Fig 1.
MDS configuration for function word unigrams in eight classes including each LLM and human.
* The distances (e.g., symmetric Jensen–Shannon distances) between each object on the two dimensions indicate relative similarity of concerned stylometric features.
Fig 2.
MDS configuration for POS bigrams in eight classes including each LLM and human.
Fig 3.
MDS configuration for phrase patterns in eight classes including each LLM and human.
Fig 4.
MDS configuration for integrated features in eight classes including each LLM and human.
* Each text marked by arc and label of LLMs was used as stimulus and presented to the participants in a web survey of Study 2.
Table 1.
Cross tabulation for text types (generated by LLMs and written by human) and judgments of survey participants.
Fig 5.
Violin plots and box-and-whisker plots of judgments of participants (a 7-point Likert scale) against texts corresponding to each LLM and human.
* Each violin plot with light blue shows the shape of the distribution via a kernel density estimate. The red diamond in a box-and-whisker plot represents the median of the distribution.
Table 2.
Results of analysis for judgments of participants using Bayesian multinomial logistic regression model with random intercepts.
Table 3.
Results of analysis for confidences of participants using Bayesian ordinal logistic regression model with random intercepts.
Fig 6.
Violin plots and box-and-whisker plots of confidence of participants (a 6-point Likert scale) against texts corresponding to each LLM and human.
* Each violin plot with light green shows the shape of the distribution via a kernel density estimate. The red diamond in a box-and-whisker plot represents the median of the distribution.