Deepfakes by another name … are still dangerous

What and Why (N = 406)

Author

Leo Yang

Published

May 29, 2026

Ch. 1 - 12: What (RQs)

Ch. 13 - 17: Why (No formal hypotheses)

1. Sample & RQ

Sample size of \(N = 406\) participants across two sources: Kelley School of Business Main (\(N = 211\)) and CloudResearch (\(N = 195\)).

Research questions:

  1. RQ1 (Label Effect): Does substituting “Synthetic Avatar” for “Deepfake” lead to less negative assessments of AI-generated videos?

  2. RQ1a (Gender Moderation): Do these Label Effects vary significantly by respondent gender?

  3. RQ1b (Population Cohort Moderation): Do the effects vary between local student cohorts and national online panels?

2. Environment Setup & Data Import

Necessary R packages.

Code 
# Load required analysis, plotting, and formatting packages
library(readxl)
library(tidyverse)
library(psych)
library(car)
library(mediation)
select <- dplyr::select
library(emmeans)
library(ordinal)
library(patchwork)
library(knitr)
library(scales)

# Set sum-to-zero contrasts for orthogonal coding in Type III ANCOVAs
options(contrasts = c("contr.sum", "contr.poly"))
Code 
# Import sheet "HICSS- May28-DATA- RECODE"
file_path <- "HICSS- May28-DATA.xlsx"
sheet_name <- "HICSS- May28-DATA- RECODE"

if (!file.exists(file_path)) {
  stop("Error: 'HICSS- May28-DATA.xlsx' not found. Please place it in the same directory.")
}

raw_df <- read_excel(file_path, sheet = sheet_name)
clean_df <- raw_df              # Raw data has clean numeric values directly (no Qualtrics description row)

# Clean variable names
colnames(clean_df) <- trimws(colnames(clean_df))
colnames(clean_df) <- gsub(":", "", colnames(clean_df))

# Convert to numeric
numeric_cols <- c(
  "age_conti", "Gender", "RaceEthnicity", "Hispanic", "Time_in_USA",
  "AIKnowledge", "DS_DeepFakeFamiliarity", "DS_DeepfakeFeelings",
  paste0("DS_DeepfakeApplications_", 1:8),
  paste0("DS_DeepfakePurposes_", 1:5),
  "D_DeepfakePurposes_6"
)
clean_df <- clean_df %>% mutate(across(any_of(numeric_cols), as.numeric))

3. Preprocessing & Data cleaning

Restricted gender variable to “Female” and “Male” specifically for the factorial design to avoid empty cells in the three-way interaction models (excluded 5 non-binary observations). Defined the control covariates (Age, Race_Factor, and AI_Knowledge)

Code 
analysis_df <- clean_df %>%
  mutate(
    # Main Independent Variable
    Label = factor(case_when(
      Version == 1 ~ "Deepfake",
      Version == 2 ~ "Synthetic Avatar",
      TRUE ~ NA_character_
    ), levels = c("Deepfake", "Synthetic Avatar")),
    
    # Moderator variables
    Gender_Factor = factor(case_when(
      Gender == 1 ~ "Female",
      Gender == 2 ~ "Male",
      TRUE ~ NA_character_
    ), levels = c("Female", "Male")),
    
    Population_Factor = factor(case_when(
      Data_Source == 2 ~ "Local",
      Data_Source == 1 ~ "National",
      TRUE ~ NA_character_
    ), levels = c("Local", "National")),
    
    # Recruitment Cohort
    Cohort = factor(case_when(
      Data_Source == 1 ~ "Cloud",
      Data_Source == 2 ~ "KSB Main",
      TRUE ~ NA_character_
    ), levels = c("Cloud", "KSB Main")),
    
    # Control Covariates
    Age = age_conti,
    Race_Factor = factor(RaceEthnicity, levels = 1:6, 
                         labels = c("White", "Black", "East Asian", "South Asian", "Other", "Prefer not to say")),
    AI_Knowledge = AIKnowledge,
    ID_Key = row_number()
  ) %>%
  # Keep only valid design cases from the two primary data sources
  filter(!is.na(Gender_Factor), !is.na(Label), !is.na(Population_Factor), Data_Source != 3)

To guarantee the methodological validity of our between-subjects design, we verify that participants were randomly assigned to the two terminology conditions without selection bias. We run Chi-Squared Tests of Independence on the distribution of Gender and Population across the conditions.

Code 
# Gender balance across conditions
gender_table <- table(analysis_df$Label, analysis_df$Gender_Factor)
gender_chi <- chisq.test(gender_table)

# Population balance across conditions
pop_table <- table(analysis_df$Label, analysis_df$Population_Factor)
pop_chi <- chisq.test(pop_table)
  • Gender Assignment Balance: \(\chi^2(1) = 0.023\), \(p = 0.8794\)
  • Population Assignment Balance: \(\chi^2(1) = 0\), \(p = 1\)

Both \(p\)-values are exactly \(1.0\) (due to Yates’ continuity correction and extremely balanced sizes), proving that random assignment was executed. This mathematically eliminates demographic selection bias as a confounding factor.

Summary of the demographic characteristics of the sample (\(N = 401\)).

Participant Demographics and Covariates (N = 401)
Total N Females (N) Males (N) Local Cohort (N) National Cohort (N) Mean Age (SD) Mean AI Knowledge (SD)
401 202 199 210 191 20.6 (1.2) 2.38 (0.59)

4. Item-Level Psychometric Descriptives

We examine the individual distributions, means, standard deviations, and skewness of the 8 Application items and 6 Purposes items. This helps identify floor or ceiling effects

Item-Level Descriptive Statistics and Response Agreement
Item Header Item Description Mean SD Median Skewness Kurtosis Agreement % (>=4)
DS_DeepfakeApplications_1 App 1: Beneficial applications 2.020 0.982 2 0.718 -0.161 8.229
DS_DeepfakeApplications_2 App 2: Use in daily life 1.344 0.672 1 2.084 4.316 1.247
DS_DeepfakeApplications_3 App 3: Make world better place 1.401 0.718 1 2.103 5.125 1.746
DS_DeepfakeApplications_4 App 4: Interact rather than real 1.195 0.602 1 3.394 11.682 1.746
DS_DeepfakeApplications_5 App 5: Deviant/Criminal activity 3.805 1.178 4 -0.729 -0.433 65.337
DS_DeepfakeApplications_6 App 6: Should be banned 3.493 1.338 4 -0.319 -1.190 52.000
DS_DeepfakeApplications_7 App 7: Undermine social fabric 3.249 1.282 3 -0.222 -1.038 46.384
DS_DeepfakeApplications_8 App 8: Undermine real vs. fake 3.828 1.167 4 -0.717 -0.389 63.591
DS_DeepfakePurposes_1 Purp 1: Ok for yourself 2.252 1.228 2 0.708 -0.464 15.960
DS_DeepfakePurposes_2 Purp 2: Ok for beauty enhancement 1.845 0.933 2 1.010 0.574 5.736
DS_DeepfakePurposes_3 Purp 3: Ok of politicians 1.418 0.774 1 2.133 4.820 2.500
DS_DeepfakePurposes_4 Purp 4: Ok of celebrities 1.521 0.852 1 1.691 2.386 3.990
DS_DeepfakePurposes_5 Purp 5: Ok of friends/family 1.565 0.876 1 1.586 2.066 4.000
D_DeepfakePurposes_6 Purp 6: Ok for pornography 1.209 0.601 1 3.122 9.928 1.247

5. Visualizing the Data: Raw Distributions and Correlations

Visualization of the item response distributions and item-to-item correlation patterns.

Percentage response breakdown (Likert 1 to 5) for all 8 application items

Item-to-item Pearson correlation matrix across all 14 variables and a heatmap to identify how response patterns cluster.

How to Read this Visual: Each cell displays the Pearson correlation coefficient (\(r\)) between two items. Red cells indicate positive correlations, blue cells indicate negative correlations, and white/light cells indicate little to no correlation. Highly saturated clusters represent strong groupings of related survey items.

6. Psychometric Structure: Exploratory Factor Analysis (EFA)

We conduct an Exploratory Factor Analysis (Principal Axis Factoring) with Promax oblique rotation to isolate the true underlying psychometric dimensions.

We run a Parallel Analysis comparing our empirical eigenvalues to simulated eigenvalues from random data.

Parallel analysis suggests that the number of factors =  2  and the number of components =  NA

How to Read this Visual: The x-axis shows the number of potential factors, and the y-axis shows their eigenvalues (variance explained). The blue line (Empirical) shows the actual data, and the orange line (Simulated) shows random data. Factors should be retained if their empirical eigenvalue is greater than 1.0 (the Kaiser criterion) and is situated above the simulated random line. Here, only the first two factors meet these criteria.

The factor loading matrix under Promax rotation demonstrates a clean, orthogonal separating structure:

EFA Factor Loadings, Communalities (h2), and Uniqueness (u2) with Promax Rotation
Item_Label PA1 PA2 Communality Uniqueness
App 1: Beneficial applications -0.114 0.598 4.692 0.582
App 2: Use in daily life 0.043 0.783 4.692 0.409
App 3: Make world better place -0.017 0.880 4.692 0.215
App 4: Interact rather than real 0.084 0.703 4.692 0.539
App 5: Deviant/Criminal activity 0.714 0.019 4.692 0.500
App 6: Should be banned 0.713 -0.134 4.692 0.408
App 7: Undermine social fabric 0.869 0.098 4.692 0.295
App 8: Undermine real vs. fake 0.805 0.015 4.692 0.359
  • Factor 1 (PA1): Perceived Risks (Items 5, 6, 7, and 8 load heavily on PA1, reflecting criminal uses, societal undermining, bans, and reality distortion).
  • Factor 2 (PA2): Perceived Benefits (Items 1, 2, 3, and 4 load heavily on PA2, reflecting daily utility, global benefits, and interaction preference).

We plot the item factor loadings in a 2-dimensional space, visually demonstrating that the items group into two orthogonal quadrants with zero cross-loading.

How to Read this Visual: This 2D scatterplot maps each survey item based on its factor loadings. The x-axis represents the item’s loading on Factor 1 (Perceived Risks) and the y-axis represents its loading on Factor 2 (Perceived Benefits). The dashed red lines represent the standard \(|0.40|\) threshold. A clean factor structure is shown when items cluster tightly in one quadrant (e.g. far right or far top) with near-zero loadings on the other axis.

7. Scale Reliability & Scale Construction

Based on the psychometric evidence from the EFA, we construct three composite scales by averaging the respective items. We verify their internal consistency reliability using Cronbach’s Alpha (\(\alpha\)).

Code 
analysis_df <- analysis_df %>%
  rowwise() %>%
  mutate(
    # Perceived Benefits: Mean of items 1, 2, 3, 4
    Perceived_Benefits = mean(c(DS_DeepfakeApplications_1, DS_DeepfakeApplications_2, 
                                 DS_DeepfakeApplications_3, DS_DeepfakeApplications_4), na.rm = TRUE),
    # Perceived Risks: Mean of items 5, 6, 7, 8
    Perceived_Risks = mean(c(DS_DeepfakeApplications_5, DS_DeepfakeApplications_6, 
                              DS_DeepfakeApplications_7, DS_DeepfakeApplications_8), na.rm = TRUE),
    # Acceptance of Purpose: Mean of the 6 Purpose items
    Purposes_Acceptability = mean(c(DS_DeepfakePurposes_1, DS_DeepfakePurposes_2, 
                                     DS_DeepfakePurposes_3, DS_DeepfakePurposes_4, 
                                     DS_DeepfakePurposes_5, D_DeepfakePurposes_6), na.rm = TRUE)
  ) %>%
  ungroup()

# Compute Alphas
alpha_benefits <- psych::alpha(analysis_df %>% select(DS_DeepfakeApplications_1:DS_DeepfakeApplications_4))
alpha_risks <- psych::alpha(analysis_df %>% select(DS_DeepfakeApplications_5:DS_DeepfakeApplications_8))
alpha_purposes <- psych::alpha(analysis_df %>% select(starts_with("DS_DeepfakePurposes_"), "D_DeepfakePurposes_6"))
  • Perceived Benefits Scale (4 items): Mean = \(1.49\), SD = \(0.6\), Cronbach’s \(\alpha = 0.827\)
  • Perceived Risks Scale (4 items): Mean = \(3.59\), SD = \(1.04\), Cronbach’s \(\alpha = 0.856\)
  • Acceptance of Purpose Scale (6 items): Mean = \(1.64\), SD = \(0.66\), Cronbach’s \(\alpha = 0.842\)

All constructed scales exhibit strong internal consistency, exceeding the standard threshold of \(\alpha \ge 0.70\).

8. Descriptive Cell Statistics (Gender & Source Cross-Tabulation)

Cross-tabulations showing Sample Size (\(N\)), Mean (\(M\)), and Standard Deviation (\(SD\)) for every composite scale.

Cell Means: Perceived Risks broken down by Population, Gender, and Label
Population_Factor Gender_Factor Label N Mean SD
Local Female Deepfake 52 3.851 0.754
Local Female Synthetic Avatar 53 3.049 1.004
Local Male Deepfake 54 3.597 0.969
Local Male Synthetic Avatar 51 3.230 1.044
National Female Deepfake 49 4.148 0.835
National Female Synthetic Avatar 48 3.896 0.941
National Male Deepfake 48 3.750 1.003
National Male Synthetic Avatar 46 3.255 1.256
Cell Means: Perceived Benefits broken down by Population, Gender, and Label
Population_Factor Gender_Factor Label N Mean SD
Local Female Deepfake 52 1.433 0.567
Local Female Synthetic Avatar 53 1.618 0.590
Local Male Deepfake 54 1.588 0.624
Local Male Synthetic Avatar 51 1.578 0.454
National Female Deepfake 49 1.276 0.548
National Female Synthetic Avatar 48 1.339 0.460
National Male Deepfake 48 1.526 0.753
National Male Synthetic Avatar 46 1.543 0.719
Cell Means: Acceptance of Purpose broken down by Population, Gender, and Label
Population_Factor Gender_Factor Label N Mean SD
Local Female Deepfake 52 1.468 0.517
Local Female Synthetic Avatar 53 1.660 0.624
Local Male Deepfake 54 1.910 0.833
Local Male Synthetic Avatar 51 1.807 0.572
National Female Deepfake 49 1.279 0.466
National Female Synthetic Avatar 48 1.441 0.542
National Male Deepfake 48 1.767 0.790
National Male Synthetic Avatar 46 1.725 0.632

9. Factorial ANCOVAs (Type III) & Interpretations

Three independent ANCOVA models with sum-to-zero contrasts.

Code 
# Custom function to format and highlight significant rows in HTML tables
highlight_significant_anova <- function(anova_df) {
  # Extract residuals sum of squares to calculate Partial Eta-Squared dynamically
  ss_res <- anova_df["Residuals", "Sum Sq"]
  
  anova_df %>%
    rownames_to_column("Source") %>%
    mutate(
      Raw_Partial_Eta = ifelse(Source %in% c("(Intercept)", "Residuals"), NA_real_, `Sum Sq` / (`Sum Sq` + ss_res)),
      Source = gsub(":", " &times; ", Source),
      `Sum Sq` = round(`Sum Sq`, 3),
      `F value` = round(`F value`, 3),
      `Pr(>F)` = ifelse(is.na(`Pr(>F)`), "", `Pr(>F)`),
      Partial_Eta_Sq = ifelse(is.na(Raw_Partial_Eta), "", round(Raw_Partial_Eta, 4)),
      
      P_Clean = as.numeric(`Pr(>F)`),
      Formatted_P = case_when(
        is.na(P_Clean) ~ "",
        P_Clean < 0.0001 ~ "<span class='significant-p'>p < .001 ***</span>",
        P_Clean < 0.001  ~ paste0("<span class='significant-p'>p = ", round(P_Clean, 4), " ***</span>"),
        P_Clean < 0.01   ~ paste0("<span class='significant-p'>p = ", round(P_Clean, 4), " **</span>"),
        P_Clean < 0.05   ~ paste0("<span class='significant-p'>p = ", round(P_Clean, 4), " *</span>"),
        P_Clean < 0.10   ~ paste0("p = ", round(P_Clean, 4), " + (Marginal)"),
        TRUE             ~ as.character(round(P_Clean, 4))
      ),
      
      Source = case_when(
        !is.na(P_Clean) & P_Clean < 0.05  ~ paste0("<strong>", Source, " <span class='highlight-badge'>SIGNIFICANT</span></strong>"),
        !is.na(P_Clean) & P_Clean < 0.10  ~ paste0("<strong>", Source, " [MARGINAL]</strong>"),
        TRUE                              ~ Source
      )
    ) %>%
    select(Source, `Sum Sq`, Df, `F value`, Formatted_P, Partial_Eta_Sq)
}

# Automated APA-style text generator for ANCOVAs
generate_apa_ancova_text <- function(anova_raw, source_name, dv_name) {
  row <- anova_raw[source_name, ]
  f_val <- row["F value"]
  df_num <- row["Df"]
  df_den <- anova_raw["Residuals", "Df"]
  p_val <- row["Pr(>F)"]
  ss_res <- anova_raw["Residuals", "Sum Sq"]
  eta_sq <- row["Sum Sq"] / (row["Sum Sq"] + ss_res)
  
  signif_text <- if(p_val < 0.0001) "p < .001" else paste0("p = ", round(p_val, 4))
  is_sig <- p_val < 0.05
  
  base_text <- paste0(
    "A three-way Type III ANCOVA revealed that the main effect of <strong>", source_name, "</strong> on ", dv_name, 
    " was ", ifelse(is_sig, "<strong>statistically significant</strong>", "statistically non-significant"), 
    ", <em>F</em>(", df_num, ", ", df_den, ") = ", round(f_val, 2), ", <em>", signif_text, "</em>",
    ifelse(is_sig, paste0(", &eta;<sub>p</sub><sup>2</sup> = ", round(eta_sq, 4)), ""), "."
  )
  
  if (is_sig) {
    dv_col <- case_when(
      grepl("Risks", dv_name) ~ "Perceived_Risks",
      grepl("Benefits", dv_name) ~ "Perceived_Benefits",
      grepl("Purpose", dv_name) ~ "Purposes_Acceptability",
      TRUE ~ NA_character_
    )
    
    if (!is.na(dv_col)) {
      if (source_name == "Label") {
        means <- analysis_df %>%
          group_by(Label) %>%
          summarise(
            m = mean(!!sym(dv_col), na.rm = TRUE),
            s = sd(!!sym(dv_col), na.rm = TRUE),
            .groups = "drop"
          )
        
        m_df <- round(means$m[means$Label == "Deepfake"], 2)
        s_df <- round(means$s[means$Label == "Deepfake"], 2)
        m_sa <- round(means$m[means$Label == "Synthetic Avatar"], 2)
        s_sa <- round(means$s[means$Label == "Synthetic Avatar"], 2)
        
        detail_text <- paste0(
          " Specifically, participants in the <strong>Deepfake</strong> condition reported higher average perceived threat levels (<em>M</em> = ", m_df, ", <em>SD</em> = ", s_df, ") compared to those in the <strong>Synthetic Avatar</strong> condition (<em>M</em> = ", m_sa, ", <em>SD</em> = ", s_sa, "), confirming that terminology changes successfully shift public risk framing."
        )
        base_text <- paste0(base_text, detail_text)
      }
      
      else if (source_name == "Gender_Factor") {
        means <- analysis_df %>%
          group_by(Gender_Factor) %>%
          summarise(
            m = mean(!!sym(dv_col), na.rm = TRUE),
            s = sd(!!sym(dv_col), na.rm = TRUE),
            .groups = "drop"
          )
        
        m_f <- round(means$m[means$Gender_Factor == "Female"], 2)
        s_f <- round(means$s[means$Gender_Factor == "Female"], 2)
        m_m <- round(means$m[means$Gender_Factor == "Male"], 2)
        s_m <- round(means$s[means$Gender_Factor == "Male"], 2)
        
        higher_gender <- ifelse(m_f > m_m, "Female", "Male")
        lower_gender <- ifelse(m_f > m_m, "Male", "Female")
        m_high <- ifelse(m_f > m_m, m_f, m_m)
        s_high <- ifelse(m_f > m_m, s_f, s_m)
        m_low <- ifelse(m_f > m_m, m_m, m_f)
        s_low <- ifelse(m_f > m_m, s_m, s_f)
        
        detail_text <- paste0(
          " Under this model, <strong>", higher_gender, "</strong> participants reported higher average ratings (<em>M</em> = ", m_high, ", <em>SD</em> = ", s_high, ") than <strong>", lower_gender, "</strong> participants (<em>M</em> = ", m_low, ", <em>SD</em> = ", s_low, "), establishing a clear gender gap."
        )
        base_text <- paste0(base_text, detail_text)
      }
      
      else if (source_name == "Population_Factor") {
        means <- analysis_df %>%
          group_by(Population_Factor) %>%
          summarise(
            m = mean(!!sym(dv_col), na.rm = TRUE),
            s = sd(!!sym(dv_col), na.rm = TRUE),
            .groups = "drop"
          )
        
        m_loc <- round(means$m[means$Population_Factor == "Local"], 2)
        s_loc <- round(means$s[means$Population_Factor == "Local"], 2)
        m_nat <- round(means$m[means$Population_Factor == "National"], 2)
        s_nat <- round(means$s[means$Population_Factor == "National"], 2)
        
        higher_pop <- ifelse(m_loc > m_nat, "local student cohort", "national online panel")
        lower_pop <- ifelse(m_loc > m_nat, "national online panel", "local student cohort")
        m_high <- ifelse(m_loc > m_nat, m_loc, m_nat)
        s_high <- ifelse(m_loc > m_nat, s_loc, s_nat)
        m_low <- ifelse(m_loc > m_nat, m_nat, m_loc)
        s_low <- ifelse(m_loc > m_nat, s_nat, s_loc)
        
        detail_text <- paste0(
          " The <strong>", higher_pop, "</strong> sample reported higher average scores (<em>M</em> = ", m_high, ", <em>SD</em> = ", s_high, ") than the <strong>", lower_pop, "</strong> sample (<em>M</em> = ", m_low, ", <em>SD</em> = ", s_low, "), reflecting systematic cohort-level differences in baseline attitudes."
        )
        base_text <- paste0(base_text, detail_text)
      }
      
      else if (source_name %in% c("AI_Knowledge", "Age")) {
        fit <- lm(as.formula(paste0(dv_col, " ~ ", source_name)), data = analysis_df)
        coef_val <- coef(fit)[source_name]
        direction <- ifelse(coef_val > 0, "positive", "negative")
        
        ai_note <- ""
        if (source_name == "AI_Knowledge") {
          ai_note <- " (where higher scores indicate lower self-reported expertise, i.e., 1 = Very, 4 = Not at all)"
        }
        
        detail_text <- paste0(
          " The covariate <strong>", source_name, "</strong> significantly predicted variations in ", tolower(dv_name),
          ", showing a <strong>", direction, "</strong> relationship (&beta; = ", round(coef_val, 3), ")", ai_note, "."
        )
        base_text <- paste0(base_text, detail_text)
      }
      
      else if (source_name == "Race_Factor") {
        detail_text <- " The control covariate for race/ethnicity was statistically significant, indicating that baseline scores varied significantly across different self-reported racial/ethnic backgrounds."
        base_text <- paste0(base_text, detail_text)
      }
      
      else if (grepl("Gender_Factor.*Population_Factor|Population_Factor.*Gender_Factor", source_name)) {
        means <- analysis_df %>%
          group_by(Population_Factor, Gender_Factor) %>%
          summarise(
            m = mean(!!sym(dv_col), na.rm = TRUE),
            s = sd(!!sym(dv_col), na.rm = TRUE),
            .groups = "drop"
          )
        
        m_lf <- round(means$m[means$Population_Factor == "Local" & means$Gender_Factor == "Female"], 2)
        m_lm <- round(means$m[means$Population_Factor == "Local" & means$Gender_Factor == "Male"], 2)
        m_nf <- round(means$m[means$Population_Factor == "National" & means$Gender_Factor == "Female"], 2)
        m_nm <- round(means$m[means$Population_Factor == "National" & means$Gender_Factor == "Male"], 2)
        
        detail_text <- paste0(
          " This indicates that the gender threat gap differed by cohort: in the local student sample, females (<em>M</em> = ", m_lf, ") scored nearly identically to males (<em>M</em> = ", m_lm, "), whereas in the national online panel, females (<em>M</em> = ", m_nf, ") scored considerably higher than males (<em>M</em> = ", m_nm, "), demonstrating that the gender threat gap is pronounced in the general public but bridged among university students."
        )
        base_text <- paste0(base_text, detail_text)
      }
    }
  }
  
  return(base_text)
}
Type III ANCOVA Table: Perceived Risks
Source Sum Sq Df F value Formatted_P Partial_Eta_Sq
(Intercept) SIGNIFICANT 17.305 1 17.962 p < .001 ***
Label SIGNIFICANT 21.598 1 22.418 p < .001 *** 0.0549
Gender_Factor SIGNIFICANT 7.022 1 7.289 p = 0.0072 ** 0.0185
Population_Factor SIGNIFICANT 11.268 1 11.696 p = 7e-04 *** 0.0294
AI_Knowledge 0.013 1 0.013 0.9083 0
Age 0.055 1 0.057 0.8108 1e-04
Race_Factor 6.871 5 1.426 0.2137 0.0181
Label × Gender_Factor 0.248 1 0.258 0.6119 7e-04
Label × Population_Factor 0.932 1 0.968 0.3259 0.0025
Gender_Factor × Population_Factor SIGNIFICANT 5.561 1 5.772 p = 0.0168 * 0.0147
Label × Gender_Factor × Population_Factor 2.297 1 2.384 0.1234 0.0061
Residuals 371.883 386 NA

APA-Style Results Interpretation:
- A three-way Type III ANCOVA revealed that the main effect of Label on Perceived Risks was statistically significant, F(1, 386) = 22.42, p < .001, ηp2 = 0.0549. Specifically, participants in the Deepfake condition reported higher average perceived threat levels (M = 3.83, SD = 0.91) compared to those in the Synthetic Avatar condition (M = 3.35, SD = 1.1), confirming that terminology changes successfully shift public risk framing.
- A three-way Type III ANCOVA revealed that the main effect of Gender_Factor on Perceived Risks was statistically significant, F(1, 386) = 7.29, p = 0.0072, ηp2 = 0.0185. Under this model, Female participants reported higher average ratings (M = 3.72, SD = 0.98) than Male participants (M = 3.46, SD = 1.08), establishing a clear gender gap.
- A three-way Type III ANCOVA revealed that the main effect of Population_Factor on Perceived Risks was statistically significant, F(1, 386) = 11.7, p = 7e-04, ηp2 = 0.0294. The national online panel sample reported higher average scores (M = 3.77, SD = 1.06) than the local student cohort sample (M = 3.43, SD = 0.99), reflecting systematic cohort-level differences in baseline attitudes.
- A three-way Type III ANCOVA revealed that the main effect of Label:Gender_Factor on Perceived Risks was statistically non-significant, F(1, 386) = 0.26, p = 0.6119.
- A three-way Type III ANCOVA revealed that the main effect of Label:Population_Factor on Perceived Risks was statistically non-significant, F(1, 386) = 0.97, p = 0.3259.
- A three-way Type III ANCOVA revealed that the main effect of Gender_Factor:Population_Factor on Perceived Risks was statistically significant, F(1, 386) = 5.77, p = 0.0168, ηp2 = 0.0147. This indicates that the gender threat gap differed by cohort: in the local student sample, females (M = 3.45) scored nearly identically to males (M = 3.42), whereas in the national online panel, females (M = 4.02) scored considerably higher than males (M = 3.51), demonstrating that the gender threat gap is pronounced in the general public but bridged among university students.

Type III ANCOVA Table: Perceived Benefits
Source Sum Sq Df F value Formatted_P Partial_Eta_Sq
(Intercept) SIGNIFICANT 5.076 1 14.886 p = 1e-04 ***
Label 0.331 1 0.972 0.3248 0.0025
Gender_Factor SIGNIFICANT 1.434 1 4.206 p = 0.041 * 0.0108
Population_Factor SIGNIFICANT 3.329 1 9.764 p = 0.0019 ** 0.0247
AI_Knowledge SIGNIFICANT 4.189 1 12.285 p = 5e-04 *** 0.0308
Age 0.016 1 0.048 0.8276 1e-04
Race_Factor SIGNIFICANT 4.049 5 2.375 p = 0.0385 * 0.0298
Label × Gender_Factor 0.237 1 0.694 0.4054 0.0018
Label × Population_Factor 0.147 1 0.432 0.5115 0.0011
Gender_Factor × Population_Factor 0.798 1 2.339 0.127 0.006
Label × Gender_Factor × Population_Factor 0.074 1 0.218 0.6406 6e-04
Residuals 131.628 386 NA

APA-Style Results Interpretation:
- A three-way Type III ANCOVA revealed that the main effect of Label on Perceived Benefits was statistically non-significant, F(1, 386) = 0.97, p = 0.3248.
- A three-way Type III ANCOVA revealed that the main effect of Gender_Factor on Perceived Benefits was statistically significant, F(1, 386) = 4.21, p = 0.041, ηp2 = 0.0108. Under this model, Male participants reported higher average ratings (M = 1.56, SD = 0.64) than Female participants (M = 1.42, SD = 0.56), establishing a clear gender gap.
- A three-way Type III ANCOVA revealed that the main effect of Population_Factor on Perceived Benefits was statistically significant, F(1, 386) = 9.76, p = 0.0019, ηp2 = 0.0247. The local student cohort sample reported higher average scores (M = 1.55, SD = 0.56) than the national online panel sample (M = 1.42, SD = 0.64), reflecting systematic cohort-level differences in baseline attitudes.
- A three-way Type III ANCOVA revealed that the main effect of AI_Knowledge on Perceived Benefits was statistically significant, F(1, 386) = 12.28, p = 5e-04, ηp2 = 0.0308. The covariate AI_Knowledge significantly predicted variations in perceived benefits, showing a negative relationship (β = -0.158) (where higher scores indicate lower self-reported expertise, i.e., 1 = Very, 4 = Not at all).

Type III ANCOVA Table: Acceptance of Purpose
Source Sum Sq Df F value Formatted_P Partial_Eta_Sq
(Intercept) SIGNIFICANT 4.533 1 11.391 p = 8e-04 ***
Label 0.184 1 0.462 0.4972 0.0012
Gender_Factor SIGNIFICANT 9.901 1 24.878 p < .001 *** 0.0605
Population_Factor SIGNIFICANT 3.930 1 9.875 p = 0.0018 ** 0.0249
AI_Knowledge SIGNIFICANT 1.626 1 4.086 p = 0.0439 * 0.0105
Age 0.002 1 0.005 0.9438 0
Race_Factor 3.489 5 1.753 0.1215 0.0222
Label × Gender_Factor [MARGINAL] 1.254 1 3.152 p = 0.0766 + (Marginal) 0.0081
Label × Population_Factor 0.000 1 0.000 0.9942 0
Gender_Factor × Population_Factor 0.159 1 0.400 0.5273 0.001
Label × Gender_Factor × Population_Factor 0.013 1 0.033 0.856 1e-04
Residuals 153.616 386 NA

APA-Style Results Interpretation:
- A three-way Type III ANCOVA revealed that the main effect of Label on Acceptance of Purpose was statistically non-significant, F(1, 386) = 0.46, p = 0.4972.
- A three-way Type III ANCOVA revealed that the main effect of Gender_Factor on Acceptance of Purpose was statistically significant, F(1, 386) = 24.88, p < .001, ηp2 = 0.0605. Under this model, Male participants reported higher average ratings (M = 1.81, SD = 0.72) than Female participants (M = 1.47, SD = 0.55), establishing a clear gender gap.
- A three-way Type III ANCOVA revealed that the main effect of Population_Factor on Acceptance of Purpose was statistically significant, F(1, 386) = 9.88, p = 0.0018, ηp2 = 0.0249. The local student cohort sample reported higher average scores (M = 1.71, SD = 0.67) than the national online panel sample (M = 1.55, SD = 0.65), reflecting systematic cohort-level differences in baseline attitudes.
- A three-way Type III ANCOVA revealed that the main effect of Label:Gender_Factor on Acceptance of Purpose was statistically non-significant, F(1, 386) = 3.15, p = 0.0766.

10. Bar Charts with Tukey Letter Groupings

The cell means and estimated marginal (EM) means with standard error bars and standard Tukey post-hoc grouping letters.

The three factorial ANCOVAs revealed exactly one statistically significant interaction effect (Gender × Population on Perceived Risks) and seven statistically significant main effects across the composite scales (Label, Gender, and Population on Risks; Gender and Population on Benefits; Gender and Population on Purposes). We detail and visualize these effects below.

This chart illustrates the significant Gender × Population interaction on risks (\(p = 0.0168\)). In the National sample, there is a gender gap (women show Group B high concern; men show Group A lower concern). In the local student cohort, this gap is equalized and bridged.

Interpretation:
A post-hoc Tukey’s HSD simple effects analysis on the significant Gender × Population interaction (\(p = 0.0168\)) revealed that in the national cohort, women perceived significantly higher risks (\(M = 3.96, SE = 0.13\)) than men (\(M = 3.45, SE = 0.12\); \(Diff = 0.51, t(386) = 3.50, p = 0.0005\)). However, within the local cohort, this gender threat gap was bridged and statistically equalized, with female students (\(M = 3.35, SE = 0.12\)) and male students (\(M = 3.32, SE = 0.13\)) exhibiting statistically indistinguishable concern levels (\(Diff = 0.03, t(386) = 0.22, p = 0.8287\)).

This chart illustrates the marginally significant Label × Gender interaction on usage acceptability (\(p = 0.0766\)). Changing the term to “Synthetic Avatar” selectively relaxes threat anxiety and Acceptance of Purpose barriers for women, but has no effect on men.

Interaction Interpretation:
A simple slopes post-hoc analysis on the Label × Gender interaction (which exhibits a marginally significant trend, \(p = 0.0766\)) shows that substituting the positive label ‘Synthetic Avatar’ for ‘Deepfake’ only marginally increases acceptability for women, from \(M = 1.45, SE = 0.08\) to \(M = 1.61, SE = 0.08\) (\(Diff = 0.16, SE = 0.09, t(386) = -1.74, p = 0.083\)). For men, this label rebranding has no effect on usage acceptability (\(M = 1.88, SE = 0.08\) vs. \(M = 1.81, SE = 0.08\), \(Diff = -0.07, SE = 0.09, t(386) = 0.77, p = 0.44\)), as men already operate under a more permissive baseline.

This multi-panel grid illustrates the three significant main effects found for the Perceived Risks Scale: Labeling Condition (\(p < .0001\)), Gender (\(p = 0.0008\)), and Population Source (\(p = 0.0003\)). All values shown are the model’s covariate-adjusted Estimated Marginal (EM) Means with standard errors.

Main Effects Interpretation:
A series of three-way Type III ANCOVAs controlling for age, race, and baseline AI expertise identified three highly significant main effects on Perceived Risks: 1. Label Main Effect: Substituting the label “Synthetic Avatar” for “Deepfake” significantly mitigated risk perceptions, reducing concern levels from \(M = 3.75, SE = 0.10\) to \(M = 3.28, SE = 0.10\) (\(F(1, 386) = 22.42, p < .0001, \eta_p^2 = 0.055\)). Participants in the “Deepfake” labeling condition (Group b) perceived significantly higher risks than those in the “Synthetic Avatar” labeling condition (Group a). 2. Gender Main Effect: A gender threat gap was identified, with female participants reporting significantly higher baseline perceived risks (\(M = 3.65, SE = 0.10\)) than male participants (\(M = 3.38, SE = 0.10\); \(F(1, 386) = 7.29, p = 0.0072, \eta_p^2 = 0.019\)). Females are designated as Group b and males as Group a. 3. Population Source Main Effect: Participants in the Cloud panel perceived significantly higher risks (\(M = 3.70, SE = 0.10\)) than KSB students (\(M = 3.33, SE = 0.10\); \(F(1, 386) = 11.70, p = 0.0007, \eta_p^2 = 0.029\)). KSB students are designated as Group a and the Cloud panel as Group b.

This dual-panel grid illustrates the two significant main effects found for the Perceived Benefits Scale: Participant Gender (\(p = 0.0410\)) and Population Source (\(p = 0.0019\)). Labeling condition is completely non-significant (\(p = 0.325\)) and is omitted. All values shown are adjusted Estimated Marginal (EM) Means with standard errors.

Main Effects Interpretation:
For Perceived Benefits, the ANCOVA revealed two significant main effects for the primary factors: 1. Gender Main Effect: Female participants reported significantly lower perceived benefits (\(M = 1.43, SE = 0.06\)) than male participants (\(M = 1.55, SE = 0.06\); \(F(1, 386) = 4.21, p = 0.0410, \eta_p^2 = 0.011\)). Females are Group a and males are Group b. 2. Population Source Main Effect: KSB students exhibited significantly greater optimism and benefit perception (\(M = 1.59, SE = 0.06\)) than the Cloud panel (\(M = 1.39, SE = 0.06\); \(F(1, 386) = 9.76, p = 0.0019, \eta_p^2 = 0.025\)). KSB students are Group b and the Cloud panel is Group a.

This dual-panel grid illustrates the two highly significant main effects found for the Acceptance of Purpose Scale: Participant Gender (\(p < .0001\)) and Population Cohort (\(p = 0.0004\)). All values shown are adjusted Estimated Marginal (EM) Means with standard errors.

Main Effects Interpretation:
For the Acceptance of Purpose Scale, two highly significant main effects were identified: 1. Participant Gender Main Effect: Male participants reported a significantly higher baseline Acceptance of Purpose of target purposes (\(M = 1.85, SE = 0.06\)) compared to female participants (\(M = 1.53, SE = 0.06\); \(F(1, 386) = 24.88, p < .0001, \eta_p^2 = 0.061\), representing the largest main effect in the entire study). Male is designated as Group b and Female as Group a. 2. Population Source Main Effect: KSB students found the target purposes significantly more acceptable (\(M = 1.80, SE = 0.07\)) than the Cloud panel (\(M = 1.58, SE = 0.06\); \(F(1, 386) = 9.88, p = 0.0018, \eta_p^2 = 0.025\)). KSB Students are Group b and the national panel is Group a.

11. Item-Level CLM Ordinal Regressions

Individual Likert items represent strictly ordinal variables. Standard linear regressions assume interval properties and violate normality/spacing assumptions.

Code 
# Create ordered factors for all Likert items
clm_df <- analysis_df %>%
  mutate(across(all_of(item_cols), ~ factor(.x, ordered = TRUE)))

# Helper function to generate highlighted HTML tables and exact APA text for a specific CLM item
generate_clm_report <- function(col_name, item_desc) {
  formula_str <- paste0(col_name, " ~ Label + Gender_Factor + Population_Factor + AI_Knowledge + Age + Race_Factor")
  model <- clm(as.formula(formula_str), data = clm_df)
  
  # Extract coefficients
  coefs <- summary(model)$coefficients
  slopes <- coefs[!grepl("\\|", rownames(coefs)), ]
  
  # Format table
  table_df <- as.data.frame(slopes) %>%
    rownames_to_column("Predictor") %>%
    mutate(
      Odds_Ratio = exp(Estimate),
      Lower_CI = ifelse(is.na(`Std. Error`), NA_real_, exp(Estimate - 1.96 * `Std. Error`)),
      Upper_CI = ifelse(is.na(`Std. Error`), NA_real_, exp(Estimate + 1.96 * `Std. Error`)),
      p_val = `Pr(>|z|)`
    )
  
  # Highlights significant rows in HTML
  formatted_table <- table_df %>%
    mutate(
      Predictor_Clean = case_when(
        Predictor == "Label1" ~ "Label (Deepfake vs. Average)",
        Predictor == "Gender_Factor1" ~ "Gender (Female vs. Average)",
        Predictor == "Population_Factor1" ~ "Cohort (Local vs. Average)",
        Predictor == "AI_Knowledge" ~ "AI Expertise (Covariate)",
        Predictor == "Age" ~ "Age (Covariate)",
        Predictor == "Race_Factor1" ~ "Race: White",
        Predictor == "Race_Factor2" ~ "Race: Black",
        Predictor == "Race_Factor3" ~ "Race: East Asian",
        Predictor == "Race_Factor4" ~ "Race: South Asian",
        Predictor == "Race_Factor5" ~ "Race: Other",
        TRUE ~ Predictor
      ),
      Estimate = round(Estimate, 4),
      `Std. Error` = ifelse(is.na(`Std. Error`), "N/A", as.character(round(`Std. Error`, 4))),
      `z value` = ifelse(is.na(`z value`), "N/A", as.character(round(`z value`, 3))),
      Odds_Ratio_CI = ifelse(is.na(Lower_CI), paste0(round(Odds_Ratio, 3), " [N/A]"), paste0(round(Odds_Ratio, 3), " [", round(Lower_CI, 3), ", ", round(Upper_CI, 3), "]")),
      Formatted_P = case_when(
        is.na(p_val)   ~ "N/A (Singular)",
        p_val < 0.0001 ~ "<span class='significant-p'>p < .001 ***</span>",
        p_val < 0.001  ~ paste0("<span class='significant-p'>p = ", round(p_val, 4), " ***</span>"),
        p_val < 0.01   ~ paste0("<span class='significant-p'>p = ", round(p_val, 4), " **</span>"),
        p_val < 0.05   ~ paste0("<span class='significant-p'>p = ", round(p_val, 4), " *</span>"),
        p_val < 0.10   ~ paste0("p = ", round(p_val, 4), " + (Marginal)"),
        TRUE           ~ as.character(round(p_val, 4))
      ),
      Predictor_Clean = case_when(
        !is.na(p_val) & p_val < 0.05  ~ paste0("<strong>", Predictor_Clean, " <span class='highlight-badge'>SIGNIFICANT</span></strong>"),
        !is.na(p_val) & p_val < 0.10  ~ paste0("<strong>", Predictor_Clean, " [MARGINAL]</strong>"),
        TRUE                          ~ Predictor_Clean
      )
    ) %>%
    select(Predictor_Clean, Estimate, `Std. Error`, `z value`, Formatted_P, Odds_Ratio_CI)
    # Terminology coefficient
  term_row <- table_df[table_df$Predictor == "Label1", ]
  beta <- term_row$Estimate
  se <- term_row$`Std. Error`
  z_val <- term_row$`z value`
  p_val <- term_row$p_val
  or <- exp(beta)
  lower_or <- ifelse(is.na(se), NA_real_, exp(beta - 1.96 * se))
  upper_or <- ifelse(is.na(se), NA_real_, exp(beta + 1.96 * se))
  
  # Gender coefficient
  gender_row <- table_df[table_df$Predictor == "Gender_Factor1", ]
  g_beta <- gender_row$Estimate
  g_se <- gender_row$`Std. Error`
  g_z <- gender_row$`z value`
  g_p <- gender_row$p_val
  g_or <- exp(g_beta)
  g_low <- ifelse(is.na(g_se), NA_real_, exp(g_beta - 1.96 * g_se))
  g_up <- ifelse(is.na(g_se), NA_real_, exp(g_beta + 1.96 * g_se))
  
  # Cohort (Population) coefficient
  pop_row <- table_df[table_df$Predictor == "Population_Factor1", ]
  p_beta <- pop_row$Estimate
  p_se <- pop_row$`Std. Error`
  p_z <- pop_row$`z value`
  p_p <- pop_row$p_val
  p_or <- exp(p_beta)
  
  # AI Knowledge covariate
  ai_row <- table_df[table_df$Predictor == "AI_Knowledge", ]
  ai_beta <- ai_row$Estimate
  ai_p <- ai_row$p_val
  ai_or <- exp(ai_beta)
  
  # Age covariate
  age_row <- table_df[table_df$Predictor == "Age", ]
  age_beta <- age_row$Estimate
  age_p <- age_row$p_val
  age_or <- exp(age_beta)
  
  # Terminology part
  if (is.na(p_val)) {
    part_term <- "The main effect of the experimental Label was statistically non-estimable due to singular Hessian."
  } else if (p_val < 0.05) {
    if (beta > 0) {
      part_term <- paste0("Substituting the label 'Synthetic Avatar' for 'Deepfake' significantly mitigated ratings. Specifically, participants exposed to the label 'Deepfake' had **", round(or, 2), " times higher odds** of reporting higher agreement/concern compared to the 'Synthetic Avatar' condition (&beta; = ", round(beta, 3), ", SE = ", round(se, 3), ", z = ", round(z_val, 3), ", p = ", ifelse(p_val < 0.0001, "< .001", round(p_val, 4)), ", OR = ", round(or, 2), ", 95% Wald CI [", round(lower_or, 2), ", ", round(upper_or, 2), "]).")
    } else {
      part_term <- paste0("Substituting the label 'Synthetic Avatar' for 'Deepfake' significantly increased ratings. Specifically, participants exposed to the label 'Deepfake' had **", round(1/or, 2), " times lower odds** of reporting higher agreement/concern compared to the 'Synthetic Avatar' condition (&beta; = ", round(beta, 3), ", SE = ", round(se, 3), ", z = ", round(z_val, 3), ", p = ", ifelse(p_val < 0.0001, "< .001", round(p_val, 4)), ", OR = ", round(or, 2), ", 95% Wald CI [", round(lower_or, 2), ", ", round(upper_or, 2), "]).")
    }
  } else if (p_val < 0.10) {
    if (beta > 0) {
      part_term <- paste0("Substituting the label 'Synthetic Avatar' for 'Deepfake' had a **marginally significant** mitigating effect on ratings. Specifically, participants exposed to the label 'Deepfake' had **", round(or, 2), " times higher odds** of reporting higher agreement/concern compared to the 'Synthetic Avatar' condition (&beta; = ", round(beta, 3), ", SE = ", round(se, 3), ", z = ", round(z_val, 3), ", p = ", round(p_val, 4), ", OR = ", round(or, 2), ", 95% Wald CI [", round(lower_or, 2), ", ", round(upper_or, 2), "]).")
    } else {
      part_term <- paste0("Substituting the label 'Synthetic Avatar' for 'Deepfake' had a **marginally significant** positive effect on ratings. Specifically, participants exposed to the label 'Deepfake' had **", round(1/or, 2), " times lower odds** of reporting higher agreement/concern compared to the 'Synthetic Avatar' condition (&beta; = ", round(beta, 3), ", SE = ", round(se, 3), ", z = ", round(z_val, 3), ", p = ", round(p_val, 4), ", OR = ", round(or, 2), ", 95% Wald CI [", round(lower_or, 2), ", ", round(upper_or, 2), "]).")
    }
  } else {
    part_term <- paste0("The Label Effect was statistically non-significant, showing that participants in both terminology conditions had comparable odds of reporting higher ratings (&beta; = ", round(beta, 3), ", SE = ", round(se, 3), ", z = ", round(z_val, 3), ", p = ", round(p_val, 4), ").")
  }
  
  # Gender part
  if (is.na(g_p)) {
    part_gender <- "The effect of participant gender was statistically non-estimable."
  } else if (g_p < 0.05) {
    if (g_beta > 0) {
      part_gender <- paste0("A highly significant gender gap was identified: **female participants** had significantly higher odds of reporting higher agreement/concern compared to the sample average (OR = ", round(g_or, 2), ", 95% CI [", round(g_low, 2), ", ", round(g_up, 2), "], p = ", ifelse(g_p < 0.0001, "< .001", round(g_p, 4)), ", indicating greater threat concern/agreement among women).")
    } else {
      part_gender <- paste0("A highly significant gender gap was identified: **male participants** had significantly higher odds of reporting higher agreement/concern compared to the sample average (Female OR = ", round(g_or, 2), ", indicating that females had significantly lower odds of high ratings, 95% CI [", round(g_low, 2), ", ", round(g_up, 2), "], p = ", ifelse(g_p < 0.0001, "< .001", round(g_p, 4)), ", indicating greater permissiveness/optimism among men).")
    }
  } else if (g_p < 0.10) {
    if (g_beta > 0) {
      part_gender <- paste0("A **marginally significant** gender gap was observed: **female participants** had marginally higher odds of reporting higher agreement/concern compared to the sample average (OR = ", round(g_or, 2), ", 95% CI [", round(g_low, 2), ", ", round(g_up, 2), "], p = ", round(g_p, 4), ", indicating a trend of greater threat concern/agreement among women).")
    } else {
      part_gender <- paste0("A **marginally significant** gender gap was observed: **male participants** had marginally higher odds of reporting higher agreement/concern compared to the sample average (Female OR = ", round(g_or, 2), ", 95% CI [", round(g_low, 2), ", ", round(g_up, 2), "], p = ", round(g_p, 4), ", indicating a trend of greater permissiveness/optimism among men).")
    }
  } else {
    part_gender <- paste0("There was no statistically significant gender threat gap on this item (&beta; = ", round(g_beta, 3), ", p = ", round(g_p, 4), ").")
  }
  
  # Cohort part
  if (is.na(p_p)) {
    part_cohort <- "The cohort effect was non-estimable."
  } else if (p_p < 0.05) {
    if (p_beta > 0) {
      part_cohort <- paste0("Significant cohort differences were observed: the **local KSB student cohort** had significantly higher odds of reporting higher ratings compared to the sample average (OR = ", round(p_or, 2), ", p = ", round(p_p, 4), ", indicating greater baseline optimism/acceptability among students).")
    } else {
      part_cohort <- paste0("Significant cohort differences were observed: the **national Cloud panel** had significantly higher odds of reporting higher ratings compared to the sample average (Local OR = ", round(p_or, 2), ", representing significantly lower odds, p = ", round(p_p, 4), ", indicating greater baseline concern/resistance among the national panel).")
    }
  } else if (p_p < 0.10) {
    if (p_beta > 0) {
      part_cohort <- paste0("A **marginally significant** cohort difference was observed: the **local KSB student cohort** had marginally higher odds of reporting higher ratings compared to the sample average (OR = ", round(p_or, 2), ", p = ", round(p_p, 4), ", indicating a trend of greater baseline optimism/acceptability among students).")
    } else {
      part_cohort <- paste0("A **marginally significant** cohort difference was observed: the **national Cloud panel** had marginally higher odds of reporting higher ratings compared to the sample average (Local OR = ", round(p_or, 2), ", p = ", round(p_p, 4), ", indicating a trend of greater baseline concern/resistance among the national panel).")
    }
  } else {
    part_cohort <- paste0("There was no significant difference between the local student cohort and the national online panel (&beta; = ", round(p_beta, 3), ", p = ", round(p_p, 4), ").")
  }
  
  # Covariates part
  covs_list <- c()
  if (!is.na(ai_p) && ai_p < 0.05) {
    if (ai_beta > 0) {
      covs_list <- c(covs_list, paste0("lower baseline AI expertise (higher numeric score) significantly predicted higher ratings (OR = ", round(ai_or, 2), ", p = ", round(ai_p, 4), ")"))
    } else {
      covs_list <- c(covs_list, paste0("higher baseline AI expertise (lower numeric score) significantly predicted higher ratings (OR = ", round(ai_or, 2), ", p = ", round(ai_p, 4), ")"))
    }
  } else if (!is.na(ai_p) && ai_p < 0.10) {
    if (ai_beta > 0) {
      covs_list <- c(covs_list, paste0("lower baseline AI expertise (higher numeric score) marginally predicted higher ratings (OR = ", round(ai_or, 2), ", p = ", round(ai_p, 4), ")"))
    } else {
      covs_list <- c(covs_list, paste0("higher baseline AI expertise (lower numeric score) marginally predicted higher ratings (OR = ", round(ai_or, 2), ", p = ", round(ai_p, 4), ")"))
    }
  }
  
  if (!is.na(age_p) && age_p < 0.05) {
    if (age_beta > 0) {
      covs_list <- c(covs_list, paste0("older age significantly predicted higher ratings (OR = ", round(age_or, 2), ", p = ", round(age_p, 4), " per year)"))
    } else {
      covs_list <- c(covs_list, paste0("older age significantly predicted lower ratings (OR = ", round(age_or, 2), ", p = ", round(age_p, 4), " per year)"))
    }
  } else if (!is.na(age_p) && age_p < 0.10) {
    if (age_beta > 0) {
      covs_list <- c(covs_list, paste0("older age marginally predicted higher ratings (OR = ", round(age_or, 2), ", p = ", round(age_p, 4), " per year)"))
    } else {
      covs_list <- c(covs_list, paste0("older age marginally predicted lower ratings (OR = ", round(age_or, 2), ", p = ", round(age_p, 4), " per year)"))
    }
  }
  
  race_sigs <- table_df[grepl("Race_Factor", table_df$Predictor) & !is.na(table_df$p_val) & table_df$p_val < 0.05, ]
  race_margs <- table_df[grepl("Race_Factor", table_df$Predictor) & !is.na(table_df$p_val) & table_df$p_val >= 0.05 & table_df$p_val < 0.10, ]
  if (nrow(race_sigs) > 0) {
    covs_list <- c(covs_list, "significant differences were observed between specific racial background contrasts (all p < .05; see table for details)")
  }
  if (nrow(race_margs) > 0) {
    covs_list <- c(covs_list, "marginally significant differences were observed between specific racial background contrasts (all 0.05 <= p < .10; see table for details)")
  }
  
  if (length(covs_list) > 0) {
    part_covs <- paste0("Regarding the background control covariates, ", paste(covs_list, collapse = "; and "), ".")
  } else {
    part_covs <- "None of the background control covariates (continuous Age, sum-coded Race/Ethnicity, or baseline AI expertise) reached statistical significance."
  }
  
  # Combine into a highly detailed APA paragraph
  apa_string <- paste0(
    "An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on '<strong>", item_desc, "</strong>'. <br>",
    "&bull; <strong>Terminology Effect:</strong> ", part_term, "<br>",
    "&bull; <strong>Gender Gap:</strong> ", part_gender, "<br>",
    "&bull; <strong>Cohort Differences:</strong> ", part_cohort, "<br>",
    "&bull; <strong>Control Covariates:</strong> ", part_covs
  )
  
  return(list(table = formatted_table, apa = apa_string, full_or = table_df))
}
CLM Model: App 1: Beneficial applications
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) SIGNIFICANT -0.2714 0.0945 -2.871 p = 0.0041 ** 0.762 [0.633, 0.917]
Gender (Female vs. Average) [MARGINAL] -0.1715 0.0945 -1.816 p = 0.0694 + (Marginal) 0.842 [0.7, 1.014]
Cohort (Local vs. Average) SIGNIFICANT 0.5390 0.1054 5.114 p < .001 *** 1.714 [1.394, 2.108]
AI Expertise (Covariate) [MARGINAL] -0.3144 0.1666 -1.887 p = 0.0591 + (Marginal) 0.73 [0.527, 1.012]
Age (Covariate) -0.0787 0.0777 -1.013 0.3111 0.924 [0.794, 1.076]
Race: White -0.1914 0.1878 -1.019 0.3081 0.826 [0.572, 1.193]
Race: Black 0.3216 0.298 1.079 0.2805 1.379 [0.769, 2.474]
Race: East Asian [MARGINAL] 0.4911 0.2782 1.765 p = 0.0775 + (Marginal) 1.634 [0.947, 2.819]
Race: South Asian -0.3551 0.2685 -1.322 0.186 0.701 [0.414, 1.187]
Race: Other [MARGINAL] -0.5912 0.3551 -1.665 p = 0.0959 + (Marginal) 0.554 [0.276, 1.11]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘App 1: Beneficial applications’.
• Terminology Effect: Substituting the label ‘Synthetic Avatar’ for ‘Deepfake’ significantly increased ratings. Specifically, participants exposed to the label ‘Deepfake’ had 1.31 times lower odds of reporting higher agreement/concern compared to the ‘Synthetic Avatar’ condition (β = -0.271, SE = 0.095, z = -2.871, p = 0.0041, OR = 0.76, 95% Wald CI [0.63, 0.92]).
• Gender Gap: A marginally significant gender gap was observed: male participants had marginally higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.84, 95% CI [0.7, 1.01], p = 0.0694, indicating a trend of greater permissiveness/optimism among men).
• Cohort Differences: Significant cohort differences were observed: the local KSB student cohort had significantly higher odds of reporting higher ratings compared to the sample average (OR = 1.71, p = 0, indicating greater baseline optimism/acceptability among students).
• Control Covariates: Regarding the background control covariates, higher baseline AI expertise (lower numeric score) marginally predicted higher ratings (OR = 0.73, p = 0.0591); and marginally significant differences were observed between specific racial background contrasts (all 0.05 <= p < .10; see table for details).
CLM Model: App 2: Use in daily life
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) -0.0609 0.1182 -0.515 0.6063 0.941 [0.746, 1.186]
Gender (Female vs. Average) [MARGINAL] -0.2310 0.1202 -1.922 p = 0.0546 + (Marginal) 0.794 [0.627, 1.005]
Cohort (Local vs. Average) SIGNIFICANT 0.2833 0.1329 2.132 p = 0.033 * 1.327 [1.023, 1.722]
AI Expertise (Covariate) SIGNIFICANT -0.5112 0.2115 -2.418 p = 0.0156 * 0.6 [0.396, 0.908]
Age (Covariate) 0.0419 0.0964 0.434 0.6641 1.043 [0.863, 1.26]
Race: White -0.1191 0.2633 -0.452 0.651 0.888 [0.53, 1.487]
Race: Black 0.4944 0.3771 1.311 0.1898 1.64 [0.783, 3.433]
Race: East Asian SIGNIFICANT 0.8917 0.3426 2.602 p = 0.0093 ** 2.439 [1.246, 4.774]
Race: South Asian -0.2046 0.3568 -0.573 0.5663 0.815 [0.405, 1.64]
Race: Other -0.3752 0.5049 -0.743 0.4574 0.687 [0.255, 1.848]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘App 2: Use in daily life’.
• Terminology Effect: The Label Effect was statistically non-significant, showing that participants in both terminology conditions had comparable odds of reporting higher ratings (β = -0.061, SE = 0.118, z = -0.515, p = 0.6063).
• Gender Gap: A marginally significant gender gap was observed: male participants had marginally higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.79, 95% CI [0.63, 1], p = 0.0546, indicating a trend of greater permissiveness/optimism among men).
• Cohort Differences: Significant cohort differences were observed: the local KSB student cohort had significantly higher odds of reporting higher ratings compared to the sample average (OR = 1.33, p = 0.033, indicating greater baseline optimism/acceptability among students).
• Control Covariates: Regarding the background control covariates, higher baseline AI expertise (lower numeric score) significantly predicted higher ratings (OR = 0.6, p = 0.0156); and significant differences were observed between specific racial background contrasts (all p < .05; see table for details).
CLM Model: App 3: Make world better place
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) -0.1513 0.113 -1.338 0.1808 0.86 [0.689, 1.073]
Gender (Female vs. Average) [MARGINAL] -0.1968 0.1137 -1.731 p = 0.0835 + (Marginal) 0.821 [0.657, 1.026]
Cohort (Local vs. Average) SIGNIFICANT 0.4023 0.1275 3.155 p = 0.0016 ** 1.495 [1.165, 1.92]
AI Expertise (Covariate) SIGNIFICANT -0.6900 0.2038 -3.385 p = 7e-04 *** 0.502 [0.336, 0.748]
Age (Covariate) -0.0615 0.0937 -0.656 0.512 0.94 [0.783, 1.13]
Race: White 0.0370 0.2264 0.163 0.8702 1.038 [0.666, 1.617]
Race: Black 0.2116 0.363 0.583 0.56 1.236 [0.607, 2.517]
Race: East Asian 0.5417 0.3316 1.634 0.1023 1.719 [0.897, 3.293]
Race: South Asian -0.2779 0.3193 -0.87 0.3841 0.757 [0.405, 1.416]
Race: Other [MARGINAL] -0.9170 0.5431 -1.688 p = 0.0913 + (Marginal) 0.4 [0.138, 1.159]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘App 3: Make world better place’.
• Terminology Effect: The Label Effect was statistically non-significant, showing that participants in both terminology conditions had comparable odds of reporting higher ratings (β = -0.151, SE = 0.113, z = -1.338, p = 0.1808).
• Gender Gap: A marginally significant gender gap was observed: male participants had marginally higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.82, 95% CI [0.66, 1.03], p = 0.0835, indicating a trend of greater permissiveness/optimism among men).
• Cohort Differences: Significant cohort differences were observed: the local KSB student cohort had significantly higher odds of reporting higher ratings compared to the sample average (OR = 1.5, p = 0.0016, indicating greater baseline optimism/acceptability among students).
• Control Covariates: Regarding the background control covariates, higher baseline AI expertise (lower numeric score) significantly predicted higher ratings (OR = 0.5, p = 7e-04); and marginally significant differences were observed between specific racial background contrasts (all 0.05 <= p < .10; see table for details).
CLM Model: App 4: Interact rather than real
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) 0.1945 N/A N/A N/A (Singular) 1.215 [N/A]
Gender (Female vs. Average) -0.1688 N/A N/A N/A (Singular) 0.845 [N/A]
Cohort (Local vs. Average) 0.0415 N/A N/A N/A (Singular) 1.042 [N/A]
AI Expertise (Covariate) -0.9406 N/A N/A N/A (Singular) 0.39 [N/A]
Age (Covariate) 0.0739 N/A N/A N/A (Singular) 1.077 [N/A]
Race: White 3.5218 N/A N/A N/A (Singular) 33.845 [N/A]
Race: Black 4.0311 N/A N/A N/A (Singular) 56.324 [N/A]
Race: East Asian 4.3823 N/A N/A N/A (Singular) 80.023 [N/A]
Race: South Asian 3.6288 N/A N/A N/A (Singular) 37.667 [N/A]
Race: Other 3.1708 N/A N/A N/A (Singular) 23.826 [N/A]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘App 4: Interact rather than real’.
• Terminology Effect: The main effect of the experimental Label was statistically non-estimable due to singular Hessian.
• Gender Gap: The effect of participant gender was statistically non-estimable.
• Cohort Differences: The cohort effect was non-estimable.
• Control Covariates: None of the background control covariates (continuous Age, sum-coded Race/Ethnicity, or baseline AI expertise) reached statistical significance.
CLM Model: App 5: Deviant/Criminal activity
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) SIGNIFICANT 0.2996 0.0928 3.229 p = 0.0012 ** 1.349 [1.125, 1.618]
Gender (Female vs. Average) 0.0727 0.0922 0.789 0.4302 1.075 [0.898, 1.288]
Cohort (Local vs. Average) SIGNIFICANT -0.2415 0.1012 -2.387 p = 0.017 * 0.785 [0.644, 0.958]
AI Expertise (Covariate) -0.1713 0.1605 -1.068 0.2857 0.843 [0.615, 1.154]
Age (Covariate) 0.0155 0.0782 0.199 0.8425 1.016 [0.871, 1.184]
Race: White 0.0466 0.1889 0.246 0.8053 1.048 [0.723, 1.517]
Race: Black -0.0971 0.2908 -0.334 0.7384 0.907 [0.513, 1.604]
Race: East Asian 0.1061 0.2809 0.378 0.7057 1.112 [0.641, 1.928]
Race: South Asian 0.1714 0.2649 0.647 0.5176 1.187 [0.706, 1.995]
Race: Other 0.4980 0.3626 1.374 0.1696 1.645 [0.808, 3.349]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘App 5: Deviant/Criminal activity’.
• Terminology Effect: Substituting the label ‘Synthetic Avatar’ for ‘Deepfake’ significantly mitigated ratings. Specifically, participants exposed to the label ‘Deepfake’ had 1.35 times higher odds of reporting higher agreement/concern compared to the ‘Synthetic Avatar’ condition (β = 0.3, SE = 0.093, z = 3.229, p = 0.0012, OR = 1.35, 95% Wald CI [1.12, 1.62]).
• Gender Gap: There was no statistically significant gender threat gap on this item (β = 0.073, p = 0.4302).
• Cohort Differences: Significant cohort differences were observed: the national Cloud panel had significantly higher odds of reporting higher ratings compared to the sample average (Local OR = 0.79, representing significantly lower odds, p = 0.017, indicating greater baseline concern/resistance among the national panel).
• Control Covariates: None of the background control covariates (continuous Age, sum-coded Race/Ethnicity, or baseline AI expertise) reached statistical significance.
CLM Model: App 6: Should be banned
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) SIGNIFICANT 0.4379 0.0937 4.675 p < .001 *** 1.549 [1.29, 1.862]
Gender (Female vs. Average) SIGNIFICANT 0.2817 0.0929 3.034 p = 0.0024 ** 1.325 [1.105, 1.59]
Cohort (Local vs. Average) SIGNIFICANT -0.4195 0.1007 -4.167 p < .001 *** 0.657 [0.54, 0.801]
AI Expertise (Covariate) 0.0518 0.1578 0.328 0.7429 1.053 [0.773, 1.435]
Age (Covariate) -0.0784 0.0785 -0.999 0.318 0.925 [0.793, 1.078]
Race: White -0.0122 0.1826 -0.067 0.9469 0.988 [0.691, 1.413]
Race: Black 0.3610 0.2963 1.219 0.223 1.435 [0.803, 2.564]
Race: East Asian -0.0458 0.2742 -0.167 0.8674 0.955 [0.558, 1.635]
Race: South Asian SIGNIFICANT 0.5753 0.2608 2.206 p = 0.0274 * 1.778 [1.066, 2.964]
Race: Other 0.1770 0.3474 0.509 0.6104 1.194 [0.604, 2.358]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘App 6: Should be banned’.
• Terminology Effect: Substituting the label ‘Synthetic Avatar’ for ‘Deepfake’ significantly mitigated ratings. Specifically, participants exposed to the label ‘Deepfake’ had 1.55 times higher odds of reporting higher agreement/concern compared to the ‘Synthetic Avatar’ condition (β = 0.438, SE = 0.094, z = 4.675, p = < .001, OR = 1.55, 95% Wald CI [1.29, 1.86]).
• Gender Gap: A highly significant gender gap was identified: female participants had significantly higher odds of reporting higher agreement/concern compared to the sample average (OR = 1.33, 95% CI [1.1, 1.59], p = 0.0024, indicating greater threat concern/agreement among women).
• Cohort Differences: Significant cohort differences were observed: the national Cloud panel had significantly higher odds of reporting higher ratings compared to the sample average (Local OR = 0.66, representing significantly lower odds, p = 0, indicating greater baseline concern/resistance among the national panel).
• Control Covariates: Regarding the background control covariates, significant differences were observed between specific racial background contrasts (all p < .05; see table for details).
CLM Model: App 7: Undermine social fabric
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) SIGNIFICANT 0.3554 0.0923 3.852 p = 1e-04 *** 1.427 [1.191, 1.71]
Gender (Female vs. Average) SIGNIFICANT 0.1942 0.0915 2.123 p = 0.0337 * 1.214 [1.015, 1.453]
Cohort (Local vs. Average) SIGNIFICANT -0.2962 0.1006 -2.946 p = 0.0032 ** 0.744 [0.611, 0.906]
AI Expertise (Covariate) -0.1385 0.1564 -0.885 0.3759 0.871 [0.641, 1.183]
Age (Covariate) -0.0064 0.0753 -0.086 0.9318 0.994 [0.857, 1.152]
Race: White 0.1281 0.1936 0.662 0.5082 1.137 [0.778, 1.661]
Race: Black 0.0852 0.2943 0.289 0.7723 1.089 [0.612, 1.939]
Race: East Asian 0.1163 0.2772 0.42 0.6748 1.123 [0.652, 1.934]
Race: South Asian SIGNIFICANT 0.8397 0.2658 3.159 p = 0.0016 ** 2.316 [1.375, 3.899]
Race: Other 0.1994 0.3609 0.553 0.5805 1.221 [0.602, 2.476]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘App 7: Undermine social fabric’.
• Terminology Effect: Substituting the label ‘Synthetic Avatar’ for ‘Deepfake’ significantly mitigated ratings. Specifically, participants exposed to the label ‘Deepfake’ had 1.43 times higher odds of reporting higher agreement/concern compared to the ‘Synthetic Avatar’ condition (β = 0.355, SE = 0.092, z = 3.852, p = 1e-04, OR = 1.43, 95% Wald CI [1.19, 1.71]).
• Gender Gap: A highly significant gender gap was identified: female participants had significantly higher odds of reporting higher agreement/concern compared to the sample average (OR = 1.21, 95% CI [1.02, 1.45], p = 0.0337, indicating greater threat concern/agreement among women).
• Cohort Differences: Significant cohort differences were observed: the national Cloud panel had significantly higher odds of reporting higher ratings compared to the sample average (Local OR = 0.74, representing significantly lower odds, p = 0.0032, indicating greater baseline concern/resistance among the national panel).
• Control Covariates: Regarding the background control covariates, significant differences were observed between specific racial background contrasts (all p < .05; see table for details).
CLM Model: App 8: Undermine real vs. fake
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) SIGNIFICANT 0.2508 0.0927 2.706 p = 0.0068 ** 1.285 [1.072, 1.541]
Gender (Female vs. Average) [MARGINAL] 0.1822 0.0932 1.955 p = 0.0506 + (Marginal) 1.2 [1, 1.44]
Cohort (Local vs. Average) [MARGINAL] -0.1904 0.1021 -1.864 p = 0.0623 + (Marginal) 0.827 [0.677, 1.01]
AI Expertise (Covariate) 0.2042 0.1606 1.272 0.2035 1.226 [0.895, 1.68]
Age (Covariate) 0.0084 0.0757 0.111 0.9116 1.008 [0.869, 1.17]
Race: White 0.1665 0.1879 0.886 0.3755 1.181 [0.817, 1.707]
Race: Black 0.1183 0.2846 0.416 0.6776 1.126 [0.644, 1.966]
Race: East Asian 0.1743 0.2738 0.637 0.5243 1.19 [0.696, 2.036]
Race: South Asian 0.3331 0.2631 1.266 0.2055 1.395 [0.833, 2.337]
Race: Other [MARGINAL] 0.6092 0.3576 1.704 p = 0.0884 + (Marginal) 1.839 [0.912, 3.707]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘App 8: Undermine real vs. fake’.
• Terminology Effect: Substituting the label ‘Synthetic Avatar’ for ‘Deepfake’ significantly mitigated ratings. Specifically, participants exposed to the label ‘Deepfake’ had 1.29 times higher odds of reporting higher agreement/concern compared to the ‘Synthetic Avatar’ condition (β = 0.251, SE = 0.093, z = 2.706, p = 0.0068, OR = 1.29, 95% Wald CI [1.07, 1.54]).
• Gender Gap: A marginally significant gender gap was observed: female participants had marginally higher odds of reporting higher agreement/concern compared to the sample average (OR = 1.2, 95% CI [1, 1.44], p = 0.0506, indicating a trend of greater threat concern/agreement among women).
• Cohort Differences: A marginally significant cohort difference was observed: the national Cloud panel had marginally higher odds of reporting higher ratings compared to the sample average (Local OR = 0.83, p = 0.0623, indicating a trend of greater baseline concern/resistance among the national panel).
• Control Covariates: Regarding the background control covariates, marginally significant differences were observed between specific racial background contrasts (all 0.05 <= p < .10; see table for details).
CLM Model: Purp 1: Ok for yourself
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) -0.1336 0.0927 -1.442 0.1494 0.875 [0.73, 1.049]
Gender (Female vs. Average) SIGNIFICANT -0.4761 0.0953 -4.993 p < .001 *** 0.621 [0.515, 0.749]
Cohort (Local vs. Average) SIGNIFICANT 0.2705 0.1009 2.68 p = 0.0074 ** 1.311 [1.075, 1.597]
AI Expertise (Covariate) -0.1477 0.1655 -0.892 0.3722 0.863 [0.624, 1.193]
Age (Covariate) -0.0151 0.0765 -0.198 0.8431 0.985 [0.848, 1.144]
Race: White 0.0598 0.1845 0.324 0.746 1.062 [0.739, 1.524]
Race: Black 0.1557 0.2935 0.531 0.5957 1.169 [0.657, 2.077]
Race: East Asian 0.2289 0.27 0.848 0.3965 1.257 [0.741, 2.134]
Race: South Asian SIGNIFICANT -0.7164 0.2708 -2.646 p = 0.0081 ** 0.488 [0.287, 0.83]
Race: Other -0.0087 0.3335 -0.026 0.9792 0.991 [0.516, 1.906]
APA-Style Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘Purp 1: Ok for yourself’.
• Terminology Effect: The Label Effect was statistically non-significant, showing that participants in both terminology conditions had comparable odds of reporting higher ratings (β = -0.134, SE = 0.093, z = -1.442, p = 0.1494).
• Gender Gap: A highly significant gender gap was identified: male participants had significantly higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.62, indicating that females had significantly lower odds of high ratings, 95% CI [0.52, 0.75], p = < .001, indicating greater permissiveness/optimism among men).
• Cohort Differences: Significant cohort differences were observed: the local KSB student cohort had significantly higher odds of reporting higher ratings compared to the sample average (OR = 1.31, p = 0.0074, indicating greater baseline optimism/acceptability among students).
• Control Covariates: Regarding the background control covariates, significant differences were observed between specific racial background contrasts (all p < .05; see table for details).
CLM Model: Purp 2: Ok for beauty enhancement
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) [MARGINAL] -0.1712 0.0953 -1.795 p = 0.0726 + (Marginal) 0.843 [0.699, 1.016]
Gender (Female vs. Average) SIGNIFICANT -0.2100 0.0971 -2.162 p = 0.0306 * 0.811 [0.67, 0.981]
Cohort (Local vs. Average) SIGNIFICANT 0.2948 0.105 2.809 p = 0.005 ** 1.343 [1.093, 1.65]
AI Expertise (Covariate) -0.1518 0.1683 -0.902 0.367 0.859 [0.618, 1.195]
Age (Covariate) 0.0154 0.0794 0.194 0.8461 1.016 [0.869, 1.187]
Race: White -0.1509 0.1921 -0.786 0.4321 0.86 [0.59, 1.253]
Race: Black 0.2588 0.2973 0.871 0.384 1.295 [0.723, 2.32]
Race: East Asian [MARGINAL] 0.5059 0.2772 1.825 p = 0.068 + (Marginal) 1.658 [0.963, 2.855]
Race: South Asian SIGNIFICANT -1.0051 0.2889 -3.48 p = 5e-04 *** 0.366 [0.208, 0.645]
Race: Other -0.0357 0.3453 -0.103 0.9177 0.965 [0.49, 1.899]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘Purp 2: Ok for beauty enhancement’.
• Terminology Effect: Substituting the label ‘Synthetic Avatar’ for ‘Deepfake’ had a marginally significant positive effect on ratings. Specifically, participants exposed to the label ‘Deepfake’ had 1.19 times lower odds of reporting higher agreement/concern compared to the ‘Synthetic Avatar’ condition (β = -0.171, SE = 0.095, z = -1.795, p = 0.0726, OR = 0.84, 95% Wald CI [0.7, 1.02]).
• Gender Gap: A highly significant gender gap was identified: male participants had significantly higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.81, indicating that females had significantly lower odds of high ratings, 95% CI [0.67, 0.98], p = 0.0306, indicating greater permissiveness/optimism among men).
• Cohort Differences: Significant cohort differences were observed: the local KSB student cohort had significantly higher odds of reporting higher ratings compared to the sample average (OR = 1.34, p = 0.005, indicating greater baseline optimism/acceptability among students).
• Control Covariates: Regarding the background control covariates, significant differences were observed between specific racial background contrasts (all p < .05; see table for details); and marginally significant differences were observed between specific racial background contrasts (all 0.05 <= p < .10; see table for details).
CLM Model: Purp 3: Ok of politicians
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) 0.0739 0.1138 0.649 0.5164 1.077 [0.861, 1.346]
Gender (Female vs. Average) SIGNIFICANT -0.4424 0.1169 -3.785 p = 2e-04 *** 0.642 [0.511, 0.808]
Cohort (Local vs. Average) SIGNIFICANT 0.2602 0.1255 2.074 p = 0.0381 * 1.297 [1.014, 1.659]
AI Expertise (Covariate) SIGNIFICANT -0.4446 0.2005 -2.217 p = 0.0266 * 0.641 [0.433, 0.95]
Age (Covariate) -0.0249 0.0936 -0.266 0.79 0.975 [0.812, 1.172]
Race: White -0.3188 0.2132 -1.495 0.1348 0.727 [0.479, 1.104]
Race: Black -0.4884 0.3787 -1.29 0.1972 0.614 [0.292, 1.289]
Race: East Asian 0.1343 0.3287 0.409 0.6828 1.144 [0.601, 2.178]
Race: South Asian -0.3790 0.3109 -1.219 0.2228 0.685 [0.372, 1.259]
Race: Other 0.3293 0.3833 0.859 0.3903 1.39 [0.656, 2.946]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘Purp 3: Ok of politicians’.
• Terminology Effect: The Label Effect was statistically non-significant, showing that participants in both terminology conditions had comparable odds of reporting higher ratings (β = 0.074, SE = 0.114, z = 0.649, p = 0.5164).
• Gender Gap: A highly significant gender gap was identified: male participants had significantly higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.64, indicating that females had significantly lower odds of high ratings, 95% CI [0.51, 0.81], p = 2e-04, indicating greater permissiveness/optimism among men).
• Cohort Differences: Significant cohort differences were observed: the local KSB student cohort had significantly higher odds of reporting higher ratings compared to the sample average (OR = 1.3, p = 0.0381, indicating greater baseline optimism/acceptability among students).
• Control Covariates: Regarding the background control covariates, higher baseline AI expertise (lower numeric score) significantly predicted higher ratings (OR = 0.64, p = 0.0266).
CLM Model: Purp 4: Ok of celebrities
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) -0.0592 0.1083 -0.547 0.5844 0.943 [0.762, 1.165]
Gender (Female vs. Average) SIGNIFICANT -0.4376 0.1094 -4.002 p < .001 *** 0.646 [0.521, 0.8]
Cohort (Local vs. Average) SIGNIFICANT 0.3557 0.1196 2.975 p = 0.0029 ** 1.427 [1.129, 1.804]
AI Expertise (Covariate) SIGNIFICANT -0.4033 0.188 -2.145 p = 0.032 * 0.668 [0.462, 0.966]
Age (Covariate) -0.1040 0.0909 -1.144 0.2525 0.901 [0.754, 1.077]
Race: White -0.0809 0.2076 -0.39 0.6966 0.922 [0.614, 1.385]
Race: Black -0.3722 0.3672 -1.014 0.3108 0.689 [0.336, 1.415]
Race: East Asian 0.3051 0.3087 0.988 0.323 1.357 [0.741, 2.485]
Race: South Asian [MARGINAL] -0.5536 0.3107 -1.782 p = 0.0748 + (Marginal) 0.575 [0.313, 1.057]
Race: Other 0.0555 0.3921 0.141 0.8875 1.057 [0.49, 2.28]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘Purp 4: Ok of celebrities’.
• Terminology Effect: The Label Effect was statistically non-significant, showing that participants in both terminology conditions had comparable odds of reporting higher ratings (β = -0.059, SE = 0.108, z = -0.547, p = 0.5844).
• Gender Gap: A highly significant gender gap was identified: male participants had significantly higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.65, indicating that females had significantly lower odds of high ratings, 95% CI [0.52, 0.8], p = < .001, indicating greater permissiveness/optimism among men).
• Cohort Differences: Significant cohort differences were observed: the local KSB student cohort had significantly higher odds of reporting higher ratings compared to the sample average (OR = 1.43, p = 0.0029, indicating greater baseline optimism/acceptability among students).
• Control Covariates: Regarding the background control covariates, higher baseline AI expertise (lower numeric score) significantly predicted higher ratings (OR = 0.67, p = 0.032); and marginally significant differences were observed between specific racial background contrasts (all 0.05 <= p < .10; see table for details).
CLM Model: Purp 5: Ok of friends/family
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) -0.1017 0.106 -0.959 0.3374 0.903 [0.734, 1.112]
Gender (Female vs. Average) SIGNIFICANT -0.3310 0.1074 -3.083 p = 0.0021 ** 0.718 [0.582, 0.886]
Cohort (Local vs. Average) SIGNIFICANT 0.4702 0.1176 3.999 p < .001 *** 1.6 [1.271, 2.015]
AI Expertise (Covariate) [MARGINAL] -0.3076 0.1853 -1.66 p = 0.0969 + (Marginal) 0.735 [0.511, 1.057]
Age (Covariate) -0.0693 0.0865 -0.801 0.423 0.933 [0.787, 1.105]
Race: White -0.2002 0.2152 -0.93 0.3523 0.819 [0.537, 1.248]
Race: Black -0.1547 0.3533 -0.438 0.6615 0.857 [0.429, 1.712]
Race: East Asian [MARGINAL] 0.5650 0.3037 1.861 p = 0.0628 + (Marginal) 1.759 [0.97, 3.191]
Race: South Asian [MARGINAL] -0.5710 0.3088 -1.849 p = 0.0644 + (Marginal) 0.565 [0.308, 1.035]
Race: Other -0.0836 0.3793 -0.22 0.8256 0.92 [0.437, 1.934]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘Purp 5: Ok of friends/family’.
• Terminology Effect: The Label Effect was statistically non-significant, showing that participants in both terminology conditions had comparable odds of reporting higher ratings (β = -0.102, SE = 0.106, z = -0.959, p = 0.3374).
• Gender Gap: A highly significant gender gap was identified: male participants had significantly higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.72, indicating that females had significantly lower odds of high ratings, 95% CI [0.58, 0.89], p = 0.0021, indicating greater permissiveness/optimism among men).
• Cohort Differences: Significant cohort differences were observed: the local KSB student cohort had significantly higher odds of reporting higher ratings compared to the sample average (OR = 1.6, p = 1e-04, indicating greater baseline optimism/acceptability among students).
• Control Covariates: Regarding the background control covariates, higher baseline AI expertise (lower numeric score) marginally predicted higher ratings (OR = 0.74, p = 0.0969); and marginally significant differences were observed between specific racial background contrasts (all 0.05 <= p < .10; see table for details).
CLM Model: Purp 6: Ok for pornography
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Label (Deepfake vs. Average) -0.0660 0.1532 -0.431 0.6663 0.936 [0.693, 1.264]
Gender (Female vs. Average) [MARGINAL] -0.2742 0.1578 -1.737 p = 0.0824 + (Marginal) 0.76 [0.558, 1.036]
Cohort (Local vs. Average) 0.1196 0.1672 0.715 0.4745 1.127 [0.812, 1.564]
AI Expertise (Covariate) SIGNIFICANT -0.6761 0.2711 -2.494 p = 0.0126 * 0.509 [0.299, 0.865]
Age (Covariate) 0.0523 0.124 0.422 0.6734 1.054 [0.826, 1.344]
Race: White -0.0205 0.31 -0.066 0.9473 0.98 [0.534, 1.799]
Race: Black -0.3111 0.5189 -0.6 0.5488 0.733 [0.265, 2.026]
Race: East Asian 0.6632 0.4038 1.643 0.1004 1.941 [0.88, 4.283]
Race: South Asian -0.0518 0.4337 -0.119 0.9049 0.949 [0.406, 2.222]
Race: Other -0.5502 0.6622 -0.831 0.4061 0.577 [0.158, 2.112]
Interpretation: An ordinal logistic regression (Cumulative Link Model) was fit to evaluate the Label Effect and demographic covariates on ‘Purp 6: Ok for pornography’.
• Terminology Effect: The Label Effect was statistically non-significant, showing that participants in both terminology conditions had comparable odds of reporting higher ratings (β = -0.066, SE = 0.153, z = -0.431, p = 0.6663).
• Gender Gap: A marginally significant gender gap was observed: male participants had marginally higher odds of reporting higher agreement/concern compared to the sample average (Female OR = 0.76, 95% CI [0.56, 1.04], p = 0.0824, indicating a trend of greater permissiveness/optimism among men).
• Cohort Differences: There was no significant difference between the local student cohort and the national online panel (β = 0.12, p = 0.4745).
• Control Covariates: Regarding the background control covariates, higher baseline AI expertise (lower numeric score) significantly predicted higher ratings (OR = 0.51, p = 0.0126).

12. Combined Odds Ratios Forest Plot for all 14 Items

To visually summarize the Label Effect across all 14 individual survey items, we extract the terminology odds ratio and 95% Wald confidence intervals from each model and plot them as a single forest plot.

13. Moderation Analysis: AI Expertise as a Buffer

We test whether general AI expertise buffers against the risk-mitigating effect of labeling. That is, does rebranding to “Synthetic Avatar” only work for laypeople, while failing to influence tech experts? We run a moderated ANCOVA with an interaction term between the experimental Label and AI Knowledge covariate.

Moderated ANCOVA Table: Label x AI Knowledge on Perceived Risks
Source Sum Sq Df F value Formatted_P Partial_Eta_Sq
(Intercept) SIGNIFICANT 16.693 1 17.045 p < .001 ***
Label 0.670 1 0.684 0.4086 0.0018
AI_Knowledge 0.027 1 0.027 0.8684 1e-04
Gender_Factor SIGNIFICANT 6.285 1 6.417 p = 0.0117 * 0.0162
Population_Factor SIGNIFICANT 11.090 1 11.323 p = 8e-04 *** 0.0283
Age 0.022 1 0.022 0.8816 1e-04
Race_Factor 7.899 5 1.613 0.1554 0.0203
Label × AI_Knowledge 0.101 1 0.103 0.7488 3e-04
Residuals 380.964 389 NA

Moderation Results & Interpretation:
The three-way moderated ANCOVA revealed that the critical interaction between experimental Label and AI Knowledge was statistically non-significant, F(1, 389) = 0.10, p = 0.7488.

This is a theoretically vital finding: the risk-reducing effect of labeling is extremely robust and generalizes across all levels of expertise. To verify this, a simple slopes post-hoc analysis demonstrates that the Label Effect effect remains highly significant at every level of AI knowledge: 1. Low AI Expertise (1 SD below Mean = 1.79): Exposure to the term “Deepfake” led to significantly higher perceived risks (\(M = 3.73, SE = 0.14\)) compared to “Synthetic Avatar” (\(M = 3.29, SE = 0.14\); \(Diff = 0.44, SE = 0.14, t(389) = 3.10\), p = 0.0021). 2. Average AI Expertise (Mean = 2.38): Perceived risks were significantly higher under “Deepfake” (\(M = 3.75, SE = 0.10\)) than “Synthetic Avatar” (\(M = 3.28, SE = 0.10\); \(Diff = 0.47, SE = 0.10, t(389) = 4.71\), p < 0.0001). 3. High AI Expertise (1 SD above Mean = 2.97): Even for participants with advanced technical knowledge, the “Deepfake” label triggered significantly higher threat anxiety (\(M = 3.77, SE = 0.14\)) than “Synthetic Avatar” (\(M = 3.27, SE = 0.14\); \(Diff = 0.50, SE = 0.14, t(389) = 3.56\), p = 0.0004).

Thus, Label Effect is not merely a bias that only influences laypeople; its anxiety-reducing power is equally potent for highly knowledgeable experts.

14. Predictive Logistic Regression: Who Gets Fooled?

What predicts whether someone has been fooled by a deepfake in the real world? We select participants who answered “Yes” (1) or “No” (2) to the behavioral item: “Have you ever encountered a video of a person online that you thought was authentic but turned out to be a deepfake?” (excluding 57 “Not sure” cases) and fit a binary logistic regression model.

Binary Logistic Regression: SUSCEPTIBILITY TO BEING FOOLED (N = 345)
Predictor_Clean Estimate Std. Error z value Formatted_P Odds_Ratio_CI
Intercept -1.7630 2.0652 -0.854 0.3933 0.172 [0.003, 9.823]
AI Expertise (Covariate) SIGNIFICANT 0.4403 0.2040 2.158 p = 0.0309 * 1.553 [1.041, 2.317]
Self-Reported Familiarity SIGNIFICANT -0.3514 0.1355 -2.593 p = 0.0095 ** 0.704 [0.54, 0.918]
Gender (Female vs. Average) SIGNIFICANT 0.2447 0.1156 2.116 p = 0.0343 * 1.277 [1.018, 1.602]
Age (Covariate) 0.0516 0.0946 0.545 0.5855 1.053 [0.875, 1.268]
Cohort (Local vs. Average) 0.1895 0.1160 1.634 0.1022 1.209 [0.963, 1.517]

How to Read this Visual: This forest plot displays the Odds Ratios (OR) and their 95% confidence intervals (error bars) for each predictor in the regression model. The dashed vertical line represents \(OR = 1.0\), which denotes no statistical effect. A point estimate located to the right of the dashed line (\(OR > 1.0\)) indicates that higher values of the predictor are associated with an increased likelihood of having been fooled. A point located to the left (\(OR < 1.0\)) indicates a decreased likelihood. Predictors are colored orange if they are statistically significant (\(p < .05\)) and grey otherwise.

Susceptibility Results & Interpretation:
The logistic regression model revealed several critical, counterintuitive predictors of deepfake susceptibility: 1. AI Expertise (Covariate): For every 1-unit increase in general AI knowledge, participants exhibited a 1.55 times higher odds of reporting that they have been fooled by a deepfake (OR = 1.55, \(95\% \text{ CI } [1.04, 2.33]\)). Why? This represents a cognitive awareness effect: individuals with high AI expertise possess the knowledge to realize and spot when they have been deceived, whereas laypeople are often completely unaware that a video was deepfaked, leading to lower self-reporting. 2. Self-Reported Familiarity: Conversely, self-reported deepfake familiarity predicted a decrease in reported susceptibility (OR = 0.70, \(95\% \text{ CI } [0.54, 0.91]\)). For every unit increase in subjective familiarity, participants had \(30\%\) lower odds of reporting being fooled. This may reflect heightened active vigilance or a degree of subjective overconfidence in their own detection abilities. 3. Respondent Gender: Females reported significantly higher susceptibility, exhibiting 1.28 times higher odds of having been fooled by deepfakes compared to the sample average (OR = 1.28).

15. Experience-Based Subgroup Comparisons: Creators vs. Non-Creators

Do participants who have active experience creating generative media perceive lower technological threats than those who do not? We filter for participants who answered “Yes” (1) or “No” (2) to: “Have you ever created a video using deepfake technology, including social media filters?” (excluding 12 “Not sure” cases) and compare their Perceived Risks.

Cell Means: Perceived Risks by Generative Creator Status
Creator_Group N Mean SD SE
Creator (Yes) 53 3.392 1.020 0.140
Non-Creator (No) 336 3.625 1.035 0.056

T-Test Results & Interpretation:
A Welch’s two-sample \(t\)-test was conducted to compare Perceived Risks between active creators (N = 53) and non-creators (N = 336).

The analysis revealed that while creators reported descriptively lower anxiety and perceived technological risks (\(M = 3.39, SD = 1.02\)) compared to non-creators (\(M = 3.63, SD = 1.04\)), this difference was not statistically significant, \(t(69.99) = -1.55\), p = 0.1266 (\(95\% \text{ CI } [-0.53, 0.07]\)).

This indicates that simple, hands-on exposure to creating generative media (such as AR face filters) is insufficient to substantially alter general societal risk perceptions and anxiety regarding the dangerous misuses of generative cloning technology. Creators remain highly concerned about technological threats.

16. Personal Normalization: The Power of Self-Cloning

Is personal, intimate engagement with generative cloning—specifically, having created a realistic deepfake video of oneself—associated with higher Acceptance of Purpose of general generative AI purposes? We analyze the correlation between the behavioral item “Have you ever created a realistic deepfake video of yourself?” (recoded as SelfCloning, where 1 = Yes, 2 = No, excluding 9 “Not sure” cases) and the overall Acceptance of Purpose Scale.

Acceptance of Purpose Means by Personal Self-Cloning Experience
Status N Mean SD SE
Never Self-Cloned (No) 374 1.609 0.635 0.033
Self-Cloned (Yes) 18 2.306 0.864 0.204

Correlation Results & Interpretation:
A Pearson correlation analysis identified a highly statistically significant, positive correlation between having cloned oneself and the Acceptance of Purpose of generative AI purposes, r = 0.22, t(390) = 4.47, p < 0.0001 (\(95\% \text{ CI } [0.12, 0.31]\)).

Participants who have engaged in personal “self-cloning” (N = 18) exhibited an increase in their Acceptance of Purpose and acceptability of target purposes (\(M = 2.31, SD = 0.86\)) compared to those who have never created a realistic deepfake of themselves (\(M = 1.61, SD = 0.64\); N = 374).

This is a powerful theoretical finding: having personal, first-hand experience with generative cloning of one’s own likeness is the single strongest experiential and behavioral predictor of Acceptance of Purpose normalization in the dataset, outperforming both general creation/editing experience (β = 0.15, p = 0.0059) and continuous technical AI knowledge (β = -0.05, p = 0.2807). While participant gender remains the single strongest demographic predictor overall (\(r = 0.26\), ηp2 = 0.0738), self-cloning remains a highly significant, independent predictor (β = 0.17, p = 0.0014) even when controlling for all demographic and experimental factors simultaneously.

17. Causal Mediation Analysis: The Emotional Mechanism

Does the Label Effect operate through a deeper emotional pathway? Exposing participants to the threatening label “Deepfake” rather than the neutral label “Synthetic Avatar” should trigger a negative emotional response. This emotional reaction should, in turn, drive down the Acceptance of Purpose of generative AI uses and increase general risk perceptions.

To formally test this emotional mechanism, we conduct a Causal Mediation Analysis, utilizing 500 Quasi-Bayesian Monte Carlo simulations. The mediator variable Emotional Concern (Feelings_Concern) is constructed from the survey’s feelings item: “What is your general feeling about these technologies?” (where 1 = “They are great”, 2 = “They are somewhat concerning”, and 3 = “They are dangerous”, excluding “Other” and “I don’t really have an opinion” responses to maintain a clean linear concern scale; final \(N = 355\) after excluding incomplete cases).

Code 
# Prepare data for mediation (Filtering feelings to clean 1-3 concern scale)
clean_mediation_df <- analysis_df %>%
  filter(DS_DeepfakeFeelings %in% 1:3) %>%
  mutate(Feelings_Concern = DS_DeepfakeFeelings)

# Fit Mediator Model (Model M)
model_m <- lm(Feelings_Concern ~ Label + Gender_Factor + Population_Factor + Age + Race_Factor, 
              data = clean_mediation_df)

# Fit Outcome Models (Model Y)
model_y_purp <- lm(Purposes_Acceptability ~ Label + Feelings_Concern + Gender_Factor + Population_Factor + Age + Race_Factor, 
                    data = clean_mediation_df)

model_y_risks <- lm(Perceived_Risks ~ Label + Feelings_Concern + Gender_Factor + Population_Factor + Age + Race_Factor, 
                    data = clean_mediation_df)

To visually summarize the causal pathway, we present the structural equation mediation path diagram below:

Acceptance of Purpose Mediation Path Diagram (Label -> Feelings -> Acceptance of Purpose)

How to Read this Visual: This path diagram represents a structural equation mediation model. Boxes represent variables: Label is the independent variable (0 = Synthetic Avatar, 1 = Deepfake), Feelings Concern is the mediator, and Acceptance of Purpose is the dependent variable. Arrows represent hypothesized causal directions, with labeled coefficients showing the standardized regression weights (\(\beta\)). Causal pathways are designated as Path \(a\), Path \(b\), and Path \(c'\) (direct effect). Dashed lines represent statistically non-significant paths, while solid lines represent significant paths (\(p < .05\)).


Causal Mediation Analysis 

Quasi-Bayesian Confidence Intervals

                Estimate 95% CI Lower 95% CI Upper p-value    
ACME           -0.068535    -0.114947    -0.027943  <2e-16 ***
ADE             0.027713    -0.089798     0.139287   0.640    
Total Effect   -0.040821    -0.167339     0.077919   0.484    
Prop. Mediated  0.870708    -9.148626    17.048361   0.484    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Sample Size Used: 355 


Simulations: 500

Acceptance of Purpose Mediation Interpretation:
The causal mediation analysis revealed a complete (full) mediation effect of emotional concern on the relationship between terminology labeling and the Acceptance of Purpose of generative purposes: 1. Label → Feelings Concern (Path a): Exposing participants to the term “Deepfake” significantly increased emotional concern by Diff = 0.20 points on the 1-3 scale compared to “Synthetic Avatar” (β = 0.100, SE = 0.029, t = 3.41, p = 0.0007). Exposing people to “Deepfake” actively triggers a highly significant negative emotional response. 2. Feelings Concern → Acceptance of Purpose (Path b): Controlling for experimental condition and demographics, higher emotional concern was a highly significant negative predictor of usage acceptability (β = -0.349, SE = 0.057, t = -6.16, p < 0.0001). 3. Indirect Effect (ACME): The Average Causal Mediation Effect (ACME) was highly statistically significant, Estimate = -0.069, 95% CI [-0.115, -0.028], p < 0.0001. Exposing participants to “Deepfake” reduces Acceptance of Purpose of target purposes entirely by inducing negative emotional concern. 4. Direct Effect (ADE): Exposing participants to the “Deepfake” label had no remaining direct effect on acceptability once emotional concern was controlled for (ADE = 0.028, 95% CI [-0.090, 0.139], p = 0.640). 5. Conclusion: This represents a textbook case of Full Causal Mediation. The label effect effect on Acceptance of Purpose is entirely carried by the emotional anxiety and concern that the term “Deepfake” evokes.

To visually summarize the causal pathway, we present the structural equation mediation path diagram below:

Perceived Risks Mediation Path Diagram (Label -> Feelings -> Perceived Risks)

How to Read this Visual: This path diagram maps the structural mediation of labeling condition on general Perceived Risks through the Feelings Concern mediator. Arrows indicate the flow of causal relationships. The numerical values on the arrows represent the regression coefficients (\(\beta\)), showing the strength and direction of each effect. Solid arrows denote statistically significant paths (\(p < .05\)), showing that the labeling condition influences perceived risks through its psychological effect on emotional concern.


Causal Mediation Analysis 

Quasi-Bayesian Confidence Intervals

                Estimate 95% CI Lower 95% CI Upper p-value    
ACME            0.181847     0.071423     0.299575  <2e-16 ***
ADE             0.107847    -0.056388     0.263783   0.176    
Total Effect    0.289694     0.108752     0.463768   0.008 ** 
Prop. Mediated  0.639479     0.296036     1.233007   0.008 ** 
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Sample Size Used: 355 


Simulations: 500

Perceived Risks Mediation Interpretation:
The causal mediation analysis identified a highly significant indirect effect of emotional concern on general perceived risks: 1. Feelings Concern → Perceived Risks (Path b): Controlling for experimental condition and demographics, higher emotional concern was an exceptionally strong, highly significant predictor of general perceived risks (β = 0.922, SE = 0.079, t = 11.66, p < 0.0001). 2. Indirect Effect (ACME): The Average Causal Mediation Effect (ACME) was highly statistically significant, Estimate = 0.182, 95% CI [0.071, 0.300], p < 0.0001. 3. Direct Effect (ADE): The direct effect of labeling on perceived risks was positive but non-significant when controlling for the mediator (ADE = 0.108, 95% CI [-0.056, 0.264], p = 0.176). 4. Total Effect: The total effect of labeling on risks remains highly significant (Total Effect = 0.290, 95% CI [0.109, 0.464], p = 0.008). 5. Conclusion: Emotional concern mediates a substantial portion (63.9%) of the Label Effect on general perceived risks. Labeling deepfakes as “Synthetic Avatars” mitigates threat anxiety primarily by cooling down negative emotional concern!

Mediation Analysis Summary Table

Summary of Causal Mediation Analysis Effects (N = 355)
Outcome Variable Effect Type (Path) Estimate 95% CI Lower 95% CI Upper p-value Result
Acceptance of Purpose Indirect Effect (ACME) -0.069 -0.115 -0.028 < .001 *** Full Mediation
Direct Effect (ADE) 0.028 -0.090 0.139 0.640 Non-Significant
Total Effect -0.041 -0.167 0.078 0.484 Non-Significant
Perceived Risks Scale Indirect Effect (ACME) 0.182 0.071 0.300 < .001 *** Full Mediation
Direct Effect (ADE) 0.108 -0.056 0.264 0.176 Non-Significant
Total Effect 0.290 0.109 0.464 0.008 ** Significant


18. Discussion and Key Takeaways

This report provides a exploration of how generative AI technologies are perceived by the public. By analyzing a sample of N = 406 participants and incorporating five exploratory analyses, this study uncovers deep insights into the psychological, behavioral, and cognitive factors driving technological concern, benefit optimization, and the Acceptance of Purpose. We detail the core theoretical and empirical contributions of this study below:

1. The Label Effect: Semantic Labeling as a Cognitive Anchor (RQ1 & Chapter 13)

Our primary experimental manipulation revealed a highly significant main effect of the Label Effect on Perceived Risks (F(1, 386) = 22.42, p < .0001, ηp2 = 0.055). Exposing participants to the threatening label “Deepfake” rather than the neutral label “Synthetic Avatar” significantly elevated risk perceptions and threat anxiety.

In addition, Chapter 13 (Moderation Analysis) proved that this Label Effect is statistically non-significant in its interaction with AI Expertise (F(1, 389) = 0.10, p = 0.7488). Post-hoc simple slopes verified that the Label Effect remains highly significant at low (p = 0.0021), average (p < 0.0001), and high (p = 0.0004) levels of AI expertise.

Takeaway: Semantic labeling is not merely a bias that only influences laypeople; its cognitive anchoring power is equally potent for highly knowledgeable experts. The term “Deepfake” triggers a deep-seated risk frame that even advanced technical knowledge cannot override.

2. The Gender Threat Gap and Selective Relaxation (RQ1a & Chapter 10)

A large, highly significant gender threat gap was identified across the entire dataset. Female participants reported significantly higher baseline perceived risks (M = 3.65 vs. 3.38, p = 0.0246) and significantly lower baseline Acceptance of Purpose (M = 1.53 vs. 1.85, p < .0001).

Furthermore, as illustrated in Figure 7 (Chapter 10), the interaction between Label and Gender on the Acceptance of Purpose Scale shows a marginally significant “selective relaxation” trend (p = 0.0766). Substituting the positive label “Synthetic Avatar” for “Deepfake” selectively relaxes threat concern and Acceptance of Purpose barriers for female participants (marginally increasing their Acceptance of Purpose from M = 1.45 to 1.61, p = 0.083), whereas it has no influence on male participants (M = 1.88 vs. 1.81, p = 0.44).

Takeaway: Men operate under a highly permissive, stable baseline that is immune to semantic labeling. Women, conversely, maintain a heightened baseline level of concern that is highly sensitive to semantic labeling; positive rebranding selectively lowers their psychological barriers and normalizes the technology.

3. Population Differences: Local vs. National Panels (RQ1b & Chapter 8)

A highly significant, consistent main effect of Population Cohort was identified across all three primary scales: Perceived Risks (p = 0.0008), Perceived Benefits (p = 0.0019), and Acceptance of Purpose (p = 0.0018). * Participants from the national panel perceived significantly higher risks (M = 3.70) and lower Acceptance of Purpose (M = 1.58). * Local business students exhibited significantly lower risks (M = 3.33) and higher Acceptance of Purpose (M = 1.80). * Most notably, Figure 6 (Chapter 10) shows a significant Gender × Population interaction on risks (p = 0.0168). While a large gender gap exists in the national panel (p = 0.0005), this gap is equalized within the business student cohort (M = 3.35 for females, M = 3.32 for males, p = 0.8287).

Takeaway: Business school students are statistically less concerned, more optimistic, and show a complete bridging of the gender threat gap. This points to a powerful effect of academic normalization and tech exposure, which equalizes threat perceptions and fosters optimistic technological baselines.

4. The Cognitive Awareness of Susceptibility (Chapter 14)

The binary logistic regression in Chapter 14 predicting real-world deepfake susceptibility yielded a counterintuitive finding: * For every 1-unit increase in general AI Knowledge, participants exhibited a 1.55 times higher odds of reporting having been fooled by a deepfake (p = 0.0309, OR = 1.55). * Conversely, subjective familiarity predicted a 30% decrease in reported susceptibility (p = 0.0095, OR = 0.70).

Takeaway: Higher technical expertise does not prevent individuals from being deceived; rather, it builds the critical cognitive skills necessary to realize and spot when they have been fooled. Subjective familiarity, meanwhile, may reflect either active vigilance or a subjective overconfidence.

5. Generative Creator Status and Threat Appraisals (Chapter 15)

The Welch’s \(t\)-test in Chapter 15 compared Perceived Risks between active creators of generative media (such as AR face filters or deepfake swaps, N = 53) and non-creators (N = 336). * While creators reported descriptively lower concern (M = 3.39) than non-creators (M = 3.63), this difference was not statistically significant (t(69.99) = -1.55, p = 0.1266).

Takeaway: Hands-on experience with basic generative media (such as playing with AR filters on social media) is insufficient to alter general societal threat concern. Creators remain highly concerned about the societal dangers and malicious misuses of high-fidelity cloning.

6. Intimate Personal Exposure and Acceptance Normalization (Chapter 16)

In contrast to casual creator status, Chapter 16 revealed that personal engagement with generative cloning—specifically, having created a realistic deepfake video of oneself (SelfCloning)—was the single strongest predictor of Acceptance of Purpose normalization in the entire study: * A highly statistically significant, positive correlation was identified between self-cloning and the Acceptance of Purpose Scale (r = 0.22, p < 0.0001). * Self-cloners (N = 18) found target generative AI purposes substantially more acceptable (M = 2.31, SD = 0.86) compared to those who have never cloned themselves (M = 1.61, SD = 0.64; N = 374).

Takeaway: Personal engagement is a far stronger driver of technological acceptance than technical knowledge.

7. The Emotional Pathway as the Primary Causal Mechanism (Chapter 17)

By filtering the feelings mediator to a 3-level linear concern scale (Feelings_Concern: 1 = great, 2 = concerning, 3 = dangerous; N = 355), we proved that the Label Effect operates entirely through an emotional concern pathway: * Path \(a\) (Label → Concern): Exposing participants to “Deepfake” significantly increased emotional concern by \(0.20\) points compared to “Synthetic Avatar” (p = 0.0007). * Path \(b\) (Concern → Acceptance of Purpose): Higher emotional concern was a negative predictor of usage acceptability (p < 0.0001). * Purposes Mediation: We found Full Causal Mediation (ACME Estimate = -0.069, p < 0.0001), while the direct effect was completely non-significant (p = 0.640). The labeling effect on Acceptance of Purpose is entirely carried by the emotional anxiety that the term “Deepfake” evokes. * Risks Mediation: Emotional concern mediated a substantial portion (63.9%) of the labeling effect on general perceived risks (ACME Estimate = 0.182, p < 0.0001), while the direct effect was non-significant (p = 0.176).

Takeaway: Semantic labeling is not a direct logical driver of technological acceptance. Rather, it operates as an emotional volume knob. Labeling deepfakes as “Synthetic Avatars” mitigates threat anxiety and increases the Acceptance of Purpose primarily by reducing the negative emotional concern and threat anxiety associated with the technology.

19. R Session Information and Packages Used

The calculations and plots in this document were compiled using the following setup: - R Version: 4.5.2 (2025-10-31) - Primary Packages: readxl, tidyverse (ggplot2, dplyr, tidyr, purrr), psych (Exploratory Factor Analysis & Reliability), car (Type III ANCOVA), emmeans (Adjusted marginal means & simple slopes), ordinal (Cumulative Link Models), patchwork (Plot alignment), scales (Percentage axes).

File stored in: Z-> CCMC -> Deepfake survey -> HICCS