PaperlayerView exact revision on arXiv

Source-linked AI summary

The As-If Machine: Personalized counterfactuals make distant policy risks feel close

Joshua Ashkinaze, Yubo Zhou, Pengyu Chen, Avi Bhattacharya, Hyebin Park, Yanman Zhang, Stephanie Preston, Eric Gilbert, Ceren Budak

arXiv:2610.33301v1cs.HCcs.AI

TL;DR

People cannot personally experience every consequence relevant to policy, and personalized stories are difficult to scale. The paper introduces AIM, a system for co-writing personalized counterfactual narratives, and finds that it reliably increased psychological closeness across two policy domains while attitude effects varied by domain.

  • Problem

    Because important risks may never be personally experienced, the paper asks whether AI-assisted simulation can make unfamiliar futures feel concrete and personally relevant.

    Source: §1 · pp. 1–2

  • Method

    AIM is a multi-agent co-writing system that scaffolds participants’ own narratives about possible policy futures.

    Source: §9 · p. 13, §10.4 · p. 16

  • Results

    AIM increased psychological closeness by 0.36 and 0.38 SDs against the vignette and 0.80 and 0.90 SDs against unrelated writing, while policy-attitude effects differed by domain.

    Source: §9 · p. 12

  • Takeaways & Limitations

    Generative AI may approximate the concrete sense of what alternative futures would mean for an individual’s own life.

    Source: §9 · p. 14

    The serial mediation decomposition is descriptive, not evidence that participants experienced closeness before transportation.

    Source: Serial mediation model · p. 50

Abstract

from arXiv · show

Life experiences can shape policy attitudes, but we only live one life. Citizens may therefore fail to support policies addressing risks they have never personally experienced. Narratives can make such risks immediate and personal, but personalized stories remain difficult to produce at scale. Here we ask whether AI can make distant scenarios feel concrete by simulating a participant's life as if that scenario occurred. We present the As-If Machine (AIM), a multi-agent system through which people co-write personalized counterfactual narratives---stories about their own lives under different policy futures. In two pre-registered and quota-matched experiments on science funding (N=1,007) and satellite cybersecurity (N=894), we compared AIM with two controls: reading an expert-written vignette on the same topic and co-writing with AIM about an unrelated topic. AIM increased psychological closeness---how close a risk feels to the self---by nearly identical margins in the two domains (0.36 and 0.38 standard deviations (SDs) against the vignette, 0.80 and 0.90 against unrelated writing). AIM increased psychological closeness even among participants with unfavorable views of AI. For policy attitudes, AIM's effects were 0.16 SDs against the vignette and 0.10 SDs (n.s.) against unrelated writing for science funding; the corresponding effects for satellite cybersecurity were 0.54 and 0.67 SDs. Across 23,496 human-AI autocomplete interactions, participants and the system played distinct roles: AIM supplied concrete detail that grounded the scenario, and participants supplied first-person perspective and emotion. Co-writing with the As-If Machine makes distant policy risks feel close.

1 Introduction

People cannot personally experience every consequence relevant to policy, and familiar interventions face a tradeoff between scalable delivery and tailored stories. AIM uses personalized counterfactual co-writing to test whether AI can make unfamiliar risks feel concrete and when that closeness changes policy attitudes.

1 Introduction

  • Unexperienced risks can remain psychologically distant and abstract, while personal experience can make consequences feel concrete and personally relevant.Many important risks cannot be experienced in advance, motivating the question of whether AI-assisted simulation can approximate what experience provides.Source: §1 · p. 2
  • AIM is a multi-agent system for co-writing personalized counterfactual narratives about participants’ lives under possible policy futures.It generates a personalized event onset and consequences, while an LLM offers continuations that participants may accept or reject.Source: §1 · p. 2
  • The introduction separates whether personalized simulation makes unfamiliar consequences feel personally relevant from whether closeness changes policy attitudes.Prior attempts to reduce psychological distance have produced inconsistent effects on attitudes and behavior.Source: §1 · p. 2
  • In two pre-registered experiments, AIM was compared with an expert-written narrative on the same issue and interactive writing about an unrelated topic.The studies measured psychological closeness and policy attitudes and examined effects by prior familiarity.Source: §1 · p. 2
  • AIM increased psychological closeness in both domains, while policy-attitude effects differed between science funding and satellite cybersecurity.Across 23,496 autocomplete interactions, AIM supplied concrete detail and participants supplied first-person perspective and emotion.Source: §1 · pp. 2–3

2 System description: An open-source engine of interventions

AIM separates scenario planning from live writing assistance through four specialized language-model agents. Its configuration-based design lets researchers vary interventions while retaining a shared architecture and workflow.

2 System description: An open-source engine of interventions

  • Four specialized agents generate a scenario onset, forecast consequences, connect them into a plot, and offer short writing suggestions.The Autocompleter maintains continuity, while participants can accept, reject, or modify its suggestions.Source: §2 · p. 3
  • The agent stages pass reusable outputs forward, from the personalized onset and predicted consequences to the plot and writing suggestions.Informal testing led the team to constrain the Autocompleter from attributing thoughts or emotions to participants.Source: §2 · p. 3
  • A plain-text YAML configuration specifies the scenario, personalization questions, deployment parameters, models, and agent instructions for each experiment.AIM uses the configuration to render participant-facing pages and create a dedicated experiment URL that can be distributed directly or incorporated into study workflows.Source: §2 · p. 4

3 Study 1: Science funding

In the science-funding experiment, AIM increased psychological closeness relative to both controls, while its policy-attitude effect was small and not significant against the placebo. Mediation estimates were positive but non-causal, and effects were generally larger among participants less familiar with the issue.

3.1 Main effects

  • 0.36 SDs against the vignette and 0.80 SDs against the placebo were the psychological-closeness effects; policy-attitude effects were 0.16 SDs and 0.10 SDs, respectively, with the placebo contrast not significant.AIM participants co-wrote a personalized account of delayed care; controls read an expert vignette or used AIM to write about an unrelated topic.Source: §3.1 · p. 5, §3 · p. 5

3.2 Mediation

  • Positive indirect effects through closeness were estimated against both controls, but the analysis does not identify causal mediation because closeness was not randomized.Both indirect effects remained positive unless an unmeasured confounder induced an approximately ρ⋆= 0.40 correlation between unexplained closeness and attitude components.Source: §3.2 · pp. 5–6

3.3 Familiarity

  • Among participants with low familiarity, AIM increased closeness and attitudes relative to both controls, while effects among higher-familiarity participants were smaller or near zero.The pattern was similar but weaker using a familiarity cutoff of 3.Source: §3.3 · p. 6

3.4 Heterogeneous treatment effects

  • Only familiarity survived correction among the explored moderators; higher familiarity predicted smaller closeness effects against the placebo.Other estimated moderator relationships did not survive multiple-comparison correction.Source: §3.4 · p. 6

4 Study 2: Satellite cybersecurity

In the satellite-cybersecurity experiment, AIM increased psychological closeness and policy support relative to both controls. Mediation estimates were positive, and lower-familiarity participants had larger estimated effects, although preregistered moderation evidence was limited.

4 Study 2: Satellite cybersecurity

  • The experiment randomized 894 participants to AIM, an expert vignette, or an unrelated-topic writing placebo.The AIM scenario concerned a satellite cyberattack disrupting everyday life, and the study used revised attitude measures spanning three subscales.Source: §4 · pp. 6–7

4.1 Main effects

  • 0.38 SDs against the vignette and 0.90 SDs against the placebo were the psychological-closeness effects; policy-attitude effects were 0.54 SDs and 0.67 SDs, respectively.Attitude effects were strong across the three subscales of importance, spending, and willingness to oppose funding cuts.Source: §4.1 · p. 7

4.2 Mediation

  • Positive indirect effects through closeness were estimated against both controls, with sensitivity thresholds of ρ⋆= 0.56 and 0.62.The estimates were b = 0.24 against the vignette and b = 0.54 against the placebo.Source: §4.2 · p. 7

4.3 Familiarity

  • All eight subgroup contrasts had confidence intervals above zero, and lower-familiarity participants had larger estimated effects across all four outcome-control combinations.At p < 0.05, preregistered moderation tests supported larger effects only in the placebo comparisons.Source: §4.3 · p. 7

4.4 Heterogeneous treatment effects

  • Only familiarity survived multiple-comparison correction among the Study 2 moderator findings; higher familiarity predicted smaller closeness effects against the placebo.No moderator of attitude effects reached p < 0.05.Source: §4.4 · p. 8

5 Exploratory distributional effects

Quantile analyses suggest AIM’s closeness effects were concentrated among participants who initially felt the issue had little to do with their lives. Figure 3 displays estimated treatment effects across outcome quantiles, controls, and studies.

5 Exploratory distributional effects

  • Quantile regression models how treatment shifts different quantiles of an outcome distribution, complementing OLS estimates of mean shifts.Source: §5 · p. 8
  • AIM’s effects on closeness were generally concentrated in the lower quantiles, suggesting fewer treated participants felt the issue had very little to do with their lives.The analysis explores treatment shifts across outcome quantiles rather than only the mean.Source: §5 · p. 8
  • Figure 3 plots estimated treatment effects across outcome quantiles 0.1–0.9, with pointwise 95% intervals; rows distinguish controls and colors distinguish studies.Source: Figure 3 · p. 8

6 Suggestive mechanisms and non-mechanisms

Suggestive analyses indicate that closeness, more than transportation, carried AIM’s indirect association with attitudes. Other participant experiences and system ratings were associated with outcomes, while AI adoption and prior AI views showed little evidence of moderation.

6 Suggestive mechanisms and non-mechanisms

  • In the pooled AIM-versus-vignette model, closeness alone carried an indirect effect of b = 0.15, and the paths mediated 59% of AIM’s effect on attitudes.The path through closeness and then transportation was b = 0.05; transportation alone was near zero (b = 0.00).Source: §6 · p. 10
  • Closeness predicted transportation and attitudes, while AIM had no detectable effect on transportation beyond its effect through closeness.The pooled model found b = 0.37 for AIM’s effect on closeness and b = 0.61 for closeness predicting transportation.Source: §6 · p. 10
  • Among AIM participants, greater transportation and more positive ratings of enjoyment, tailoring, realism, helpfulness, and ease were associated with greater closeness and more supportive attitudes.These analyses are correlational and do not establish that positive system experiences caused the outcomes.Source: §6 · p. 10
  • AI-text acceptance was unrelated to closeness or attitudes, and prior AI views did not significantly moderate AIM’s effects on either outcome.All eight AI-attitude moderation coefficients were small and nonsignificant, ranging from −0.14 to 0.09 SD per SD of AI attitudes.Source: §6 · p. 10
  • Perceived agency was unrelated to attitudes and only weakly related to closeness, the smallest association among system ratings.Participants nevertheless rated AIM’s ease, fun, tailoring, realism, and helpfulness above the scale midpoint.Source: §6 · p. 10

7 Qualitative feedback reveals success modes and failure modes

Participant feedback highlighted personal relevance and recognition of everyday dependencies as successes, alongside concerns about unsuitable suggestions, implausible details, and attention shifting from policy to personal predicaments.

7 Qualitative feedback reveals success modes and failure modes

  • Personal relevance was a leading success theme: it accounted for 30.1% of Study 1 success responses, while recognizing disruption’s scale led Study 2 at 31.4%.Source: §7 · p. 10
  • Recognizing everyday satellite dependencies was a major success theme, appearing in 34.7% of Study 2 success responses.About a third of those responses concerned reliance on card and phone payments rather than cash.Source: §7 · pp. 10–11
  • Participants valued suggestions that continued a narrative as intended, but the predictive-ability theme appeared in 6.7% of Study 1 success responses.Source: §7 · p. 11
  • In Study 1, some participants felt AI suggestions were too repetitive because they restated the prompt and scenario; the autocomplete prompt was revised for Study 2 to reduce repetition.Source: §7 · p. 11
  • Implausibility was the largest Study 2 failure theme, accounting for 37.2% of substantive negative responses.Participants steering their own stories meant a detail plausible in one narrative could seem implausible in another.Source: §7 · p. 11
  • Some participants said personal predicaments drew attention away from policy, reported in 12.8% of Study 1 and 16.7% of Study 2 failure responses.Study 2 comments included stories of disruption that did not explain how an attack caused it or how cybersecurity policy could prevent it.Source: §7 · p. 11

8 Consistent yet distinct roles of humans and AI

Across both studies, AI and participants contributed distinct kinds of narrative content: AI supplied scenario structure and concrete detail, while participants supplied first-person perspective and emotion. This division replicated across domains.

8 Consistent yet distinct roles of humans and AI

  • Participants received 23,496 AI suggestions and accepted 4,068; linguistic features distinguished human from AI writing with mean AUCs of .914 and .940 across the two studies.The AUCs were .914 in science funding and .940 in satellite cybersecurity.Source: §8 · p. 11
  • AI supplied scenario structure and physical detail, while participants contributed emotional and interior perspective across both domains.AI’s distinctive terms named the science-funding scenario or established the satellite setting; human terms included “feel,” “think,” and “worried.”Source: §8 · pp. 11–12
  • The system supplied scenario description, while participants supplied first-person perspective and emotion.Humans used more first-person pronouns (dz = 0.55 in science funding and 0.72 in satellite cybersecurity); AI used more nouns, participial clauses, and longer words.Source: §8 · p. 12
  • The human–AI feature division replicated across domains: feature contrasts correlated r = .78 between studies, with 83 of 109 in the same direction.The division was intentional: the Autocompleter was designed to avoid being too sentimental, didactic, or prescriptive.Source: §8 · p. 12

9 Discussion

Across two experiments, AIM made distant risks feel closer, but whether closeness changed policy attitudes depended on the domain. The discussion connects this pattern to hidden dependencies and distinct human–AI roles, while identifying limits around specificity, political framing, and causal interpretation.

Results

  • AIM increased psychological closeness by 0.36 and 0.38 SDs versus an expert-written vignette and 0.80 and 0.90 SDs versus unrelated writing; attitude effects varied by domain.For science funding, attitude effects were 0.16 SDs versus the vignette and 0.10 SDs versus placebo (not significant); for satellite cybersecurity, they were 0.54 and 0.67 SDs, respectively.Source: §9 · p. 12

Hidden dependencies

  • In the satellite study, 34.7% of success-theme responses described discovering everyday dependencies or previously unnoticed vulnerability.Participants identified connections among satellite infrastructure, payments, navigation, and communication.Source: §9 · p. 12

Human–AI roles

  • Across domains, AIM supplied scenario structure and physical detail, while participants supplied first-person perspective and emotion.The 109 human–AI linguistic contrasts correlated r = .78 between studies; AIM scaffolded stories participants developed themselves.Source: §9 · pp. 12–13

Conditions and interpretation

  • AIM effects held among participants with unfavorable AI views, but the larger satellite attitude effect may reflect less crystallized attitudes; crystallization was not manipulated.AI attitudes did not moderate either outcome; other study differences could explain the attitude-effect divergence.Source: §9 · p. 13

Specificity

  • Specific details can make counterfactuals concrete but also give participants claims to contest when they seem implausible.The fixed 14-month wait in Study 1 helped some participants imagine consequences but was rejected by others; participants could also steer stories into contexts where details no longer fit.Source: §9 · pp. 13–14

Personal and political framing

  • AIM’s episodic stories sometimes left institutions and policy causes abstract, and 13% of Study 1 failure-theme responses focused on personal predicaments instead of science funding policy.The discussion argues scenarios should connect personal consequences to politically credible and typical causes.Source: §9 · p. 14

Misuse potential

  • Active writing and suggestion choices may reduce manipulation risk, but deployers still control the scenario and how strongly agents steer.A deployer could omit scenario disclosure; active engagement therefore reduces but does not remove manipulation risk.Source: §9 · p. 14

Limitations

  • The studies measured self-reported outcomes immediately, did not isolate AIM components, and tested only negative futures with Prolific samples potentially more comfortable with AI writing.Closeness and attitudes were measured together, so results are consistent with—but do not identify—an indirect effect through closeness.Source: §9 · p. 14

10 Methods

The methods describe two preregistered randomized experiments, AIM’s implementation, outcome measures, and analyses. They also report formative design choices and procedures for qualitative and linguistic comparisons.

10.1 Study design and preregistration

  • The experiments randomized participants to AIM, an expert vignette, or an unrelated-topic writing placebo, with quota sampling and preregistered hypotheses and analyses.Participants completed consent and screening, then read the vignette or co-wrote before answering outcome and experience questions.Source: §10.1 · p. 15, §10.2 · p. 15

10.3 System implementation

  • AIM’s current Autocompleter called GPT-4.1-mini after a writing pause and offered short completions; formative testing led to a personalized starting prompt and constraints against attributing feelings to users.The prompt gave participants a plausible entry point, and participants rated it helpful for starting their stories.Source: §10.3 · p. 15, §10.4 · pp. 15–16

10.5 Outcome measures

  • Closeness and attitude indices combined standardized study-specific items; satellite attitudes equally weighted importance, spending, and action subscales.Pretreatment measures included topic familiarity, writing experience, imagery, interpersonal tendencies, and AI attitudes; post-task measures included usability, enjoyment, story quality, agency, and transportation.Source: §10.5 · p. 16

10.6 Pretreatment and posttreatment measures

  • Participants also reported system experience and narrative transportation after treatment, while the preregistered exclusions removed failed checks, unusually fast completions, and technical-error reports.Quota sampling targeted U.S. population distributions on sex, political party, and age.Source: §10.6 · p. 17, §10.7 · p. 17

10.8 Primary effects and familiarity

  • Primary effects came from preregistered OLS models with AIM as reference; the analyses reported AIM-minus-control contrasts and tested familiarity groups at two cutoffs.The models used HC1 robust standard errors and reported sharpened false-discovery-rate q values for the specified contrasts.Source: §10.8 · p. 17

10.9 Mediation and exploratory models

  • The preregistered mediation analysis tested whether closeness mediated attitude effects, while an exploratory serial model included transportation and study number.Exploratory adoption was summarized with an inverse-covariance weighted index of suggestion acceptance, accepted-suggestion count, and AI-origin words.Source: §10.9 · p. 18

10.10 Causal forests and treatment-effect heterogeneity

  • Exploratory causal forests estimated heterogeneous effects separately for each outcome and AIM–control comparison, with honest trees and doubly robust-score summaries.The Supplementary Information reports CATE panels, BLP estimates, and q values adjusted within each forest.Source: §10.10 · p. 18

10.11 Qualitative and linguistic analysis

  • Qualitative themes were clustered with BERTopic2, while paired human–AI text analyses compared linguistic features; accepted suggestions were removed approximately to reconstruct human-authored text.The researchers also used generative AI for code generation, writing assistance, and stimulus refinement.Source: §10.11 · pp. 18–19

11 Author contributions

The author contributions identify roles in conceiving, designing, programming, running, analyzing, and writing the research.

Author contributions

  • Joshua Ashkinaze led conception, programming, experiments, analyses, and manuscript writing; collaborators contributed to design, programming, analysis, supervision, and feedback.The listed collaborators were Yubo Zhou, Pengyu Chen, Avi Bhattacharya, Hyebin Park, Yanman Zhang, Stephanie Preston, Eric Gilbert, and Ceren Budak.Source: §11 · p. 19

12.2 Funding

The funding statement names two sources supporting both studies.

Funding

  • Both studies received funding from the Cooperative AI Foundation and the University of Michigan’s Propelling Original Data Science program.Source: §12.2 · p. 25

13 Extended Data

The extended data documents outcome and treatment-effect tables, subgroup and mediation analyses, interface examples, human–AI language comparisons, and narrative-transportation practices.

13.1 Outcome distributions and primary effects

  • Extended Data Table 1 reports outcome means by study and condition, distinguishing z-scored outcome indices from their constituent items in original units.Study 2’s attitude index averages importance, spending, and action subscales.Source: Table 1 · p. 26
  • Extended Data Table 2 reports AIM-minus-control estimates for the z-scored closeness and attitude indices.The estimates come from condition-only OLS models and are expressed in standard deviations.Source: Table 2 · p. 27

13.2 Pre-registered familiarity subgroups

  • Extended Data Tables 3 and 4 report AIM-minus-control estimates and 95% confidence intervals for familiarity splits at thresholds of 2 and 3.The tables cover the pre-registered subgroup cutoffs.Source: Table 3 · p. 28, Table 4 · p. 28

13.3 Study-specific mediation and satellite attitude subscales

  • Extended Data Table 5 presents single-mediator models by study and control, with bias-corrected bootstrap intervals for indirect effects.The mediation estimates do not identify causal mediation because closeness was not randomized.Source: Table 5 · p. 29
  • Extended Data Table 6 reports secondary satellite attitude-subscale estimates from three-arm OLS models with HC1-robust intervals and two-sided tests.Source: Table 6 · p. 29

13.4 Interface and scenario configuration

  • Extended Data Figures 1 and 2 pair an illustrative gravity-disappearance interface with its YAML configuration; that scenario was used in neither experiment.Source: Figure 1 · p. 30, Figure 2 · p. 31

13.5 Satellite interface and autocomplete prompt

  • Extended Data Figure 3 shows the satellite co-writing interface and step-five autocomplete request, with colors indicating each element’s source.Source: Figure 3 · p. 32

13.6 Human and AI lexical differences

  • Extended Data Figures 4 and 5 compare human- and AI-associated terms and linguistic features across paired narratives or documents in both studies.The lexical figure ranks terms by log-odds z score; the linguistic-feature figure uses positive dz for higher participant-written values.Source: Figure 4 · p. 33, Figure 5 · p. 34

14 Supplementary Information

The supplementary analyses document study procedures and measurement, validate attitude coding, and examine explanations for differences between studies, including ceiling effects and story–measure overlap. They also assess closeness across prior AI attitudes, mediation and moderator patterns, and sensitivity to unmeasured confounding, with some analyses qualified by concurrent measurement, exploratory designs, or passages that describe scope without numerical findings.

14.3 Screening procedure, deviations, complications

  • The authors skipped pre-registered manual classification of custom technical-error reports, applied reCAPTCHA-score exclusions only where scores were recorded, and did not pool the studies’ causal forests.They retained participants reporting predefined serious errors; the pooled forest was not used because moderator coefficients correlated weakly and inconsistently across studies (r ranged from −0.41 to 0.54).Source: §14.3 · p. 38
  • Dropping the 146 Study 2 participants who began before the timer change left the main results unchanged, with no significant effect-by-period interaction (smallest p = .28).The remaining 748 participants began after the timer was revised to remain visible while scrolling.Source: §14.3 · p. 38, Table S1 · p. 39

Closeness differences

  • Each SD increase in closeness predicted a 0.42 SD attitude difference in Study 1 and a 0.59 SD difference in Study 2 among placebo participants (interaction: b = 0.16, 95% CI [0.05, 0.28], p = .007).Placebo participants provide the cleanest contrast because they wrote about the same unrelated scenarios in both experiments.Source: Closeness differences · p. 41, Table S4 · p. 41

Media Cloud procedure

  • 493 of 8,050,883 U.S. national news stories mentioned science funding, versus 17 mentioning satellite or space cybersecurity, in the two-year search window.The corresponding rates were 0.61 and 0.02 per 10,000 stories; adding related funding phrases increased the science-funding count to 6,579, while related satellite phrases added no stories.Source: Media Cloud procedure · p. 42

Validation of the LLM coding

  • LLM–human agreement was lower on the 100 randomly sampled responses than across all 200, while raw agreement remained 93%–96%.On the random subset, κ was 0.71 with Rater A and 0.66 with Rater B; across all responses, κ was 0.91 and 0.80, respectively. The human raters’ κ was 0.80 on the random subset and 0.83 across all responses.Source: Validation of the LLM coding · pp. 44–45, Table S7 · p. 45

14.7 Alternative explanations for Study 1 and Study 2 attitude differences

  • Neither improved participant ratings nor sample composition explains Study 2’s larger attitude effects, and story–item overlap points in the opposite direction.No placebo system rating differed after correction, and adding 13 pretreatment covariates, Safari use, and story length left the effect gap essentially unchanged against both controls.Source: UX differences · p. 45, Table S8 · p. 46, Composition differences · p. 46

Measure overlap differences

  • Study 1 stories were more similar to their attitude items than Study 2 stories, but within either study, story similarity did not significantly predict attitudes.The between-study similarity difference was Welch t(591) = 3.04, p = .002.Source: Measure overlap differences · p. 47, Table S10 · p. 47

Ceiling effects

  • Study 1 ceiling effects did not explain its smaller attitude effects: Tobit estimates matched least-squares estimates for the top-heavy item and barely changed for the NIH budget item.For personal benefit of science, 27.4% of placebo and 27.6% of AIM participants were at the maximum, and both estimates were b = 0.02; for the NIH budget item, the estimate changed from b = 0.12 to b = 0.13. In Study 2, AIM increased maximum responses by 7.0–11.2 percentage points across all seven attitude items, and Tobit estimates exceeded least-squares estimates.Source: Ceiling effects · p. 47

Ceiling effects; 14.10 Qualitative themes; 14.11 Linguistic classification performance; 14.12 Satellite attitude measurement

  • Supplementary materials cover ceiling effects, qualitative themes from both studies, human–AI text classification, and satellite attitude measurement, but the supplied captions do not report their findings.The captions identify S11 as ceiling-effect analysis; S15 and S16 as open-text themes; S17 as linguistic classification; and S18 as satellite attitude measurement.Source: Table S11 · p. 48, Table S15 · p. 52, Table S17 · p. 52

Vignette difference (some support)

  • Study 1’s vignette was more emotionally affecting than Study 2’s, while AIM stories were similarly affecting; this difference explains about a fifth of the larger Study 2 vignette contrast.The account applies only to the vignette contrast, not the larger difference against placebo.Source: Vignette difference (some support) · p. 48, Table S12 · p. 48

14.8 Prior attitudes toward AI

  • Every closeness contrast held among participants on both sides of the AI-attitude midpoint, with no significant arm-by-attitude interaction.The authors caution that this shows no variation by prior AI attitude, not that AIM worked for every participant who dislikes AI.Source: §14.8 · p. 49

Serial mediation model

  • The primary serial model places closeness before transportation, but concurrent measurement makes this a descriptive ordering assumption rather than evidence of causal sequence.Model selection combined prior work with consistency with the data; the dual-mediator models report only the vignette contrast because placebo stories were unrelated to the attitude topic.Source: Serial mediation model · pp. 49–50, §14.9 · p. 49

Considering Directionality

  • The observed coefficients were inconsistent with transportation preceding closeness and consistent with closeness preceding transportation.Controlling for transportation did not reduce treatment’s effect on closeness to zero, while controlling for closeness reduced its effect on transportation to zero.Source: Considering Directionality · p. 50, Table S14 · p. 50

Parallel mediation model

  • Across studies and comparisons, closeness carried more of the treatment’s effect on attitudes than narrative transportation under both serial and parallel mediation models.The models fit the same number of parameters as observed moments, so fit statistics alone could not distinguish them.Source: Parallel mediation model · p. 50, Figure S2 · p. 51, Figure S3 · p. 51

14.13 Experiment event configurations

  • Experiment configurations added intake answers to participant character descriptions and used them to personalize the science-funding and satellite-cybersecurity story prompts.The science-funding prompt centered on a specialist appointment delay linked to lost NIH training funding; the satellite prompt centered on a cyberattack disrupting services participants rely on.Source: §14.13 · pp. 54–55

14.14 Causal-forest heterogeneity analyses

  • Causal-forest heterogeneity analyses were exploratory, used complete-case samples, and projected conditional treatment effects onto moderators with HC3 standard errors.The supplied passages describe methods and figures for Study 1 comparisons, but do not report moderator-specific findings.Source: §14.14 · p. 58, Table S19 · p. 59, Figure S4 · p. 59
  • The causal-forest analyses examine conditional effects and best linear projections for psychological closeness and policy attitudes across the science-funding and satellite-cybersecurity studies and their AIM control comparisons.The supplied captions identify Study 1 AIM versus vignette and Study 2 AIM versus placebo and AIM versus vignette analyses; they do not provide coefficient values or moderator findings.Source: Figure S10 · p. 65, Figure S11 · p. 66, Table S21 · p. 67

14.15 Consistency of moderators across studies

  • Moderator-coefficient consistency was assessed by standardizing moderators within each study and correlating coefficients across moderators for each of four forests.The comparison figure reports coefficients with 95% confidence intervals; its vignette comparison omits four app-usage moderators available only for AIM and placebo participants.Source: §14.15 · p. 75, Figure S20 · p. 75

14.16 Sensitivity of the mediation analyses

  • Mediation sensitivity analyses plot indirect effects against the mediator–outcome error correlation, with 95% intervals and the correlation value where each effect crosses zero.The figure caption states that estimates used the R package mediation and 1,000 simulations.Source: Figure S21 · p. 76, §14.16 · p. 76

Related papers

Loading 2610.33301v1…