Nazarovets, Serhii The Marker Drift Problem: Measuring LLM-Assisted Writing with Unstable Lexical Indicators., 2026 [Preprint]

Lexical markers are increasingly used to estimate the prevalence of large language model (LLM)-assisted writing across large publication corpora. However, longitudinal applications of this approach rely on an important assumption: that the relationship between these markers and LLM use remains sufficiently stable over time. Recent research suggests that once publicly identified, individual markers may lose their diagnostic value due to adaptation by users and the evolution of language models. Furthermore, lexical markers may reflect not only LLM use, but also linguistic competence, AI-assisted editing, and broader forms of stylistic convergence. This paper introduces the Marker Drift Problem to describe a methodological challenge in which indicators change alongside the phenomenon they are intended to measure. Marker drift raises questions about temporal comparability, as observed changes in marker prevalence may reflect changes in LLM use, changes in the indicators themselves, or both. Without attention to the temporal stability and construct validity of lexical indicators, increasingly precise estimates of LLM-assisted writing may be based on unstable measurement instruments.

Kehkashan, T., Riaz, R. A., Al-Shamayleh, A. S., Akhunzada, A., Ali, N., Hamza, M., & Akbar, F. (2025). AI-generated text detection: A comprehensive review of methods, datasets, and applications. Computer Science Review, 58, 100793. https://doi.org/10.1016/j.cosrev.2025.100793

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779

View Item