Beyond binary AI text detection: Partial authorship, source-text similarity, and an integrated threshold model of detectability
Abstract
This study investigates AI text detection under conditions that more closely reflect real-world writing practices than conventional binary human-versus-AI benchmarks. Rather than treating authorship as a dichotomy, detectability is examined as a function of partial AI authorship, source-text similarity, and generator heterogeneity. Ten pre-LLM, human-written English scientific abstracts were used as reference texts, from which four authorship states (H100, AI50, AI75, and AI100) were constructed through controlled rewriting and paragraph-level mixing. AI-generated content was produced using four large language models: ChatGPT, Gemini, Claude, and LLaMA. Detection outputs from Turnitin and GPTZero were evaluated using non-parametric statistics, correlation analysis, and logistic threshold modeling. The results demonstrate that fully human and fully AI-rewritten texts are statistically distinguishable: median AI scores increased from 0 in H100 to 53 in AI100 for Turnitin ( U = 49.5, P < 0.001) and from 0 to 11 for GPTZero ( U = 61.0, P < 0.001), although AI-generated texts did not constitute a homogeneous class. Detectability also varied significantly across the four nominal authorship states for both Turnitin (H = 30.39, P < 0.001) and GPTZero (H = 21.77, P < 0.001), and systematically with source-text similarity and the generating model. Turnitin showed greater sensitivity than GPTZero to changes in nominal AI proportion and transformation intensity, particularly under mixed-authorship conditions. Significant generator-dependent differences were also observed, with Gemini outputs generally receiving higher AI-detection scores than ChatGPT or LLaMA outputs. The main methodological contribution is the Integrated Detectability Threshold Model (IDTM), which represents detection probability as a joint function of nominal AI proportion and transformation intensity. Within the present sample, the model identified a similarity transition range of approximately 63–66%, below which Turnitin-based detection became more likely than non-detection. These findings support a shift from binary AI classification toward a threshold- and risk-sensitive interpretation of AI text detectability.