Fig 1.
Illustrations (screenshots) of the interface.
Illustrations (screenshots) of the interface. From top left to bottom right: (i) a correct answer to a type-B question, an incorrect answer to a type-A question, (iii) an incorrect answer to a type-B questions, and (iv) a correct answer to a type-A question. Similar visual feedback is provided in both situations: (1) the number of (in)correct answers is shown in the top (green and red circles); (2) immediately below, the progress bar shows the time elapsed; (3) provenance information (artist name and album title) appear below the authentic fragment.
Table 1.
Statistics on the number of stimuli.
Table 2.
Corpus statistics.
Table 3.
Song statistics.
Table 4.
Text generation model details.
Table 5.
Examples of generated samples.
Table 6.
Linguistic feature statistics.
Fig 2.
Marginal effects plot of trial number on the authentication accuracy.
A: Marginal effects plot of trial number for both type-A and type-B questions; B: Marginal effects plot showing the effect of trial number on the accuracy of classifying text fragments as “Authentic” or “Generated” in type-B questions alone.
Fig 3.
Marginal effects plot, showing the effect of Trial number on the perceived authenticity of text fragments.
Fig 4.
Marginal effects plot showing the effect of different language models.
Marginal effects plot showing the effect of different language models in interaction with conditioning on the accuracy of classifying text fragments as “Authentic” or “Generated”.
Table 7.
WAIC scores as estimates of out-of-sample deviance.
Fig 5.
Estimates of objective (a) and subjective (b) feature importance.
Fig 6.
Estimates of (subjective) feature importance for high-scoring participants.