GEDE: AI-Generated Essay Detection in Education

• Generative Essay Detection in Education (GEDE) is a benchmark framework that defines and quantifies varied levels of AI and human contribution in student essays.
• It employs multiple detection methods—including statistical, embedding-based, and heuristic approaches—to analyze both final text and writing-process data.
• Empirical findings reveal high detection accuracy for purely human or entirely AI-generated texts, while hybrid-authored essays pose significant challenges.

Generative Essay Detection in Education (GEDE) is a benchmark setting for detecting LLM-generated text in educational contexts, introduced with a dataset containing over 900 student-written essays and over 12,500 LLM-generated essays from various domains. Its defining contribution is the explicit modeling of a spectrum of student and model involvement rather than a binary opposition between “human” and “AI.” In this setting, essays may be purely human-written, slightly LLM-improved, fully LLM-generated, or deliberately “humanized” to evade detectors. The resulting research area connects detector benchmarking, hybrid-authorship analysis, writing-process logging, academic integrity policy, and assessment redesign (Gehring et al., 11 Aug 2025).

1. Benchmark structure and the notion of contribution levels

GEDE was introduced as a benchmark dataset tailored for detecting LLM-generated text in educational contexts. It contains 900+ student-written essays and 12,500+ LLM-generated essays, spanning 886+ unique task descriptions across AAE, PERSUADE 2.0, and BAWE. The generated essays were produced with GPT-4o-mini and Llama-3.3-70B-Instruct, and the dataset and code were released publicly (Gehring et al., 11 Aug 2025).

A central concept in GEDE is the “contribution level,” which represents students’ contribution to a given assignment. This is a methodological departure from earlier detector evaluations that treated each document as wholly human or wholly machine. In GEDE, the detection problem is explicitly conditioned on how much of the final artifact originates from the student, how much from the model, and whether the text was post-processed to evade detection (Gehring et al., 11 Aug 2025).

Contribution level	Characterization
Human	purely human-authored
Improve-Human	small LLM-driven grammar/language fixes
Rewrite-Human	LLM rewrites the human text
Summary-based Generation	LLMs generate essays from human or model-generated summaries
Task+Summary	LLMs write essays from task & summary
Task-based	LLMs generate essays from just the assignment prompt
Rewrite-LLM	LLMs rewrite LLM-generated essays to evade detectors
Humanize	adversarial paraphrasing to actively evade detection

This structure makes GEDE unusually relevant to educational settings, because student use of generative AI is often partial and iterative rather than total. A plausible implication is that any detector evaluated only on “pure” human versus “pure” AI text is benchmarking an easier problem than the one actually faced by teachers and institutions.

2. Detector families, scoring criteria, and benchmark methodology

The GEDE benchmark evaluates several detector families. The zero-shot methods include DetectGPT, Fast-DetectGPT, and Intrinsic-Dim; the supervised methods include Ghostbuster and RoBERTa; GPTZero appears as a proprietary detector. DetectGPT is described as measuring log probability curvature, with score

$$\mathbf{d}(x, p_\theta, q) \triangleq \log p_\theta(x)-\mathbb{E}_{\tilde{x}\sim q(\cdot \mid x)}\log p_\theta(\tilde{x}),$$

while Fast-DetectGPT is presented as a speedup variant and Intrinsic-Dim as an embedding-based persistent-homology approach (Gehring et al., 11 Aug 2025).

Evaluation in GEDE uses ROC-AUC, Macro F1, and specificity, with threshold optimization studied through maximizing F1, Youden’s Index,

$J = \text{Sensitivity} + \text{Specificity} - 1,$

and FPR-based thresholding with strict upper bounds on FPR such as $\leq 5\%$ to limit false accusations (Gehring et al., 11 Aug 2025). The broader survey literature similarly treats AUC, false positive rates, and false negative rates as the core metrics for educational deployment, and emphasizes that writing-process features, watermarking, and similarity matching occupy distinct methodological niches alongside feature-based and end-to-end supervised classifiers (Hao, 2 Mar 2026).

This landscape includes both opaque and explainable systems. “AIDetection” is a JavaScript-based tool that does not use semantic modeling; instead, it scans for ASCII quotation-mark traces, inconsistent character encoding, and explicit mentions of tools such as ChatGPT, Claude, Gemini, Grok, Llama/Meta, Microsoft Copilot, and Grammarly AI. It processes .pdf, .doc, and .docx files, works client-side, and outputs color-coded Excel and CSV reports. Its own paper characterizes the findings as “potential AI traces,” not forensic proof (Buschmann, 12 Mar 2025).

The methodological contrast is important. Statistical detectors attempt to infer provenance from distributional properties of the final text, whereas heuristic tools such as AIDetection attempt to identify artifact traces of copy-paste workflows. This suggests that “detection” in education is not a single technical problem but a family of partially overlapping inference tasks.

3. Hybrid authorship, process data, and localized attribution

A major limitation of early AI-text detection research is the assumption that a document is either entirely human-written or entirely AI-generated. One line of work addresses this by formalizing hybrid essay detection as a boundary-detection problem: given a hybrid text $\langle s_1, s_2, \dots, s_n \rangle$, the task is to identify indices where adjacent sentences are written by different authors. The proposed two-step TriBERT approach first separates AI-generated and human-written content during encoder training and then detects boundaries by locating the largest distances between adjacent local prototypes. On single-boundary hybrid essays, a relatively large prototype size led to a 22% improvement in the In-Domain evaluation and an 18% improvement in the Out-of-Domain evaluation (Zeng et al., 2023).

The main evaluation metric in that work is

$$F1@K = 2 \cdot \frac{|L_{topK} \cap L_{Gt}|}{|L_{topK}| + |L_{Gt}|},$$

with $K=3$ because the dataset contains at most three boundaries per essay (Zeng et al., 2023). This formulation is more fine-grained than whole-document classification and is better aligned with the educational reality of partial assistance.

A complementary direction uses writing-process data rather than final-text statistics. NIRVANA records keystroke-level writing behavior, full ChatGPT conversation histories, and all text copied from ChatGPT for 77 university students writing an analytical essay. It introduces two process metrics:

$HCR = \frac{HA - HD}{(HA - HD) + (GP - GD)}$

and

$HER = \frac{HA + HD + GD}{HA + HD + GP + GD},$

where $HA$ is words added by humans, $HD$ is human-written words later deleted, $J = \text{Sensitivity} + \text{Specificity} - 1,$0 is words pasted from ChatGPT, and $J = \text{Sensitivity} + \text{Specificity} - 1,$1 is GPT-pasted words later deleted (Jelson et al., 8 Apr 2026).

Using these measures, NIRVANA identifies four writing profiles: Lead Authors, Collaborators, Drafters, and Vibe Writers. It also reports that query frequency correlates with essay length ($J = \text{Sensitivity} + \text{Specificity} - 1,$2), time spent, and readability ($J = \text{Sensitivity} + \text{Specificity} - 1,$3) (Jelson et al., 8 Apr 2026). This type of process-aware instrumentation changes the evidentiary basis of GEDE: instead of inferring authorship only from style, it becomes possible to reconstruct when assistance was sought, what was copied, and how heavily the copied text was revised.

Related evidence comes from a study of 1,445 GAI-assisted writing sessions. It distinguishes three behaviors—seeking suggestions and not accepting them, accepting them as they are, and accepting them with modification—and reports that writers who frequently modified GAI-generated text consistently improved lexical sophistication, syntactic complexity, and text cohesion, whereas writers who often accepted suggestions without changes saw a decrease in essay quality. The paper explicitly argues that, for GEDE implementations, behavioral cues and process logs are more informative than linguistic features alone (Yang et al., 2024).

4. Empirical performance, model drift, and adversarial fragility

GEDE’s benchmark results show that detector performance is highly contingent on contribution level. For “pure” Human versus “pure” Task-based LLM essays, top zero-shot detectors, especially Fast-DetectGPT, achieve ROC-AUC $J = \text{Sensitivity} + \text{Specificity} - 1,$4. Performance drops substantially on intermediate levels: detectors struggle with LLM-improved or rewritten human texts, ROC-AUC drops to $J = \text{Sensitivity} + \text{Specificity} - 1,$5 or lower, and even minor LLM-enhanced grammatical changes to human texts often yield false positives. Performance also deteriorates greatly on short essays of roughly 50–100 words, and supervised models generalize poorly to unseen prompts, contribution types, and LLMs unless retrained on all contribution levels (Gehring et al., 11 Aug 2025).

Cross-LLM generalization is likewise unstable. A study on GRE essays generated by GPT-4, GPT-4o, GPT-o1, GPT-o3-mini, GPT-o4-mini, and GPT-5 trained Gradient Boosting Machine detectors on perplexity-derived features. It found within-model AUCs close to 1, high mutual generalizability among GPT-4, GPT-4o, GPT-o1, GPT-o3-mini, and GPT-4o-mini with AUC $J = \text{Sensitivity} + \text{Specificity} - 1,$6 and often $J = \text{Sensitivity} + \text{Specificity} - 1,$7, but poor transfer to GPT-o4-mini and GPT-5, which behaved as outliers. A unified “GPT-all” detector trained on all LLMs achieved balanced performance of approximately 0.9 AUC or better across the full set (Hao, 2 Mar 2026).

Adversarial evaluation further destabilizes the picture. “AIG-ASAP” was constructed to test word substitution, sentence substitution, and paraphrasing attacks on AI-generated student essays. Unperturbed essays were easily detected, with RoBERTa-QA exceeding 90% in Accuracy and AUROC for almost all generators, but word substitution reduced detection accuracy close to random chance, often around 50–60%, while preserving essay quality at roughly the same automated score. Fine-tuning helped against paraphrasing attacks but not against word or sentence substitution (Peng et al., 2024).

A larger study of six major GenAI detectors reported already low accuracy rates of 39.5% and major reductions in accuracy of 17.4% when the content was manipulated. The most effective evasion techniques were adding spelling errors and increasing burstiness, with accuracy after those techniques reported as 12.9% and 15.9% respectively; paraphrasing produced 18.4%, decreasing complexity 21.0%, writing as NNES 27.7%, and increasing complexity 37.0% (Perkins et al., 2024). These findings support the recurring conclusion that detector evaluations on clean, unedited model outputs can substantially overstate real-world performance.

Human readers are also limited. In ArguGPT, 43 novice and experienced ESL instructors achieved 61.6% accuracy when first exposed to machine-generated essays and 67.7% after one round of minimal self-training. They were much better at identifying human essays than machine essays, and the paper reports that machines produced sentences with more complex syntactic structures while human essays tended to be lexically more complex (Liu et al., 2023).

5. False positives, bias, privacy, and the ethics of accusation

Across this literature, false positives are treated as the defining institutional risk. GEDE states directly that detectors are particularly likely to produce false positives and that this is problematic in educational settings where false suspicions can severely impact students’ lives (Gehring et al., 11 Aug 2025). Curriculum-theoretical work sharpens the claim into an ethical principle: “It is more acceptable for 100 students to commit plagiarism via ChatGPT than for one student to suffer a wrongful accusation due to AI-detection software.” The same paper cites an incident in which an entire class was falsely accused by an automated detector, with graduation threatened, and reports that GPTZero marked the US Constitution and the Bible as “most likely generated by AI” (Healy, 2023).

The higher-education critique of detector use extends this argument quantitatively. If the per-assignment false positive rate is $J = \text{Sensitivity} + \text{Specificity} - 1,$8 over $J = \text{Sensitivity} + \text{Specificity} - 1,$9 written assignments, then

$\leq 5\%$0

With $\leq 5\%$1 and $\leq 5\%$2, the probability becomes $\leq 5\%$3, meaning that even a nominally low false positive rate compounds into a high cumulative risk over a degree program (Ardito, 2023). That paper also argues that detectors may be more likely to flag non-native English writers, that students may be pushed to upload work to third-party services before submission, and that such practices effectively introduce surveillance into assessment workflows (Ardito, 2023).

Inclusivity research reinforces the point. In one detector comparison, Copyleaks had a false accusation score of 50%, Turnitin had 0% false accusation but 84% undetected cases, and the average proportion of AI-generated texts left undetected after adversarial manipulation was 65.7%. The same study emphasizes that NNES-style writing both decreases detectability for machine-generated text and overlaps with the population most exposed to wrongful suspicion (Perkins et al., 2024).

The misconception that detector output constitutes proof is rejected repeatedly. AIDetection presents its reports as a basis for intervention or discussion rather than disciplinary certainty (Buschmann, 12 Mar 2025). The broader literature similarly recommends triangulating detector output with other evidence, especially writing-process data and revision logs, and stresses that mixed-origin text is inherently difficult to classify reliably (Hao, 2 Mar 2026).

6. Pedagogical adaptation, assessment redesign, and research infrastructure

A large share of the literature concludes that educational response cannot be reduced to better detection alone. One prominent position argues for a strategic shift toward robust assessment methods and educational policies that embrace generative AI usage while ensuring academic integrity and authenticity in assessments. Its practical framework divides assessment redesign into three categories: Remove/Reduce memory recall, rote composition, and factual recall; Adapt/Integrate AI-assisted essay creation, critical analysis and co-creation, and group assessments or presentations; Fortify in-person exams, oral exams, and controlled labs (Ardito, 2023).

Curriculum theory reaches a parallel conclusion. Generative AI is treated as an “inflection point” in the life of curriculum, and the act of relegating it to the null curriculum is described as “tantamount to willful blindness.” The same work recommends moving teaching with and about GenAI from the “null” to the “explicit” side of the curriculum ledger, rejecting overreliance on detectors and modeling openness, adaptability, and trust rather than “algorithms over human trust” (Healy, 2023).

Empirical work on instruction supports this shift toward critical engagement. In a pilot study in introductory Computational and Data Science courses, a 15-minute prerecorded video lecture on LLMs, hallucinations, and GAI limitations was paired with guided exercises and a four-prompt short-answer assignment in which students analyzed, critiqued, and revised GAI-generated solutions. Using a Kruskal-Wallis nonparametric test, the intervention produced a statistically significant improvement in assignment scores in CDS 101 with $\leq 5\%$4 at significance level $\leq 5\%$5, although the significance vanished when non-submitters were excluded (Lamberti et al., 29 Aug 2025).

The longer-term research infrastructure for this agenda is increasingly process-rich. ChEDDAR, built from a semester-long longitudinal experiment involving 212 college students in EFL writing courses, includes conversation logs, utterance-level essay edit history, self-rated satisfaction, students’ intent, and session-level pre-and-post surveys. It establishes baseline results for intent detection and satisfaction estimation in educational dialogue systems (Han et al., 2023). NIRVANA adds keystroke logging, full ChatGPT histories, copy-paste provenance, and replay interfaces for analytical essays (Jelson et al., 8 Apr 2026). Together with GEDE, these datasets indicate a shift from output-only auditing toward process-sensitive analysis of how students plan, solicit help, revise, and integrate model text.

A plausible implication is that the future of GEDE lies in combining text-based detectors with process-based evidence, continual retraining across evolving LLM families, and institutionally governed interpretation standards. The literature is increasingly consistent on one point: the educational problem is not only whether a detector can assign an AI label, but whether the surrounding assessment system can distinguish superficial outsourcing from meaningful engagement, while preserving fairness, privacy, and due process (Gehring et al., 11 Aug 2025).