A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions

Motivation

• LLM-generated text has expanded across advertising, news, stories, code, education, law, biology, and medicine, with AI-generated news increasing 55.4% on mainstream sites and 457% on misinformation sites.The reported increases cover January 1, 2022, to May 1, 2023.Source: §1 · p. 2
• LLM-generated text creates risks through fabrication, outdated information, prompt sensitivity, disinformation, fraud, spam, plagiarism, and academic dishonesty.These risks can spread erroneous knowledge, undermine expertise, and facilitate malicious exploitation, including students using LLMs for essay writing.Source: §1 · p. 2
• Recursive use of LLM-generated data may reduce the quality and diversity of future models, as warned by the Model Autophagy Disorder analysis (Alemohammad et al. 2023).As models improve, generated text may increasingly replace human-curated training data, creating an AI data feedback loop.Source: §1 · p. 2

Problem

• Current commercial detectors are unreliable and biased toward human-written classifications, while human-assisted detection performs only slightly better than random classification.Human identification can also be less accurate than detectors or algorithms across environmental settings.Source: §1 · pp. 2–3

Contributions

• The survey reviews detection mechanisms, datasets, benchmarks, watermarking, statistical and neural methods, human-assisted approaches, attacks, out-of-distribution challenges, real-world data issues, evaluation gaps, and future research directions.It aims to guide researchers toward efficient detectors and responsible AI development.Source: §1 · p. 3

2.1 LLM-generated Text Detection Task

• Human discrimination between LLM-generated and human-written text barely exceeds random classification, although LLM text often differs in organization, syntax, length, vocabulary, and emotional intensity.These differences include greater organization, formality, objectivity, and comprehensiveness, alongside occasional nonsensical or fabricated details; detectability has also been reported for GPT-3.5-Turbo and GPT-4 (Chakraborty et al. 2023b; Helm, Priebe, and Yang 2023).Source: §2.1 · pp. 3–5

2.2.1 Generation Mechanisms of LLMs.

• LLMs generate text one token at a time from probabilities conditioned on the input and previously decoded tokens, while greedy, beam, top-k, and top-p decoding trade off speed, quality, coherence, and diversity.Greedy search is simple but lacks diversity; beam search improves quality but may duplicate; top-k increases diversity with weaker quality control; and top-p couples diversity to parameter P.Source: §2.2.1 · pp. 5–6, Table 2 · p. 6

2.2.2 Sources of LLMs’ Strong Generation Capabilities.

• LLM generation capabilities arise from in-context learning, human-preference alignment, and enhanced reasoning procedures such as Chain of Thought, Tree of Thoughts, and Graph of Thoughts.In-context learning adapts outputs through prompts without changing model parameters, while RLHF uses human evaluations, reward modeling, and PPO to align behavior with human values and instructions (Ouyang et al. 2022; Lambert et al. 2022; Schulman et al. 2017a).Source: §2.2.2 · p. 7

2.3 Why Do We Need to Detect Text Generated by LLMs?

• Detection is needed because aligned LLM content is spreading across everyday platforms, creating concerns about legal ownership, user trust, academic integrity, model self-consumption, and the diversity of human communication.The survey frames these concerns through regulation, users, developments, science, and human society; examples include copyright debates (Epstein et al. 2023; Wikipedia 2023), self-consuming training and LLM Autophagy Disorder (Alemohammad et al. 2023), and threats to academic validation and linguistic diversity.Source: §2.3 · pp. 7–9
• Detection is not universally necessary: it may be superfluous when LLM use poses minimal risk or remains confined to predictable, limited domains, because unnecessary detection increases inefficiency and development costs.The authors nevertheless argue that detection and filtering can preserve trust, academic integrity, human independent thinking, and linguistic and informational diversity.Source: §2.3 · pp. 8–10

3.1 Related Works

• Prior surveys established foundations for machine-generated text detection, but this review argues that existing LLM-focused reviews need broader method coverage, critical evaluation, and analysis of motivation, data, and evaluation.The survey positions itself as more systematic and comprehensive, covering detection mechanisms, attacks, benchmarks, evaluation frameworks, challenges, and future directions.Source: §3.1 · pp. 10–11
• The review adds pragmatic analysis of model-size effects, mixed human-LLM text, and the lack of effective evaluation frameworks.It also evaluates detection strategies’ effectiveness, strengths, and weaknesses in the current LLM environment.Source: §3.1 · p. 11

3.2 Systematic Investigation and Implementation

• Using Kitchenham’s System for Literature Review framework, the study searched authoritative databases before November 2023 for research on LLM-generated text detection and related terminology.The search combined five topic-specific phrases with the Boolean operator OR and applied criteria covering reviews, detection methodologies, challenges or future directions, and necessity or applications.Source: §3.2 · pp. 11–13
• 83 pertinent works remained after de-duplication and manual screening, with most relevant research published in 2023 and later sections organized around data, detectors, metrics, issues, and future directions.The search process acknowledged inaccurate or duplicated retrievals and weakly related results as limitations.Source: §3.2 · pp. 11, 13, Figure 3 · p. 12

4. Data

• High-quality datasets enable efficient detector training, calibration, and standardized evaluation, but current resources remain nascent, costly to construct, and limited in volume and sample complexity.The survey organizes detection datasets, repurposable datasets, and benchmarks while evaluating their shortcomings to guide future data-resource design.Source: §4 · pp. 13–14

4.1.1 Detection Datasets.

• Detection datasets span paired human–LLM corpora, multiple generators, and application domains, but many remain narrow, monolithic, or insufficiently cross-domain and multilingual.HC3 compares human and ChatGPT responses in English and Chinese; CHEAT models polishing and blending; HC3 Plus adds semantic-invariance tasks; OpenLLMText covers GPT-3.5, PaLM, LLaMA-7B, and GPT2-1B with 344k samples, yet lacks cross-domain and multilingual nuance.Source: §4.1.1 · pp. 14–15, Table 5 · p. 14

4.1.2 Potential Datasets.

• Researchers commonly repurpose Q&A, scientific-writing, story, news, web, social-media, and reasoning datasets as human text, then generate LLM counterparts through prompting or bootstrap complementation.Examples include ELI5, Peer Read, WritingPrompts, XSum, Wiki-40B, Yelp Reviews, and SQuAD, reflecting varied domains and task structures.Source: §4.1.2 · pp. 16, 17, 18, 20

4.2 Evaluation Benchmarks

• Evaluation benchmarks increasingly target diverse generators, domains, languages, tasks, and modalities, but important gaps remain in recency, adversarial coverage, and scenario breadth.TuringBench requires updates for contemporary models; GPABenchmark covers title-based generation, draft completion, and polishing; MULTITuDE evaluates 11 languages; HANSEN targets spoken text; M4 spans domains, generators, and more than ten languages.Source: §4.2 · pp. 21–22, Table 7 · p. 21

4.3 Data Challenges.

• Robust datasets and benchmarks remain insufficient, driving continued reliance on datasets built for unrelated tasks and synthetic LLM text generated from them.This practice reflects the inability of existing resources to comprehensively address diverse research perspectives.Source: §4.3 · p. 23

4.3.1 Comprehensiveness of Evaluation Frameworks.

• Reliable detector evaluation requires benchmarks covering multiple attack types, domains, tasks, models, and languages rather than only superficial challenges.The survey presents these five dimensions as essential for holistic assessment before detectors can gain trust.Source: §4.3.1 · p. 23
• Evaluation datasets should include texts produced through diverse attack mechanisms, such as sophisticated prompt attacks, rather than relying only on simple prompts.The survey identifies attack diversity as important for assessing detector efficacy and echoes this limitation in Guo et al. (2023).Source: §4.3.1 · p. 23
• Datasets should span multiple domains and tasks to assess detectors’ robustness, usability, and credibility across applications such as academic writing, news, arithmetic reasoning, and Q&A.The survey also advocates higher-quality datasets for these evaluations.Source: §4.3.1 · p. 23
• Evaluation should cover multiple LLMs, including LLaMa (Touvron et al. 2023), PaLM (Chowdhery et al. 2022a), and Claude-2,13, rather than focusing only on ChatGPT.The survey argues that emerging models may introduce additional risks that detectors must address.Source: §4.3.1 · p. 23

4.3.2 Temporal.

• Benchmark datasets should be multilingual and regularly updated with outputs from advanced LLMs, because older datasets based on backward models may not reflect current real-world dynamics.Multilingual datasets can reveal detector weaknesses on underrepresented languages, while temporal updates should track rapid field evolution.Source: §4.3.1 · pp. 23–24, §4.3.2 · p. 24

• The section covers four detector families: watermarking, statistics-based, neural-based, and human-assisted methods.Source: §5 · p. 24
• It focuses on the most recently proposed detector designs and detection algorithms.Source: §5 · p. 24
• The discussion is organized according to the underlying principles of each method, with the taxonomy presented in Figure 4.Source: §5 · p. 24

5.1.1 Data-Driven Watermarking.

• Data-driven watermarking embeds backdoors in training data to verify ownership and resist unauthorized fine-tuning, but requires substantial retraining and has limited payload capacity and generalizability.Triggers can be exposed through generated text analysis, especially when they use common words; adding 1% watermarked samples can inject a traceable feature.Source: §5.1.1 · pp. 24, 26

5.1 Watermarking Technology

• The survey classifies LLM-generated text detectors into watermarking, statistics-based, neural-based, and human-assisted methods, with watermarking detection relying on embedded signals and their processing pathways.Figure 4 distinguishes human-written text (HWT) from LLM-generated text (LGT) and depicts detector capability sources and detection processes.Source: Figure 4 · p. 25

5.1.2 Model-Driven Watermarking.

• Model-driven watermarking manipulates logits or token sampling during inference, offering plug-and-play deployment with minimal quality impact but requiring stronger robustness against rewriting, paraphrasing, and key-related attacks.Kirchenbauer et al. (2023a) uses green-red token selection and p-values, while Kuditipudi et al. (2023) reports robustness after modifying approximately 40–50% of tokens.Source: §5.1.2 · pp. 26–27

5.1.3 Post-Processing Watermarking.

• Post-processing watermarking adds watermarks after generation through character, synonym, or sequence-to-sequence transformations, enabling independent watermarking of black-box LLM outputs.Easymark preserves BLEU and perplexity, while REMARK-LLM embeds more signature bits than AWT and improves resistance to removal and detection attacks.Source: §5.1.3 · pp. 27–28

5.2 Statistics-Based Methods

• Statistics-based detection derives thresholds and other regularities from text features or accessible LLM information without supervised training.The approach assumes access to LLMs or extractable text features and evaluates distinctive statistical properties.Source: §5.2 · pp. 28–29

5.2.1 Linguistics Features Statistics.

• Linguistic-feature statistics use grammatical properties, word and phrase frequencies, and repeated n-grams or bigram similarities to distinguish machine- and human-generated text.Hamed and Wu (2023) found that only 23% of ChatGPT-generated-text bigrams were unique and identified 98 of 100 LLM-written academic papers; these methods depend heavily on extensive corpus statistics and multiple LLM types.Source: §5.2.1 · p. 29

5.2.2 White-Box Statistics.

• White-box methods require source-model access and typically compare zero-shot logits-, entropy-, perplexity-, or perturbation-based metrics against statistical thresholds.Log-Likelihood and Rank are established logits-based baselines, while perturbation methods compare original text probabilities with semantically similar alternatives; intrinsic-dimension methods can handle cross-domain, model-shift, and adversarial scenarios but weaken for suboptimal or high-temperature generators.Source: §5.2.2 · pp. 29–30
• An average F1 score of 99.0 was achieved by GHOSTBUSTER, improving by 41.6 F1 points over GPTZero and DetectGPT.GHOSTBUSTER uses weaker language models to obtain token probabilities and searches combinations of their outputs with a linear classifier.Source: §5.2.2 · p. 31

5.2.3 Black-Box Statistics.

• Black-box methods query an LLM to compare original text with continuations, rewrites, paraphrases, or denoised versions, achieving up to 94% accuracy or 91.8% AUROC but incurring substantial access overhead and response latency.These methods exploit differences in similarity or revision behavior between human- and LLM-generated text.Source: §5.2.3 · pp. 31–32

5.3.1 Features-Based Classifiers.

• Linguistic-feature classifiers extract stylistic, complexity, semantic, psychological, and information features before machine-learning classification, achieving reported accuracies of 97% and 93% and F1-scores of 98.36% and 98.29%.Feature examples include word and punctuation frequencies, lexical diversity, semantic dependencies, sentiment, named entities, opinions, and relations.Source: §5.3.1 · pp. 32–33
• These linguistic classifiers can be brittle under ambiguous semantics, weaker than neural features, and limited for detecting LLM-generated misinformation, although statistical features may add adversarial robustness.Stylistic consistency in language-model outputs limits misinformation detection, motivating integrated models that combine statistical and neural signals.Source: §5.3.1 · p. 33
• Model-feature classifiers use token-level perplexity, contrastive features, likelihood, rank, entropy, or LLM bias, reaching 86.0% accuracy or a 98.41% average F1 score, but require source-model logits.This dependency limits effectiveness for powerful closed-source models whose logits are inaccessible.Source: §5.3.1 · p. 33

5.3.2 Pre-Training Classifiers.

• Fine-tuned RoBERTa-based detectors average 95% accuracy in-domain, outperforming zero-shot and watermarking methods, but degrade on cross-domain, unseen, or different-model data; mBERT remains robust in English and Spanish tasks.Encoder-based detectors can overfit training distributions, whereas mBERT performs consistently in document-level classification and model-attribution settings; XLM-RoBERTa and TinyBERT show larger disparities.Source: §5.3.2 · p. 34
• Adversarial and feature-enhanced training improve robustness: OUTFOX surpasses statistical and RoBERTa methods against TF-IDF and DIPPER attacks, while J-Guard’s average performance decrease under attacks is as low as 7%.OUTFOX uses attacker-detector interaction with in-context examples, whereas J-Guard incorporates journalism features for AI-generated news detection.Source: §5.3.2 · pp. 35–36

5.3.3 LLMs as Detectors.

• Direct LLM detection is unreliable: ChatGPT misclassifies about 50% of generated text as human-written, GPT-4 misclassifies about 95% of human text as generated, and GPT-4-Turbo remains below 50% accuracy.These findings indicate diminishing reliability for self-generated-text detection as LLM complexity increases.Source: §5.3.3 · p. 36
• In-context learning improves LLM-based detection without additional fine-tuning, and OUTFOX’s example-driven strategy outperforms traditional zero-shot methods and RoBERTa-based detectors.ICL supplies task examples in the prompt, enabling the model to learn the detection task during inference.Source: §5.3.3 · pp. 36–37

5.4 Human-Assisted Methods

• Human-assisted methods leverage human prior knowledge and analytical skills, providing notable interpretability and credibility in LLM-generated-text detection.Human classification primarily relies on visual observation of features indicative of generated text.Source: §5.4 · p. 37, §5.4.1 · p. 37

5.4.1 Intuitive Indicators.

• Humans identify LLM-generated text through coherence, consistency, relevance, detail, style, factuality, and common-sense reasoning, but grammatical errors alone are unreliable.Indicators vary by domain: generated news is often generic, stories irrelevant, and academic summaries sparse in research motivation and methodology.Source: §5.4.1 · p. 37

5.4.2 Imperceptible Features.

• GLTR supports human vetting as a visual forensic tool that exposes interpretable features and helps humans detect manipulations of statistical thresholds.Humans can recognize errors and unseen features that zero-shot statistical detectors may miss or that adversaries may manipulate.Source: §5.4.2 · p. 37

5.4.3 Enhancing Human Detection Capabilities.

• Manual annotation outperformed detection models on half of SCARECROW’s ten error types, while incentives and motivation improved annotator performance over time.Exposing evaluators to examples also enhanced detection, especially for longer samples, although revealing true boundaries did not immediately improve accuracy.Source: §5.4.3 · pp. 37–38

5.4.4 Mixed Detection: Understanding and Explanation.

• Weng et al. (2023) combine expert labeling, iterative machine-learning refinement, and visual statistical analysis to select intuitive detectors and strengthen trust.The approach also enables experts to learn from model behavior while efficiently identifying LLM-generated samples.Source: §5.4.4 · p. 38

Confusion Matrix

• The confusion matrix defines four outcomes—TP, TN, FP, and FN—that underpin the evaluation metrics for LLM-generated text detection.TP and TN are correct classifications of machine- and human-written text, while FP and FN are the corresponding misclassifications.Source: §6 · p. 38

Accuracy

• Accuracy measures correct classifications overall but is less useful on unbalanced datasets, while paired and unpaired accuracy assess detection in matched and independent test settings.In paired evaluation, the detector compares real and generated samples sharing metadata; unpaired evaluation requires independent classification, making it relatively more difficult.Source: §6 · pp. 38–39

Precision

• Precision emphasizes the correctness of LLM-generated predictions and is important when false positives could damage user trust or create business harm.It is the proportion of correctly detected LLM-generated samples among all samples classified as LLM-generated.Source: §6 · p. 39

Recall

• Recall measures correctly identified LLM-generated texts, whereas HumanRec, LLMRec, and AvgRec expose class-specific and mean performance across categories.Recall is especially useful when underreporting machine-generated text should be minimized.Source: §6 · p. 39

AUROC

• AUROC evaluates discrimination across varying thresholds using true- and false-positive rates, which is especially useful for unbalanced binary detection and threshold-sensitive zero-shot methods.It supports evaluation when particular false-positive or miss rates matter and is commonly used to assess zero-shot detection across possible thresholds.Source: §6 · pp. 40–41

7.1 Out of Distribution Challenges

• Out-of-distribution performance deteriorates across domains, languages, and unknown LLMs, with some classifiers approaching random classification and the OpenAI detector reaching only 74.74% AUROC on GPT-3.5-Turbo and GPT-4 text.Cross-lingual failures can disproportionately penalize non-native or non-standard linguistic styles, while surrogate-model comparisons, model identification, and diverse training data can improve cross-LLM robustness.Source: §7.1 · pp. 41–42

7.2 Potential Attacks

• Paraphrase, adversarial, prompt, and training-based attacks can sharply reduce detector accuracy, although adversarial training and some detector designs provide partial resilience.Adversarial perturbations reduced fine-tuned classifier, watermarking, and DetectGPT performance by more than 18%, 10%, and 25%, respectively; Llama-2-7B fine-tuning reduced OpenAI RoBERTa-Large AUROC from 0.84 to 0.62.Source: §7.2 · pp. 43–44

7.3 Real-World Data Issues

• Real-world text may mix human and LLM-written content or contain human edits, requiring fine-grained detection while remaining difficult for current detectors, especially on short or paraphrased text.Data ambiguity also risks mislabeling LLM-generated training data as human-written, weakening detector research and potentially recycling synthetic data into future model training.Source: §7.3 · p. 45

7.4 Impact of Model Size on Detectors

• Larger generative models produce harder-to-detect text, whereas larger supervised detectors improve in-distribution performance but may generalize less effectively.Training on medium-sized models can generalize to larger versions without additional samples, while aligning training and test model distributions improves detection.Source: §7.4 · p. 46

7.5 Lack of Effective Evaluation Framework

• Detector results vary across independently constructed test sets because generation, environment, distribution, and preprocessing choices undermine reproducibility, making objective shared benchmarks essential.The survey advocates comprehensive evaluation frameworks and openly shared test sets to establish consistent and fair detector comparisons.Source: §7.5 · p. 46

8.1 Building Robust Detectors with Attacks

• Detectors should be tested against diverse attack types and comprehensive benchmarks, because existing robustness studies and benchmarks cover limited threats.Relevant attacks include paraphrase, adversarial, and prompt attacks; current methods often address only specific forms.Source: §8.1 · p. 47

8.2 Enhancing the Efficacy of Zero-shot Detectors

• Black-box zero-shot detectors merit further study because they offer stability, interpretability, and broader applicability than white-box methods.Future work should identify low- to high-dimensional distinctions between LLM-generated and human-written text to improve discrimination and decision justifications.Source: §8.2 · p. 47

8.3 Optimizing Detectors for Low-resource Environments

• Detector research should quantify performance across data volumes and identify minimum resource requirements, while evaluating domain, language, and attack adaptability.These analyses can guide dependable detector selection in resource-constrained environments.Source: §8.3 · p. 47

8.4 Detection for Not Purely LLM-Generated Text

• Researchers should organize datasets and define tasks for detecting mixed-source and human-edited text, which may be the most common application setting.Potential solutions for these real-world cases remain pending verification.Source: §8.4 · pp. 47–48

8.5 Constructing Detectors Amidst Data Ambiguity

• Future studies should verify the authenticity of real-world training data to prevent unrecognized LLM-generated text from compromising detector research.Contaminated blogs and web comments can create a detrimental feedback loop in detector development.Source: §8.5 · p. 48

8.6 Developing Effective Evaluation Framework Aligned With Real-World

• Evaluation should reflect practical detection needs through fair frameworks that account for application domains, black-box text generation, attacks, and post-editing.Existing test sets may favor their creators’ detectors, while benchmarks often represent idealized scenarios.Source: §8.6 · p. 48

8.7 Constructing Detectors with Misinformation Discrimination Capabilities

• Detectors should assess factual veracity alongside text-generation signatures to improve source attribution and help mitigate misinformation.Promising directions include integrating LLM-based factual detection with external knowledge bases or search engines (Gao et al. 2023; Chern et al. 2023; Asai et al. 2023; Liang et al. 2023b).Source: §8.7 · p. 48

Conclusion

• LLM-generated text detectors are crucial defenses against deceptive news, scams, and educational inequality caused by misuse of LLMs.The conclusion frames detection as essential because LLM-generated text has become pervasive in daily life.Source: §9 · p. 49
• The survey reviews the detection task, sources of LLM capabilities, popular and promising datasets, detector challenges, and future research directions.It highlights out-of-distribution data, potential attacks, real-world data issues, and the lack of an effective evaluation framework as critical limitations.Source: §9 · p. 49

Applications and evaluation

• Application-oriented references examine detection in essays, tweets, fake science, misinformation, academic authorship, and conversational AI across university courses.The cited studies include deepfake tweet detection, ChatGPT authorship attribution, fake-science detection, machine-made media monitoring, and conversational AI detectability.Source: 2023. A deep fusion model for human · pp. 52, 53, 55

Detection methods

• Referenced detection approaches include statistical visualization, unsupervised distributional methods, Bayesian surrogate modeling, adversarial learning, denoising, and emotion recognition.The references include GLTR, unsupervised detection, Bayesian detection, RADAR, AuthentiGPT, and synthetic-text detection via emotion recognition.Source: 2023. A deep fusion model for human · pp. 52–54

Robustness and watermarking

• The references also address watermarking and robustness, including watermark reliability, paraphrastic robustness, backdooring, and adversarial robustness of neural-statistical detection features.Cited works include Kirchenbauer et al. (2023a, 2023b), SemStamp, Gu et al. (2022), and Crothers et al. (2022).Source: 2023. A deep fusion model for human · pp. 52, 54, 55

Related work

• Related studies examine detector robustness through paraphrasing, soft prompts, mutation-based attacks, adversarial optimization, and guided evasion.Examples include Karpinska, Wieting, and Iyyer (2023), Kumarage et al. (2023b), Liang, Guerrero, and Alsmadi (2023), Lu et al. (2023), and Nicks et al. (2023).Source: 2023. A deep fusion model for human · pp. 56, 57, 59
• Watermarking research develops robust distortion-free, private, semantic-invariant, deep-learning-based, and code-generation watermarking methods for detecting or identifying large-language-model-generated content.Source: 2023. A deep fusion model for human · pp. 56–58
• Applications address detecting generated content in journalism, education, computing, academia, and scientific or medical communication.Examples include J-guard (Kumarage et al., 2023a), computing-education detection (Orenstrakh et al., 2023), academic-writing detection (Liu et al., 2023d), and medical-text differentiation (Liao et al., 2023a, 2023b).Source: 2023. A deep fusion model for human · pp. 56, 57, 59
• Several works study detector behavior under distribution shift and explainability, including zero-shot generalization, coherence-enhanced detection, linguistic patterns, and sentence-level decisions.These directions are represented by Pu et al. (2023b), Liu et al. (2022), Muñoz-Ortiz et al. (2023), and Mitrovic, Andreoletti, and Ayoub (2023).Source: 2023. A deep fusion model for human · pp. 57–59
• The references cover statistical, stylistic, neural, explainable, multilingual, authorship, and zero-shot approaches to detecting machine-generated text.Examples include DetectLLM (2023a), GPT-who (2023), Ghostbuster (2023), M4 (2023b), and LLmDet (2023).Source: 2023. A deep fusion model for human · pp. 60–62
• The references also document evaluation resources, attacks, human-assisted detection, and real-world applications such as education, medicine, scientific writing, and university homework.Examples include TURINGBENCH (2021), HANSEN (2023), detector red teaming (2023), human collaboration (2023), and context-aware homework detection (2023).Source: 2023. A deep fusion model for human · pp. 61–62
• The bibliography includes watermarking and provenance-tracing methods covering black-box language models, robust invariant features, context-aware lexical substitution, Unicode homoglyph substitution, and synonym substitutions.Examples include Liu et al. (2023a), Kwak (2023), Zhang et al. (2022), Montesi (2016), and Atallah (2006).Source: 2023. A deep fusion model for human · pp. 60, 61, 63
• The bibliography cites CHEAT, a large-scale dataset for detecting ChatGPT-written abstracts, and GPT paternity test, which detects GPT-generated text through genetic inheritance.The supplied entries identify these works but provide no methods, experiments, or results beyond their titles and citations.Source: 2023. A deep fusion model for human · pp. 63–64
• Additional references cover Wordcraft for story writing with large language models, HellaSwag for sentence completion, and neural defenses against fake news.These entries are bibliographic references and do not state comparative findings in this part.Source: 2023. A deep fusion model for human · p. 64
• The cited literature also includes automatic human–AI essay boundary detection, robust LLM watermarking, hallucination surveys, few-shot calibration, multilingual code generation, and factual-structure deepfake detection.The supplied passages list these topics through publication titles without reporting their technical details or results.Source: 2023. A deep fusion model for human · p. 64
• A further reference presents zero-shot LLM-generated text detection by querying ChatGPT.No evaluation outcome or comparison is provided in the supplied citation entry.Source: 2023. A deep fusion model for human · p. 64