimodal stimuli, humans give more intuitive System 1 responses while twins give more analytical System 2 explanations.The discussion cautions that longer, richer interviews alone may not yield better predictions; closely related calibration tasks may help, but gains are modest, especially for multimodal ratings.Source: §5 · p. 23
  • For customer-need discovery, AI moderation is a scalable, cost-efficient substitute for human moderation, but predictive accuracy may require calibration data from tasks closely related to the target task.The discussion suggests using responses to similar ads, mailers, or claims rather than relying only on a rich interview transcript, while noting that twins may still reason differently from humans.Source: §5 · p. 23
  • The study focuses on one relatively high-involvement window-and-door replacement category, so AI moderation may perform differently for low-involvement, symbolic, affect-laden, or sensitive topics.The authors call for systematic domain comparisons to identify when AI moderation is effective and when human rapport remains essential.Source: §5 · p. 24

Appendix A: Interview Content and Moderation Protocol

  • Appendix A describes three conditions with identical interview content but different moderation: static responses receive no probing, while AI and human moderators probe toward shared objectives.The appendix says the study was preregistered and participants were randomly assigned; responses were spoken in all conditions.Source: Appendix A · p. 29

1. Age. “Which of the following best describes your age? Select one option.”

  • The age item includes “Prefer not to answer” as a response option.Source: §1 · p. 30

2. Homeownership. “Which of the following best describes your current living situation? Select one option.”

  • Homeownership response options include owning outright or with a mortgage, renting, and living with family or friends without owning the home.Source: §2 · p. 30

4. Household income. “What is your total annual household income before taxes? Select one option.”

  • The approximately two-minute demographic intake included household and project screening, including decision-making role and comfort recording audio.The instrument also asked about recent or planned home-improvement projects and requested a brief project description.Source: §4 · pp. 30–31, §2 · p. 30

5. Household composition. “Which of the following people live in your household? Select all that apply.”

Section 6. Survey experience (AI-moderated and static conditions only)

Appendix B: Example Moderated Exchanges for the Purchase Journey Block

  • The purchase-journey examples show AI and human moderators addressing the same objective; the selected exchanges covered eight and nine of the nine journey themes.The excerpts are verbatim with light trimming and come from different participants, one per condition.Source: Appendix B · pp. 36–37, Table 16 · p. 37

Appendix C: Example of LLM Coding for Breadth and Depth

Appendix C describes fixed, objective-derived theme coding that measures interview breadth and depth; Appendix D describes a separate extraction of customer needs from respondent text.

Appendix C: Example of LLM Coding for Breadth and Depth

  • The depth rubric ranges from a bare mention (1) to rich, multi-part elaboration (4), with intermediate scores reflecting increasing detail and explanation.A score of 0 means the theme is absent.Source: Appendix C · p. 38, Coder and reviewer prompts · p. 39
  • Breadth counts themes present, while mean depth averages their 0–4 elaboration scores; themes are coded against a fixed codebook derived from each block’s objective.The codebook is specific to each open-ended block, and responses are coded independently.Source: Appendix C · pp. 37–38
  • Theme-level agreement across blocks was 98.5–99.9% after five coders’ judgments were adjudicated by a reviewer using majority rule.A deterministic majority computation served as a cross-check.Source: Appendix C · p. 37, Coder and reviewer prompts · p. 39

Appendix D: Customer-need extraction prompt

  • Customer needs were extracted in one gpt-4o pass from each respondent’s four open-ended answers, with instructions to return only needs grounded in what they expressed.A customer need is framed as an abstract job to be done, rather than a feature, solution, or opinion.Source: Appendix D · p. 40

Appendix E: Customer needs elicited in each condition

Appendix E presents the most common customer needs across the three interview conditions and needs most frequently unique to each condition.

Appendix E: Customer needs elicited in each condition

  • The table’s frequency column counts mentions across the full sample.Source: Appendix E · p. 40, Table 19 · p. 40
  • Table 19 lists the ten most frequently mentioned needs across all conditions, followed by the most frequently mentioned needs unique to each condition.The table uses a master-codebook exemplar.Source: Table 19 · p. 40, Appendix E · p. 40
  • A full list of customer needs is available to qualified academic researchers upon request for research purposes.Source: Appendix E · p. 40

Appendix F: Word Count Breakdown by Open-Ended Block

Appendix F reports interviewee-only word counts by open-ended block, and Appendix G describes regressions controlling for response length.

Appendix F: Word Count Breakdown by Open-Ended Block

  • AI-moderated interviews produced significantly more text in every block except category usage, context and feelings; human-moderated interviews produced more in two blocks.The human condition produced more text in category usage, context and feelings and brand awareness, and statistically similar amounts in purchase journey and brand perception and satisfaction.Source: Table 20 · p. 41, Appendix F · p. 41
  • Table 20 counts interviewee-only spoken words, excluding validation questions and closed-ended items, with its all-open-ended total aligned to Table 6.Block counts correspond to the per-block word counts in Table 7.Source: Table 20 · p. 41, Appendix F · p. 41

Appendix G: Length-Controlled Interview-Quality Regressions

  • Appendix G regresses breadth on the AI indicator and block word count, and mean depth per theme on the AI indicator and words per covered theme.Table 21 reports the AI-moderated coefficient relative to a static baseline.Source: Table 21 · p. 41, Appendix G · p. 41

Appendix H: Pairwise LLM-as-a-Judge Robustness Check

Appendix H describes a blinded LLM-judge comparison of matched AI-moderated and static responses. The judge preferred AI responses across all four dimensions in every block.

Appendix H: Pairwise LLM-as-a-Judge Robustness Check

  • The robustness check sampled 1,000 matched pairs per block and instructed the judge not to reward verbosity unless added content improved answer quality.The comparison covered four open-ended blocks.Source: Appendix H · pp. 41–42
  • The judge compared matched responses on breadth, depth, richness, and informativeness, and the presentation order was randomized to blind it to condition.Pairs were matched on consumer status or prior-purchaser status, depending on the block.Source: Appendix H · pp. 41–42
  • AI-moderated responses were preferred on breadth, depth, richness, and informativeness in every block (all p<.001), with the largest margins in purchase journey and brand perception and satisfaction.The smallest margin was in category usage, context and feelings, where volume-based measures showed no difference.Source: Appendix H · pp. 41–42, Table 22 · p. 42

Appendix I: Bootstrap-matched customer-need rate

The appendix covers supplementary analyses of customer-need rates at a matched sample size and prediction prompts and persona comparisons across modalities.

Appendix I: Bootstrap-matched customer-need rate

  • AI moderation exceeded static in >.99 of draws on both measures, while its exceedance rates over human moderation were 77% for needs per interview and 61% for total needs.The bootstrap used 2,000 iterations, drawing 24 AI and 24 static interviews per iteration and using all 24 human interviews.Source: Appendix I · pp. 42–43, Table 23 · p. 43

Appendix J: Digital-Twin Construction and Prediction Prompts

  • Digital twins received a shared system prompt and either a demographics-only profile or a fuller profile with category and brand context, individual-difference scales, and raw response logs.The validation section was excluded from the profile data.Source: Persona system prompt · p. 43

3. Due to vinyl warp, your vinyl windows won’t latch over time (security risk)

  • The reproduced claim-ranking prompts asked respondents to rank window-benefit claims by attractiveness and vinyl-window statements by concern.The tasks included benefit claims about durability, drafts, and home value, as well as concerns about vinyl-window performance and fading.Source: Task 1: Claim ranking — Set A (benefit… · p. 44, Task 1: Claim ranking — Set B… · p. 44
  • The ranking prompts predict a related task using prior rankings, prior rankings plus explanations, or reasoning about the target claim set.The templates distinguish 1Val, which uses prior ranking and explanation, from 2Val, which uses reasoning about the target set itself.Source: Ladder-condition inserts · p. 46

Appendix K: Individual Measures by Modality and Persona Type

  • Appendix K tests paired Full-versus-Demo differences within Audio and AI and tests whether those differences vary by modality.It identifies Table 24 as pre-registered measures, Table 25 as Kendall τ-b alternatives, and Table 26 as standardized composites; the supplied passages do not give outcome values.Source: Appendix K · pp. 46–47, Table 26 · p. 47

Appendix L: Robustness to Twin Model Choice

A GPT-5-mini replication largely reproduces the main digital-twin patterns on claims and mailers; commercials are not replicated because the model does not accept video inputs.

Appendix L: Robustness to Twin Model Choice

  • The replication uses GPT-5-mini and covers claims and mailers, but not commercials, because the model accepts text and image inputs but not video.Table 27’s measures are limited to claims and mailers, so its levels are not directly comparable to Table 10’s values.Source: Appendix L · pp. 47–48, Table 27 · p. 48
  • Full personas improve over Demo personas on Accuracyall and claim-ranking τall; AI and static moderation are indistinguishable on closed-ended measures, with Cosineall the sole AI–static difference.The reported patterns include no significant condition-by-persona interaction.Source: Appendix L · p. 47, Table 27 · p. 48
  • The Full-over-Demo change in Cosineall is small and negative in this replication, significantly so for the AI condition.The pattern is unchanged when open responses are re-embedded with all-MiniLM-L6-v2.Source: Appendix L · pp. 47–48
  • Twins are significantly more System-2 than their own humans for both tasks and both models.Table 28 compares each twin with its own human using independent scoring among AI-moderated participants.Source: Appendix L · p. 48, Table 28 · p. 48

Appendix M: Coding Examples for Training Data Relevance and Reasoning Fidelity

Appendix M illustrates the coding of twin prediction errors, while Appendix N supplies the coding prompts, scoring rubrics, and analysis sample boundaries.

Appendix M: Coding Examples for Training Data Relevance and Reasoning Fidelity

  • Examples juxtapose the participant’s interview, held-out response, twin response, and an LLM judge’s rationales and scores.The examples are Full-persona claim rankings from the AI-moderated condition.Source: Appendix M · p. 49
  • The examples include high and low training-data relevance and reasoning fidelity, plus a mixed case where the twin re-prioritizes claims.The mixed case retains high training-data relevance despite lower reasoning fidelity.Source: Appendix M · p. 49, Table 29 · p. 49, Table 30 · p. 49

Appendix N: Coding Prompts for Training Data Relevance, Reasoning Fidelity, and Response Style

  • Open-ended coding excludes three participant–stimulus cells, leaving samples of 276 claims, 277 mailers, and 278 commercials; closed-ended accuracy analyses retain those cells.The exclusions concern missing open-ended explanations, not the corresponding closed-ended responses.Source: Appendix N · p. 50
  • Appendix N’s rubric separately scores whether human validation reasoning traces to interview data and whether twin reasoning resembles the human’s, each from 0 to 4.The prompt defines traceability and similarity through ordered anchors, from no tracing or entirely different reasoning to direct tracing or essentially identical reasoning.Source: Training data relevance and reasoning… · p. 51
  • The style prompt scores each response from System 1 to System 2, while reasoning points are classified into six categories later collapsed to four.For the style scale, 0 is pure intuitive/reactive response and 4 is fully deliberative and systematic.Source: System 1 vs. System 2 response style · p. 51, Response-focus taxonomy · p. 51

Appendix O: Response-Focus Coding of Human and Twin Rationales

Human and twin open-ended rationales differ in their focus: humans more often react to the stimulus, while twins more often draw on person-level interview context.

Appendix O: Response-Focus Coding of Human and Twin Rationales

  • The coding distinguishes sensory reactions, evaluative emotions, product features or claims, and person-level memories, experience, values, or social proof.The reported four categories derive from a six-category taxonomy.Source: Appendix O · p. 52
  • Humans are more likely to explain responses through immediate sensory reactions, whereas twins more often retrieve person-level context from interview data.Table 34 reports the response-focus category distributions, with Δ defined as Twin minus Human.Source: Appendix O · p. 52, Table 34 · p. 52

Appendix P: Alternative Measure of the Reasoning Difference

The appendix tests the reasoning-style gap as an alternative measure of differences between human and twin reasoning. A larger gap is associated with lower prediction accuracy for Claims, while the measure captures only part of the reasoning differences.

Appendix P: Alternative Measure of the Reasoning Difference

  • The reasoning-style gap and reasoning fidelity both measure differences between a twin’s and a human’s open-ended reasoning about the same stimulus.Source: Appendix P · p. 52
  • Reasoning-style gap and reasoning fidelity are weakly correlated (r = −.23, −.15, and −.25 for Claims, Mailers, and Commercials), so other reasoning differences remain unaddressed.Source: Appendix P · p. 52
  • Training-data relevance and reasoning fidelity are substantially correlated (r = .56, .50, and .61 for Claims, Mailers, and Commercials), so relevance partly absorbs fidelity’s association in this specification.The specification replaces reasoning fidelity with the reasoning-style gap.Source: Appendix P · pp. 52–53
  • For Claims, a larger reasoning-style gap is significantly associated with lower prediction accuracy; coefficients are small and nonsignificant for Mailers and Commercials.The regression concerns Full personas in the AI-moderated condition.Source: Appendix P · p. 52, Table 35 · p. 53

Appendix Q: Per-Task Condition Ladder

Appendix Q reports condition-ladder means separately for each stimulus measure, across the listed persona conditions for AI-moderated participants.

Appendix Q: Per-Task Condition Ladder

Loading 2609.29143v1…