Paper Reading | Linguistic Markers of Inherently False AI Communication and Intentionally False Human Communication: Evidence From Hotel Reviews

Markowitz, D. M., Hancock, J. T., & Bailenson, J. N. (2023). Linguistic Markers of Inherently False AI Communication and Intentionally False Human Communication: Evidence From Hotel Reviews. Journal of Language and Social Psychology. https://doi.org/10.1177/0261927x231200201

Table of contents

• Problem statement
• Literature review
• Data
• Automated text analysis
• Measures
• Analytic plan
• Results
• Reproduction

Problem statement

• Diff in linguistic signals between AI-generated content and human-generated content
• More specifically, human-generated hotel reviews were compared to AI-generated hotel reviews to identify how such text types are different in terms of content (e.g., emotion), style (e.g., analytic writing, adjectives), and structure (e.g., readability).
• Discuss two types of false: inherently false (AI) vs. intentionally false (human)

Literature review

AI-mediated communication

• H1: AI-generated text will be more affective than human-generated text.
• [Style] Analytic Writing Index (e.g. function words) → proxy for elaborate thinking
• [Style] Descriptiveness (rate of adjectives)
• [Structural complexity] Readability

Deception Research

• Deception has a different linguistic signature than honesty - channel matters.
• AI - inherently false ~ misinformation, different from intentionally false of human-generated content

Data

Equivalent prompts:

• 400 truthful human-written reviews
• 400 deceptive human-written reviews
• 400 AI-generated hotel reviews (ChatGPT 3.5)

Automated text analysis

• LIWC
• Coh-Metrix
• R package quanteda.textstats

Measures

• [content] rate of affect terms
• [style] analytic writing
• [style] adjectives
• [structure] readablity

Analytic plan

• Linear mixed model
  • fixed effect = review type (AI or decep or truth)
  • random effect = hotel name
• Classification model
  • classification acc of distinguishing between AI-generated text and human text

Results

linguistic signals:

• AI-generated content is more affective than deceptive human content (and truthful).
• AI-generated content is more analytic than deceptive human content.
• AI-generated content is more descriptive than deceptive human content (and truthful).
• AI-generated content is less readable than deceptive human content (and truthful).

classification accuracy:

• “88.5%”

Reproduction

library(readr)
library(tidyverse)
library(lme4)
library(lmerTest)

# https://osf.io/nrjcw/
data = read_csv("./data/hotel_reviews_AIMC_OSF_FINAL.csv")
subdata = data %>%
            transmute(readability, Analytic, adj, Affect,
                      type = as.factor(type), hotel = as.factor(hotel))

# DV: Affect
# IV: type
# random intercept: hotel
mixed.mod1 = lmer(Affect ~ type + (1|hotel), data = subdata)
summary(mixed.mod1)

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: Affect ~ type + (1 | hotel)
   Data: subdata

REML criterion at convergence: 6056.1

Scaled residuals:
    Min      1Q  Median      3Q     Max
-2.4440 -0.6484 -0.0988  0.5421  6.0932

Random effects:
 Groups   Name        Variance Std.Dev.
 hotel    (Intercept) 0.149    0.386
 Residual             8.985    2.998
Number of obs: 1200, groups:  hotel, 20

Fixed effects:
              Estimate Std. Error       df t value Pr(>|t|)
(Intercept)     12.137      0.173   74.974   70.17   <2e-16 ***
typedeceptive   -4.461      0.212 1178.000  -21.05   <2e-16 ***
typetruthful    -4.857      0.212 1178.000  -22.92   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
            (Intr) typdcp
typedeceptv -0.613
typetruthfl -0.613  0.500

# sig01: the random effect of between group variation
# sigma: the residual random effect
confint(mixed.mod1)

subdata2 = subdata %>%
    mutate(type2=relevel(type, ref='truthful'))
mixed.mod11 = lmer(Affect ~ type2 + (1|hotel), data = subdata2)
summary(mixed.mod11)

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: Affect ~ type2 + (1 | hotel)
   Data: subdata2

REML criterion at convergence: 6056.1

Scaled residuals:
    Min      1Q  Median      3Q     Max
-2.4440 -0.6484 -0.0988  0.5421  6.0932

Random effects:
 Groups   Name        Variance Std.Dev.
 hotel    (Intercept) 0.149    0.386
 Residual             8.985    2.998
Number of obs: 1200, groups:  hotel, 20

Fixed effects:
                Estimate Std. Error        df t value Pr(>|t|)
(Intercept)       7.2795     0.1730   74.9735  42.089   <2e-16 ***
type2chatGPT      4.8573     0.2120 1178.0000  22.916   <2e-16 ***
type2deceptive    0.3961     0.2120 1178.0000   1.869   0.0619 .
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
            (Intr) ty2GPT
type2chtGPT -0.613
type2decptv -0.613  0.500

Interpretations:

• The mean of Affect for ChatGPT messages is 12.14
• The mean of Affect for Deceptive messages is 12.137-4.461=7.68
• The mean of Affect for Truthful messages is 12.137-4.857=7.28
• The difference in the mean Affect scores between ChatGPT messages and Deceptive messages is 4.46, which is significantly greater than 0
• The difference in the mean Affect scores between ChatGPT messages and Truthful messages is 4.86, which is significantly greater than 0
• The difference in the mean Affect scores between Deceptive messages and Truthful messages is 0.40, which is not significantly greater than 0

# DV: Analytic
# IV: type
# random intercept: hotel
mixed.mod2 = lmer(Analytic ~ type + (1|hotel), data = subdata)
summary(mixed.mod2)

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: Analytic ~ type + (1 | hotel)
   Data: subdata

REML criterion at convergence: 10548.5

Scaled residuals:
    Min      1Q  Median      3Q     Max
-3.0281 -0.6291  0.1752  0.7354  1.9173

Random effects:
 Groups   Name        Variance Std.Dev.
 hotel    (Intercept)   7.008   2.647
 Residual             382.915  19.568
Number of obs: 1200, groups:  hotel, 20

Fixed effects:
              Estimate Std. Error       df t value Pr(>|t|)
(Intercept)     71.146      1.144   71.441  62.215  < 2e-16 ***
typedeceptive   -9.633      1.384 1178.000  -6.962 5.58e-12 ***
typetruthful    -1.766      1.384 1178.000  -1.277    0.202
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
            (Intr) typdcp
typedeceptv -0.605
typetruthfl -0.605  0.500

# DV: adj
# IV: type
# random intercept: hotel
mixed.mod3 = lmer(adj ~ type + (1|hotel), data = subdata)
summary(mixed.mod3)

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: adj ~ type + (1 | hotel)
   Data: subdata

REML criterion at convergence: 6168.9

Scaled residuals:
    Min      1Q  Median      3Q     Max
-2.4988 -0.6994 -0.0880  0.6096  4.6913

Random effects:
 Groups   Name        Variance Std.Dev.
 hotel    (Intercept) 0.06235  0.2497
 Residual             9.93075  3.1513
Number of obs: 1200, groups:  hotel, 20

Fixed effects:
               Estimate Std. Error        df t value Pr(>|t|)
(Intercept)     11.6243     0.1672  110.5426   69.54   <2e-16 ***
typedeceptive   -3.5195     0.2228 1178.0000  -15.79   <2e-16 ***
typetruthful    -2.4894     0.2228 1178.0000  -11.17   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
            (Intr) typdcp
typedeceptv -0.667
typetruthfl -0.667  0.500

# DV: readability
# IV: type
# random intercept: hotel
mixed.mod4 = lmer(readability ~ type + (1|hotel), data = subdata)
summary(mixed.mod4)

Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: readability ~ type + (1 | hotel)
   Data: subdata

REML criterion at convergence: 8933.4

Scaled residuals:
    Min      1Q  Median      3Q     Max
-6.8669 -0.5154  0.0740  0.6653  2.6982

Random effects:
 Groups   Name        Variance Std.Dev.
 hotel    (Intercept)  3.255   1.804
 Residual             98.784   9.939
Number of obs: 1200, groups:  hotel, 20

Fixed effects:
               Estimate Std. Error        df t value Pr(>|t|)
(Intercept)     55.2123     0.6401   52.7172   86.26   <2e-16 ***
typedeceptive   11.4165     0.7028 1178.0000   16.24   <2e-16 ***
typetruthful    14.3212     0.7028 1178.0000   20.38   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
            (Intr) typdcp
typedeceptv -0.549
typetruthfl -0.549  0.500