Detecting AI-Generated Essays in Writing Assessment: Responsible Use and Generalizability Across LLMs

Students increasingly use the assistance of large language models (LLMs) in their academic writing. While slight assistance (e.g., grammar and style correction, as well as feedback) is permitted under most institutional policies, it is usually forbidden to offload entire writing tasks to LLMs. Unfortunately, current approaches to LLM-generated text detection predominantly assume a binary distinction between human-written and LLM-generated text, ignoring the breadth of realistic human-AI collaboration practices and limiting the validity of detection systems for educational assessment. In this paper, we propose a contribution-aware evaluation framework for LLM-based detection systems in education. We introduce a scale of eight student contribution levels that model realistic writing scenarios ranging from fully human-written texts to LLM-assisted revisions to fully LLM-generated and adversarially humanized texts. Institutional policies regarding LLM use can then be translated to thresholds of acceptable LLM assistance on this scale. We further present Generative Essay Detection in Education (GEDE), a novel benchmark dataset comprising more than 900 human-written and over 12,500 generated essays across 886 tasks and all contribution levels. Using this benchmark, we conduct a systematic evaluation of four state-of-the-art zero-shot and supervised detection methods across policy boundaries, contribution levels, generative models, out-of-distribution data, and text length. We show that most detectors struggle to accurately classify texts at intermediate student contribution levels, in particular LLM-assisted revisions of human-written texts. Such errors pose a substantial risk of false accusations, indicating that current text detection systems are (still) unsuitable to reliably support the enforcement of institutional policies regarding LLM assistance in education.

Automated Essay Scoring (AES) has been explored for decades with the goal to support teachers by reducing grading workload and mitigating subjective biases. While early systems relied on handcrafted features and statistical models, recent advances in Large Language Models (LLMs) have made it possible to evaluate student writing with unprecedented flexibility. This paper investigates the application of state-of-the-art open-weight LLMs for the grading of Austrian A-level German texts, with a particular focus on rubric-based evaluation. A dataset of 101 anonymised student exams across three text types was processed and evaluated. Four LLMs, DeepSeek-R1 32b, Qwen3 30b, Mixtral 8x7b and LLama3.3 70b, were evaluated with different contexts and prompting strategies. The LLMs were able to reach a maximum of 40.6% agreement with the human rater in the rubric-provided sub-dimensions, and only 32.8% of final grades matched the ones given by a human expert. The results indicate that even though smaller models are able to use standardised rubrics for German essay grading, they are not accurate enough to be used in a real-world grading environment.

Automated detection of LLM-generated texts (LGTs) is critical, yet dedicated detectors often struggle to generalize across domains and models. While general-purpose LLMs offer flexible zero-shot authorship classification with explanatory rationale, their detection behavior, especially regarding self-detection versus cross-detection across model generations, remains poorly understood. We systematically evaluate 15 LLMs spanning three model generations as both generators and detectors. Using a benchmark of 1,000 human-written texts and 15,000 LGTs (1,000 per model), we collected over 233,000 binary classifications alongside natural-language explanations. Our results reveal that detection efficacy is primarily driven by detector capability rather than generator provenance, although outputs from newer generators remain notably harder to detect. Crucially, statistical comparisons show no systematic advantage or disadvantage for self-detection across models. Error analysis further exposes generational bias shifts: first-generation detectors under-detect LGTs (high false-negative rates), second-generation detectors over-flag human texts (high false-positive rates), and the latest models achieve balanced trade-offs. Finally, we highlight significant inconsistencies in how different LLMs apply textual cues to justify their decisions. Code: https://github.com/hyyuan/detect-llm-generated-texts.