Aug 2026· Engineering, Technology & Applied Science Research· Vol 16, pp. 38145-38151· 0 citations· 27 references
TL;DR
This systematic literature review synthesizes peer-reviewed and high-quality studies published between 2023 and 2026 on AI-obfuscated, AI-refined, and humanized text and suggests that AI-text detection should be treated as one supportive signal rather than a stand-alone judgment, particularly in high-stakes academic or professional contexts.
Abstract
Large Language Models (LLMs) are now widely used to draft, revise, paraphrase, and polish text, making the detection of AI-generated writing increasingly difficult. This systematic literature review synthesizes peer-reviewed and high-quality studies published between 2023 and 2026 on AI-obfuscated, AI-refined, and humanized text. From 1,002 records, 26 primary studies were retained after screening and quality assessment. The review organizes the literature through a seven-dimensional taxonomy. Overall, the evidence shows that many detectors perform well on clean or in-distribution AI-text but become less reliable when the text is paraphrased, humanized, or collaboratively edited. The review also highlights recurring fairness concerns, especially for non-native English writers, and finds that current benchmarks often do not fully capture realistic mixed-authorship and adversarial settings. These results suggest that AI-text detection should be treated as one supportive signal rather than a stand-alone judgment, particularly in high-stakes academic or professional contexts.
It is concluded that AI-text detection, in its current form, cannot serve as a sole, dispositive basis for academic-integrity decisions, and a set of institutional and technical recommendations are proposed that better address the underlying problem than detector accuracy alone can.
A. Kumar, Ayush Kumar, Danish Maqbool Chopan· International Journal of Res...· 0 citations
With the widespread adoption of large language models in academic writing, the detection of artificial intelligence-generated content has become an important research topic. This paper reviews the detection of AI-generated text in academic papers. It defines the relevant concepts, surveys detection methods and commercial detection tools, and introduces experimental datasets and evaluation metrics. The review shows that current detection approaches face several challenges, including a shortage of suitable datasets, ambiguous labels for human–AI collaborative text, insufficient coverage of generative models, limited generalization, and inadequate interpretability. Future research should develop datasets covering multiple languages, disciplines, models, and generation methods; strengthen evaluation on unseen models and adversarial samples; and improve explainable detection and human-review mechanisms, thereby supporting the governance of academic integrity.
Institutions use commercial AI detectors for academic integrity, yet detectors cannot distinguish AI editing from full LLM drafts and may treat both as misconduct. In a controlled study of published English abstracts (four domains; 2013 to 2015 vs. 2023 to 2025), we quantify this policy failure under proxy human/AI labels at tau=0.50. Light"refine abstract only"edits, a proxy for guideline-compliant AI assistance, are flagged at 38 to 80%. Unmodified 2023 to 2025 originals are flagged at 9 to 15%, with non-STEM rates far above STEM (p<0.001); elevated scores track long-token and Academic Word List density, not authorship intent alone. After Undetectable AI humanization, evasion is near-total: fewer than 4% of AI-labeled rewrites remain flagged (post-humanization detection rate<4%; FNR>96%). Honest AI-editing results in a higher sanction risk than humanizer-assisted evasion. Therefore, detector scores should not serve as standalone misconduct evidence.
Jonathan A. Karr, Grigorii Khvatskii, T. Hua et al.· 0 citations
Dear Editor,
The rapid integration of AI into academic writing has necessitated tools for detecting AI-generated content such as iThenticate ZeroGPT, Turnitin, Phrasly AI, Open AI text Classifier, Writer, Copy leaks to differentiate between human and machine-generated text.1 However, current AI detection tools suffer from significant misclassification rates, generating false positives that wrongly accuse human authors and false negatives that allow AI-generated text to evade detection.2 This unreliability unduly impacts non-native English speakers, who often utilise AI tools for paraphrasing and grammar correction to ensure their work meets academic standards; however, these legitimate linguistic refinements are frequently misconstrued by detection software as evidence of AI-generated content.3 This problem is heightened by the fact that many AI detection tools, primarily trained on English corpora, struggle to accurately assess texts with diverse linguistic structures and stylistic conventions, thereby increasing the risk of false positives for non-English speaking scholars.2 Such false positives, where human-written manuscripts are incorrectly labelled as AI-generated , severely threaten scholarly psychological safety by fostering distrust and creating an environment of anxiety and unfair allegations. This systemic issue fundamentally challenges academic integrity, undermining the credibility of authors and the foundation of human-based scholarly writing.4,5 This editorial highlights the urgent need for human centric approach to AI detection, advocating for strategies that prioritise human-AI collaboration over sole reliance on fallible automated systems.
A fundamental flaw of current AI detectors lies in their documented inconsistency and unreliability. Studies consistently demonstrate high rates of both false positives, where human-written text is erroneously identified as AI-generated, and false negatives, where AI-generated content evades detection.6,7 For instance, literature indicated that both free and commercially available AI detection tools can incorrectly classify human-written content as AI-generated with rates ranging from 43.3% to 83.3%.7,8
The ethical concerns arise directly because of incorrectly labelling human-written manuscripts as AI-generated and vice versa. Such errors lead to unfair allegations, rejection of genuine work, unwarranted accusations of academic misconduct, and reputational damage for authors.9,10 When authors must modify their writing or use "humaniser" tools to avoid false detection, it paradoxically increases AI involvement and further obscures human-machine authorship boundaries.8 Furthermore, the ease with which AI-generated text can be altered allows it to bypass current detection methods, turning the process into a counterproductive 'cat-and-mouse' game.1 This eventually weakens the very goal of identifying AI misuse while simultaneously unjustly burdening diligent human scholars.4
---Continue
Plagiarism-detection and AI-detection tools are now widely used in academic publishing. These systems were introduced to support research integrity by helping journals identify potential plagiarism, inappropriate text reuse, and concerns related to undisclosed use of artificial intelligence (AI). When used appropriately, they can serve as useful screening tools and support editorial decision-making. However, their growing use has also created new challenges. Similarity scores are often interpreted as direct measures of plagiarism, even though they only indicate matching text and require contextual evaluation. Likewise, AI-detection tools can produce uncertain or incorrect classifications, yet their output may influence perceptions of authorship and manuscript quality. As a result, researchers may spend considerable time reducing similarity scores or worrying about AI-detection reports, even when the underlying writing is appropriate. These challenges may be particularly relevant for early-career researchers and authors writing in a second language. This article discusses the benefits and limitations of both similarity-detection and AI-detection systems and argues that their output should be viewed as screening indicators rather than definitive judgments. Human interpretation should remain central to the evaluation of originality, authorship, and research quality.
The proposed Recursive Self-Correction approach raises model performance from a Political Neutrality Likert scale baseline of 2.14 to 4.56, averaged across all models, demonstrating effective inference-time mitigation of political bias in LLM-generated summaries.
Tejaswi V. Panchagnula, Bruce Coburn, Bryce J. Dietrich et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
