Table 4.

P-values of the Wilcoxon signed-rank tests adjusted for multiple comparisons using the Holm-Bonferroni method. Effect sizes measured with Cohen’s d reported for significant results.

Criterion/Linguistic characteristic	Human vs. ChatGPT-3	ChatGPT-3 vs. ChatGPT-4	ChatGPT-3 vs. ChatGPT-4
Topic and completeness	$<0.001(d=-0.77)$	$<0.001(d=-1.09)$	0.095
Logic and composition	$<0.001(d=-0.84)$	$<0.001(d=-1.20)$	$0.025(d=-0.45)$
Expressiveness and compr.	$0.008(d=-0.57)$	$<0.001(d=-0.88)$	0.055
Language mastery	$<0.001(d=-1.15)$	$<0.001(d=-1.43)$	0.105
Complexity	$0.025(d=-0.52)$	$<0.001(d=-0.99)$	$0.025(d=-0.48)$
Vocabulary and text linking	$<0.001(d=-0.76)$	$<0.001(d=-1.27)$	$0.012(d=-0.50)$
Language constructs	$<0.001(d=-0.82)$	$<0.001(d=-1.15)$	0.105
Lexical diversity	$<0.001(d=1.06)$	$0.001(d=-0.60)$	$<0.001(d=-1.93)$
Syntactic complexity (depth)	$0.001(d=-0.59)$	0.055	0.105
Syntactic complexity (clauses)	$<0.001(d=-0.93)$	$0.004(d=-0.54)$	$0.024(d=0.49)$
Nominalizations	$<0.001(d=-0.88)$	$<0.001(d=-1.35)$	$0.020(d=-0.29)$
Modals	$0.025(d=0.39)$	$<0.001(d=1.08)$	$<0.001(d=0.76)$
Epistemic markers	$<0.001(d=1.01)$	$<0.001(d=1.53)$	$0.005(d=0.65)$
Discourse markers	0.150	$<0.001(d=0.98)$	$<0.001(d=0.85)$