In brief: AI-generated text detection is the attempt to tell whether a piece of writing was produced by a large language model, a task usually framed as binary classification (Wu et al., 2025). A detector returns a probability or percentage, not proof: in one test, Turnitin’s AI detector gave all 160 student-written responses a score of zero (Gosling et al., 2024), yet in a test on rehabilitation articles it identified only 30% of ChatGPT-written articles once they had been rephrased with Wordtune (J. Q. J. Liu et al., 2024). The same percentage can therefore mean very different things depending on the text, the tool and how common AI use is.
AI-generated text detection and Turnitin: an introduction (101)
This guide introduces the detection of AI-generated text for students, teachers, supervisors, editors and reviewers. It is built on 185 records (180 articles and 5 reviews) from the Web of Science Core Collection, published from 2021 to 2027, and every empirical statement is traced to one of them. It covers definitions, how detectors work, tool-by-tool accuracy (including Turnitin and GPTZero), false positives, bias against non-native writers, evasion through paraphrasing and “humanizer” tools, human reviewers, policy for education and scientific publishing, Türkiye, debates and a practical guide for handling an “AI detected” result. Results refer to the tool versions and models named in each study and may not hold for newer ones.
Try it yourself: Decide what to do with six fictional cases that carry an AI score, see how base rates turn a small false positive rate into many false accusations, and check a course or journal policy against four fairness domains.
Open the interactive module: AI Detection Lab
Learning objectives
By the end of this guide you should be able to:
- Define AI-generated text detection and distinguish it from plagiarism detection, watermarking, stylometry and authorship judgement.
- Explain how classifiers, perplexity and burstiness measures, watermarks, stylometry and human readers try to separate human from machine text, and where each fails.
- Appraise accuracy studies of Turnitin, GPTZero and other tools, separating high, low and mixed results.
- Analyse false positives, bias against second language writers, edited and hybrid texts, and evasion.
- Apply a step-by-step procedure and decision table when a paper or manuscript is flagged as AI-generated.
Key concepts
- AI-generated text detection. A binary classification task that seeks to determine whether a given text was produced by a large language model (Wu et al., 2025).
- AI score. Online tools report a percentage score of AI use (Cheng et al., 2025); GPTZero scans for statistical signatures of machine generation and returns a probability estimate (Giray, Sevnarayan, & Maphoto, 2026).
- False positive and false negative. A false positive classifies a human-written text as machine-generated, a false negative the reverse (Habibzadeh, 2023); because score distributions for genuine and AI texts overlap, both errors are possible (Dalalah & Dalalah, 2023).
- Perplexity. A measure derived from a language model’s next-token predictions (Miralles-González et al., 2026); machine-generated text tends to show lower perplexity (Ahmad et al., 2026).
- Burstiness. A computed linguistic feature used, with lexical diversity and textual complexity, to separate human from generated texts (Nitu & Dascalu, 2024).
- Watermarking. Embedding imperceptible signals during generation through token selection so that the text can later be identified (Zhang et al., 2025).
- Humanised AI text. AI-generated text rewritten with paraphrasing tools such as GPT-Humanizer or QuillBot (Hua & Yao, 2024); YouTube videos promote such rewriters to bypass detectors (Foltynek & Newton, 2025).
- Hybrid text. Text that combines AI-generated and human-written content (X. Liu et al., 2025).
Contents
- How this guide was produced; introduction and history
- Definitions and neighbouring concepts; how detectors work
- Accuracy evidence; false positives, bias, evasion and human reviewers
- Policy and integrity in education and scientific publishing
- Themes and bibliometric profile; AI text detection in Türkiye
- Debates, research gaps and implications for practice
- Summary, discussion questions, further reading, glossary, frequently asked questions and references
How this guide was produced
Database and date. Web of Science Core Collection, Advanced Search, 18 November 2026; no language or year limits; articles and reviews.
Query (verbatim). TI=("AI detect*" OR "AI-generated text" OR "AI text detect*" OR "AI content detect*" OR "AI writing detect*" OR "AI-generated content" OR Turnitin OR GPTZero OR "detecting AI" OR "detection of AI" OR "detect AI" OR "machine-generated text" OR "AI-generated writing" OR "AI-written" OR "ChatGPT-generated" OR "LLM-generated text" OR "generated by ChatGPT") AND TI=(detect* OR identif* OR distinguish* OR plagiarism OR integrity OR Turnitin OR GPTZero OR classifier* OR authorship OR "false positive*") AND DT=("Article" OR "Review")
Criteria. Included: records that substantively address the detection of AI-generated text (Turnitin’s AI indicator, GPTZero, other detectors or human reviewers), including methods, accuracy, errors, bias, evasion, integrity and policy. Excluded: detection of AI-generated images, audio, video or deepfakes only, or “AI detection” meaning AI used to detect a disease or object; unrelated records, such as classic plagiarism detection without AI text; protocols without results; editorials without substance.
Flow. See Figure 1: 339 identified, 1 duplicate removed, no retracted publications, 338 screened, 153 excluded (98 other meanings of “AI detection”, 55 unrelated), 185 included (180 articles and 5 reviews).
Screening and extraction. AI-assisted screeners worked on titles and abstracts; the lead reviewer resolved 2 unsure records. Evidence was extracted from the abstracts of a subset of 175 records into a structured template; bibliometric counts cover all 185 records.
Lessons from screening. “AI detection” is a homonym: 98 of the 153 exclusions concerned deepfakes or AI that detects a disease or object. Studies that build a new detector were kept, but their in-house benchmark figures are reported separately from independent tests of commercial tools.
Limitations. Single database; title search; abstract-level reading without quality appraisal; AI-assisted screening; English-language bias of Web of Science. The underlying studies test specific tool versions and models at one point in time, often on small or synthetic text sets.
Introduction: why AI text detection matters
Imagine marking essays after the public release of ChatGPT. Among 153 students in an introductory microbiology course, 46.9% used large language models in their college work, 39% admitted using them to answer assessments and 7% to write entire papers (Paustian & Slinger, 2024). Telling the difference by eye is hard: in a paired-essay test, 140 college instructors picked the ChatGPT essay only 70% of the time (Waltzer et al., 2024). In publishing, ChatGPT writes believable scientific abstracts with completely generated data (Gao et al., 2023), and many publishers reject language models as authors (Desaire et al., 2023).
Institutions responded with software. Many universities have implemented Turnitin AI, GPTZero, Copyleaks or ZeroGPT (Das Deep et al., 2025), and many teachers turned to AI detection as a first-line response (Giray, Sevnarayan, & Maphoto, 2026). This guide concentrates on education and scholarly publishing, where a score can trigger a misconduct case. The central difficulty, as one position paper argues, is that unlike plagiarism detection, which matches verifiable sources, AI detection relies on unverifiable probabilistic estimates (Bassett et al., 2026).
History and development
The oldest record, from 2021, is a deep learning model for detecting machine-generated text in a Chinese biomedical question-and-answer dataset (Wang & Feng, 2021). The field changed scale after OpenAI released ChatGPT to the public in late 2022 (Paustian & Slinger, 2024): 38 records appeared in 2023 and 2024, and 144 in 2025 and 2026 (Figure 5).
Early studies, published in 2023, found tools more accurate on GPT 3.5 than on GPT 4 content (Elkhatat et al., 2023), and a broad test of 12 public tools plus Turnitin and PlagiarismCheck judged them neither accurate nor reliable (Weber-Wulff et al., 2023). Regulation followed: obligations arising mainly from the EU AI Act, enhanced by a US Executive Order, require providers of AI generators to support reliable detectors (Knott et al., 2024). Turnitin added an AI content detector to its plagiarism platform (Ibrahim et al., 2025), and comparisons over time began: 6% of 2023 doctoral dissertations had Turnitin AI scores over 20%, against none above 20% in 2013 (Murch et al., 2025).
By 2025 and 2026, studies tested newer models (GPT-4o, GPT-4.5, DeepSeek-R2, Gemini, Grok) and harder conditions such as hybrid, humanised and AI-polished texts (Almohaimeed et al., 2026; Cheng et al., 2025; Makhmutova et al., 2026; Rosenthal et al., 2026; Van Vlasselaer et al., 2026; Villa et al., 2026). Attention also turned to interpretation: threshold frameworks (Rosenthal et al., 2026), the design of detection reports (Du et al., 2026) and process-based alternatives such as short vivas piloted in 2025 and 2026 (Hutson et al., 2026). Figure 2 summarises milestones stated in the corpus.
Definitions
Definitions differ in what they emphasise: the technical task, the source of the signal, the social purpose or the meaning of the output (Table 1).
| Source | Year | Discipline | Definition (paraphrased) | Emphasised dimension |
|---|---|---|---|---|
| Crothers et al., 2023 | 2023 | Computer science (survey) | A key countermeasure against abuse of natural language generation, set in its cybersecurity and social context | Threat models |
| Wu et al., 2025 | 2025 | Computational linguistics (survey) | A binary classification task: was a given text produced by a large language model? | Binary classification |
| Xiang et al., 2026 | 2026 | Computer science (review) | Passive detection uses intrinsic textual features; active detection uses embedded signals such as watermarks | Source of the signal |
| Knott et al., 2024 | 2024 | Ethics and information technology | Reliable ways of identifying AI content that supplement the institutions enabling social trust | Trust and regulation |
| Giray, Sevnarayan, & Maphoto, 2026 | 2026 | Writing assessment | GPTZero scans for statistical signatures of machine generation and returns a probability score | Probability estimate |
| Bassett et al., 2026 | 2026 | Higher education policy | Unlike source-matching plagiarism detection, AI detection rests on unverifiable probabilistic estimates | Verifiability |
| Tamuka, 2026 | 2026 | Computing | AI-written and AI-rewritten passages differ; a single AI-likelihood score says little about which applies | Type of AI involvement |
| Maphoto et al., 2026 | 2026 | Second language writing assessment | Turnitin’s score repositioned from a verdict of misconduct to a formative cue for dialogue | Pedagogical use |
Two points follow. The object of detection is no longer simply “AI versus human”: texts may be polished by AI, with the degree of modification measurable as a “Polish Ratio” (Yang et al., 2024), partly AI-written (X. Liu et al., 2025) or rewritten after generation (Tamuka, 2026). And the output is a score whose meaning depends on its use. Figure 3 places these elements in sequence.
Distinguishing AI text detection from neighbouring concepts
Several terms are used loosely in policies and media reports. Table 2 separates them; Figure 4 compares them on six features.
| Concept | What it is | Difference from AI text detection | Source |
|---|---|---|---|
| Plagiarism (similarity) detection | Matching a text against existing sources | Finds copied text; AI-generated abstracts were 100% unique in one study and highly original in another | Erol et al., 2025; Makiev et al., 2023 |
| Turnitin AI indicator | An AI detector offered as an option with Turnitin, a widely used plagiarism platform | One commercial implementation, with mixed independent results | Gosling et al., 2024; Ibrahim et al., 2025 |
| GPTZero | A platform returning a probability that a text is machine-generated | One tool; empirical findings on its efficacy are mixed | Giray, Sevnarayan, & Maphoto, 2026 |
| Watermarking | Signals embedded during generation through token selection | Active detection that depends on a signal placed by the generator | Xiang et al., 2026; Zhang et al., 2025 |
| Stylometry | Distance measures between stylistic features such as n-grams | A text attribution approach applied without training | Salnikov & Bonch-Osmolovskaya, 2025 |
| False positive | A human-written text classified as machine-generated | An error type, not a method; it risks falsely accusing authors | Almohaimeed et al., 2026; Habibzadeh, 2023 |
| Academic integrity | Norms and procedures protecting honest scholarship | A normative goal; one position paper argues that investigations need evidence meeting the balance of probabilities, which detection scores do not | Bassett et al., 2026 |
| Authorship | Who produced and is responsible for a text | One position paper argues that a human/AI split ignores work created with, not by, AI | Bassett et al., 2026 |
How detectors work
Surveys group detectors as watermarking, statistics-based, neural-based and human-assisted methods (Wu et al., 2025), or split passive methods using intrinsic features from active methods using embedded signals (Xiang et al., 2026). Table 3 summarises the families.
Trained classifiers. Many new detectors are supervised models trained on labelled human and machine texts, from a random forest on TF-IDF features (Pantic & Mugosa, 2025) to fine-tuned transformers such as RoBERTa (Oghaz et al., 2025). Some are built for one domain, such as graduate admission letters, because a general, universally effective model appears beyond current methods (Zhao et al., 2024). Their weakness is generalisation: fine-tuned detectors reached an AUROC of 1.0000 in distribution, but domain shift dropped two code detectors to 0.85 and 0.67, and matched train-test conditions can overestimate real-world reliability (Alluri et al., 2026). In e-commerce reviews, all detectors collapsed under domain shift, with AUC falling to 0.48, although the authors’ training-free residual score reached an AUC of 0.989 and transferred to six unseen generator families at 0.83 (Balci et al., 2026).
Perplexity, burstiness and other statistical signals. These methods ask how predictable a text is to a language model. Machine text shows signatures such as lower perplexity and simpler morphology (Ahmad et al., 2026), and lower entropy with higher compressibility (Lv & Sun, 2026). Perplexity has been combined with linguistic features to detect ChatGPT essays (Jiang et al., 2024), and burstiness with complexity indices for Romanian texts (Nitu & Dascalu, 2024). Feature-based detectors are particularly vulnerable to paraphrasing (Rydell et al., 2026).
Watermarking. One framework embeds imperceptible watermarks through entropy-aware token selection and reports 95.4% detection accuracy (Zhang et al., 2025). Because watermarking relies on signals embedded at generation (Xiang et al., 2026), it depends on generator providers; new obligations require providers to support reliable detectors for their content (Knott et al., 2024), and implementation challenges persist (Das Deep et al., 2025).
Stylometry. A trigram-cosine stylometric delta clustered Russian human and model diary texts without training (Adjusted Rand Index about 0.70 against about 0.28 for a fine-tuned RuModernBERT), though validation is limited to Russian diaries and selected models (Salnikov & Bonch-Osmolovskaya, 2025).
Human judgement and process evidence. Teachers combine linguistic and content-related strategies (De Wilde, 2024); a think-aloud study with teachers in Bahraini higher education describes detection as a professional judgement task rather than a purely technical one (Teraif, 2026). Process-based approaches pair a written artefact with a short adaptive viva, as a diagnostic layer rather than a detector (Hutson et al., 2026), or share instructors’ own AI-generated “pseudo-AI submissions” so that look-alike code is easy to flag (Bashir, 2025).
| Approach | What it measures | Known weakness | Citation |
|---|---|---|---|
| Trained classifier | Patterns learned from labelled human and AI texts | Falls under domain shift; matched tests overestimate reliability | Alluri et al., 2026; Balci et al., 2026 |
| Commercial score (for example Turnitin, GPTZero) | A percentage or probability of AI generation | Scores for the same text varied from 5% to 67% across six (unnamed) systems | Giray, Sevnarayan, & Maphoto, 2026; Savelev et al., 2026 |
| Perplexity, burstiness, entropy | How predictable or regular the text is | Feature-based detectors particularly vulnerable to paraphrasing | Ahmad et al., 2026; Rydell et al., 2026 |
| Watermarking | A signal embedded during generation | Needs an embedded signal; implementation challenges | Das Deep et al., 2025; Xiang et al., 2026 |
| Stylometry | Distances between stylistic features | Validated on few languages; simple features vulnerable to attacks | Salnikov & Bonch-Osmolovskaya, 2025; Sciandra et al., 2026 |
| Human reviewers | Judgements of language, content and context | Close to chance in several studies; confidence misleads | Cheng et al., 2025; Fiedler & Döpke, 2025; Milicka et al., 2025 |
| Process evidence | Whether the writer can explain the work | Pilot evidence only, mixed student perceptions | Hutson et al., 2026 |
Accuracy evidence, tool by tool
Two kinds of study should not be mixed. Studies that build a new detector and test it on benchmarks usually report very high figures; a survey of 44 empirical studies reports near-perfect accuracy of about 100% for transformer models (Ahmad et al., 2026). Independent tests of commercial or free tools on realistic texts range from near-perfect to worse than chance, depending on text type, model, tool version and editing (Table 4).
Turnitin. It gave all 160 human class responses a score of zero, while 160 ChatGPT responses scored significantly higher (Gosling et al., 2024). Across 81 scripts mixing human and LLM words, it gave no AI score to human scripts and its percentage rose with the actual LLM share, but low LLM shares received inaccurately high scores and high shares inaccurately low ones (Atamhenwan, 2026). It misclassified no human rehabilitation article but identified only 30% of AI-rephrased ones (J. Q. J. Liu et al., 2024). On 120 dentistry manuscripts its performance was moderate and four detectors outperformed it (Villa et al., 2026); on 192 texts including EFL student writing its accuracy was 0.61 against 0.69 for Originality (Hadra et al., 2026); and it underestimated GenAI content compared with Pangram (Van Vlasselaer et al., 2026).
GPTZero. On medical texts it had a sensitivity of 0.65 and a specificity of 0.90, a low false positive and high false negative rate (Habibzadeh, 2023). It showed the highest discrimination accuracy (90.0%) on 150 dental abstracts (AL-Rawas et al., 2026) but 63% accuracy on Arabic texts (Alshammari et al., 2024). A critical article notes that empirical findings on GPTZero are mixed, from strong diagnostic value under controlled conditions to near-random performance on some AI essays (Giray, Sevnarayan, & Maphoto, 2026).
Other tools. Originality.ai detected 100% of ChatGPT-generated and AI-rephrased rehabilitation texts (J. Q. J. Liu et al., 2024); Copyleaks and Originality.ai were near-perfect in a test of nine tools, while some free tools fell as low as 63.0% (Makhmutova et al., 2026). A free detector labelled a median 27.2% of human behavioural health text as AI-generated (Popkov & Barrett, 2025).
What changes the result. The model: tools did better on GPT 3.5 than GPT 4 content (Elkhatat et al., 2023), and DeepSeek-R2 manuscripts were easier to detect than GPT-4.5 or GPT-4o ones (Villa et al., 2026). The text type: 13 detectors performed acceptably on long theses but failed systematically on engineering code and short coursework (Sun et al., 2026). The tool: three tools discriminated well (AUC 0.962) but agreed poorly (intraclass correlation -0.118) (Fine et al., 2026), whereas five other tools agreed highly (ICC 0.92) (Rosenthal et al., 2026).
| Study | Tools | Texts | Main result | Citation |
|---|---|---|---|---|
| High accuracy reported | ||||
| Class assignment | Turnitin AI detector | 160 student vs 160 ChatGPT responses | All human responses scored zero; AI responses scored significantly higher (no detection rate given) | Gosling et al., 2024 |
| Anatomy and physiology essays | Four online detectors | 50 student and AI essays | Best three detectors 93% to 98% correct; false positives about 1.3% | Hyatt et al., 2025 |
| Low accuracy reported | ||||
| Academic setting | 12 public tools, Turnitin, PlagiarismCheck | Human vs ChatGPT, with obfuscation | Neither accurate nor reliable; biased towards “human-written” | Weber-Wulff et al., 2023 |
| Orthopaedic abstracts | Two detectors | 21 abstracts | 42.9% and 66.6% correctly identified | Makiev et al., 2023 |
| German thesis excerpts | Detectors and 63 lecturers | 200 to 300 words | Both only slightly better than chance | Fiedler & Döpke, 2025 |
| Scientific papers | Six systems | Academic texts | Same text scored 5% to 67% | Savelev et al., 2026 |
| Mixed results | ||||
| Dentistry manuscripts | Six detectors incl. Turnitin | 120 (GPT-4.5, GPT-4o, DeepSeek-R2, human) | Four high; Turnitin moderate; DetectingAI ineffective; no false positives among 30 human texts at 60% | Villa et al., 2026 |
| Academic excerpts | Five tools | 500: human, AI, humanised, hybrid | AUC 0.994 to 1.000 for human vs AI; substantial decline for hybrid and humanised | Rosenthal et al., 2026 |
| Medical texts | GPTZero | 20 ChatGPT vs 30 published | Sensitivity 0.65; specificity 0.90 | Habibzadeh, 2023 |
| EFL-focused higher education | Turnitin, Originality | 192 incl. EFL writing and hybrids | Accuracy 0.61 vs 0.69; both poor on hybrids | Hadra et al., 2026 |
| Student work at scale | 13 detectors | Over 280,000 coursework, thesis and code samples | Acceptable on theses; failures on code and short tasks | Sun et al., 2026 |
False positives, bias and evasion
A false positive harms an honest writer. Some tests found almost none: zero in a Turnitin class test (Gosling et al., 2024), none among 30 human dentistry texts at a 60% threshold (Villa et al., 2026), and rare false positives across four tools, which the authors read as an improvement on earlier studies (Van Vlasselaer et al., 2026). Others found many: false positives on human-written manuscript passages were frequent at optimal cut-offs of 25% to 50% (Fine et al., 2026), ZeroGPT and GPTZero gave high false positive rates on human abstracts in Persian-language library and information science journals (Shamsi et al., 2026), and literature review paragraphs were more likely than abstracts to be falsely flagged (Dalalah & Dalalah, 2023).
Non-native and multilingual writers. A systematic review found systemic biases that disproportionately affect non-native English writers (Caner, 2025), and a synthesis found frequent false positives and little transparency, especially for multilingual or non-native English speakers (Das Deep et al., 2025). Critical papers argue that the syntactic regularity and formulaic phrasing of second language (L2) writers are often misclassified by Turnitin and similar tools (Maphoto et al., 2026), and that careful L2 English can look less natural to models trained on standard English (Giray, Sevnarayan, & Maphoto, 2026). Empirical tests point in different directions: detectors built on Graduate Record Examinations essays showed no evidence of bias against non-native English speakers (Jiang et al., 2024); accuracy-bias trade-offs in GPTZero, ZeroGPT and DetectGPT disproportionately affected non-native speakers and certain disciplines (Pratama, 2025); and Originality showed a borderline trend towards higher accuracy on professional than on EFL student writing, which the authors read as a potential fairness concern (Hadra et al., 2026). Adding linguistic features to a multilingual model for Indonesian cut the false positive gap between first language and L2 writers from 0.0233 to 0.0150 (Guspian et al., 2026). Suspicion has costs: students in Uzbekistan saw strong writing as a trigger for suspicion (Mirzayev et al., 2026).
Other languages. Existing detectors were notably hampered on Arabic (Alshammari et al., 2024) and out-of-the-box solutions did not work for Romanian (Nitu & Dascalu, 2024). Slight AI polishing of human Arabic articles was especially damaging: the human-only accuracy of Originality.AI fell from 92% to 12% for articles slightly polished by Mistral (Almohaimeed et al., 2026).
Edited, hybrid and short texts. A leading detector often flagged abstracts in which GPT had only fixed grammar, although most of 271 surveyed academics thought such use need not be reported (Chemaya & Martin, 2024). Tools had limited effectiveness on AI-edited essays (Barrot & Aranda, 2025) and performed poorly on hybrid texts (Hadra et al., 2026; Rosenthal et al., 2026). Short coursework produced systematic failures, and formulaic technical writing exposed STEM disciplines to significant algorithmic bias (Sun et al., 2026); the stylistic features of academic writing drove false results in six systems (Savelev et al., 2026).
Paraphrasing and humanizer tools. Obfuscation significantly worsened the performance of the 14 tools in one test (Weber-Wulff et al., 2023). Turnitin’s accuracy dropped slightly for paraphrased and significantly for humanised text (Ibrahim et al., 2025), and fell to 45.7% in some cases of paraphrasing and non-native-style rewriting, while Copyleaks, GPTZero and Sapling held up (Makhmutova et al., 2026). Hybrid editing let up to 88% of AI content evade 13 detectors (Sun et al., 2026), and a low-resource paraphraser cut the average AUROC of six detectors from 0.9611 to 0.5981 while preserving text quality in 94% of cases (Rydell et al., 2026). YouTube videos advise students how to bypass detectors, for example by humanising text through rewriters (Foltynek & Newton, 2025). Results are not uniformly negative: a chemistry-journal detector stayed highly accurate when ChatGPT was told to write like a chemist (Desaire et al., 2023), and a fine-tuned detector kept high F1 under ten adversarial transformations (Murgia et al., 2026). Yet slight phrasing changes flipped verdicts in six tools (Wakjira et al., 2025).
Human reviewers compared with tools
Human reviewers often do no better. Blinded reviewers identified 68% of ChatGPT medical abstracts but labelled 14% of originals as generated (Gao et al., 2023); experienced shoulder and elbow reviewers misclassified 38% of originals, often because of writing style (Stadler et al., 2025); five blinded raters reached 19%, indistinguishable from chance (Cheng et al., 2025); and 24 medical raters identified origin 26.5% of the time (kappa 0.078) (Fine et al., 2026). Expertise did not reliably help: orthopaedic residents (34.9%) and professors (31.7%) did worse than a non-orthopaedic control (76.2%) (Makiev et al., 2023).
Some results favour humans. Professorial reviewers caught at least 96% of AI-rephrased articles but misclassified 12% of human ones (J. Q. J. Liu et al., 2024); human raters and the best detectors performed similarly on anatomy and physiology essays (Hyatt et al., 2025). Immediate feedback improved accuracy and confidence calibration, and without feedback participants erred most when most confident (Milicka et al., 2025).
Tools also shape human judgement. When 214 university teachers rated the same Chinese social-science course paper with a fictitious report showing an 87% rather than a 7% AI rate, perceived AI authorship rose and quality, originality, language and logic ratings fell; red highlighting also mattered, and the authors call these findings preliminary and context-specific (Du et al., 2026).
Policy and integrity: education and scientific publishing
Education. A systematic review judged detection tools unsuitable for high-stakes integrity decisions in their current form (Caner, 2025), and other authors see detectors only as indicators that prompt further inquiry, never the sole basis for misconduct decisions (Hadra et al., 2026), or as initial flags within a broader evaluation strategy (Van Vlasselaer et al., 2026). One position paper argues that detection should not be used in education at all, because outputs cannot be verified, procedural fairness is violated and scores do not meet the balance of probabilities standard (Bassett et al., 2026).
The recommendations can be grouped into four moves (the grouping is the author’s). First, treat the score as a cue for dialogue, triangulated with drafts, reflective commentaries and oral explanations (Maphoto et al., 2026) or used as a starting point for metacognitive dialogue (Giray, Sevnarayan, & Maphoto, 2026). Second, match interpretation to the decision, distinguishing screening from high-specificity diagnostic uses (Rosenthal et al., 2026); aggregating several detectors reduced false positives to nearly 0% in one STEM course (Hyatt et al., 2025). Third, protect judgement through neutral report design, independent teacher judgement, human oversight and training on automation bias (Du et al., 2026), including how to interpret unreliable results and handle insufficient evidence (Teraif, 2026). Fourth, redesign assessment and teach AI literacy: more diverse assessment methods (Foltynek & Newton, 2025), process-based assessment (Caner, 2025), AI-integrated rather than punitive pedagogies (Das Deep et al., 2025) and critical AI literacy instead of policing (Ardana & Christiana, 2026; Giray, Roe, & Diesta Espiritu, 2026).
Students and teachers at a university in Uzbekistan wanted clearer institutional guidance (Mirzayev et al., 2026), and a study of Omani college students and lecturers concluded that institutions shape integrity when they have clear AI policies and integrate AI literacy (Tayebipour et al., 2026). Policies alone are not enough: despite an AI policy, students in one online course still used AI when it was not allowed (Burden, 2026). Process-based verification is promising but untested at scale: after 52 vivas, only 14.3% of seven survey respondents agreed that the oral assessment helped them think more deeply, and 57.1% disagreed (Hutson et al., 2026).
Scientific publishing. Depending on publisher guidelines, detectors may serve as an editorial tool (Gao et al., 2023) or an additional screening tool in peer review (J. Q. J. Liu et al., 2024). Many authors add limits: tools should not be relied on alone (Cheng et al., 2025), error rates risk unnecessary penalties for human authors (Popkov & Barrett, 2025), and outputs call for standardised, cautious interpretation (Bohler et al., 2026; Porto et al., 2025). Clear guidelines are needed for authors’ reporting of AI use and for using detection software in review (Pan & Florian-Rodriguez, 2024), although enforcing nuanced reporting rules with detectors may be difficult because academics disagree on which uses must be reported (Chemaya & Martin, 2024).
What a score can and cannot show. Under controlled conditions, scores can separate fully AI-generated from fully human text well (Rosenthal et al., 2026). A score cannot show whether AI only fixed grammar (Chemaya & Martin, 2024), whether a passage was AI-written or AI-rewritten (Tamuka, 2026), or whether work was created with rather than by AI; one position paper adds that it cannot be checked against the true origin of real submissions (Bassett et al., 2026). One study therefore supports AI as decision support, not automated proof of misconduct (Tamuka, 2026), and another warns that relying solely on tools such as Turnitin’s AI detection to penalise individuals risks unjust outcomes (Popescu-Apreutesei et al., 2025).
Themes and bibliometric profile
Counts rose fast: 1 record in 2021, none in 2022, 12 in 2023, 26 in 2024, 56 in 2025 and 88 in 2026, with 2 already assigned to 2027 (Figure 5). Keyword matches in titles, abstracts and author keywords (Figure 6) are led by detector accuracy (105 records), academic integrity (84) and education (67); Turnitin appears in 18 and GPTZero in 22. A record can match several groups, so the counts overlap.
The most frequent Web of Science categories are Computer Science, Artificial Intelligence (53); Computer Science, Information Systems (44); Education & Educational Research (32); Computer Science, Interdisciplinary Applications (20); Linguistics (17); Engineering, Electrical & Electronic (15); Computer Science, Theory & Methods (11); Multidisciplinary Sciences (11); Language & Linguistics (8); and Surgery (8). Twelve journals published three or more records: Scientific Reports (8), Neurocomputing (7), International Journal for Educational Integrity (6), Information (6), IEEE Access (5), Expert Systems with Applications (4), International Journal of Machine Learning and Cybernetics (4), and Mathematics, Procesamiento del Lenguaje Natural, Education and Information Technologies, Journal of Advances in Information Technology and CMC: Computers, Materials & Continua (3 each). By affiliation address, the leading countries are China (46), the USA (40), India (14), the UK (12), Saudi Arabia (9), Italy (9), Türkiye (8) and Canada (7); a record counts for every country in its addresses.
AI text detection in Türkiye
Eight included records have at least one Türkiye-affiliated author, but only one mentions Turkey, Türkiye or Turkish in its title, abstract or keywords, and it refers to the Turkish language, not a Turkish setting. It proposes a hybrid DistilBERT and graph convolutional network model, notes that detection is especially challenging in Turkish, Arabic and English because of their distinct structures, such as agglutinative morphology in Turkish, introduces new multilingual datasets and reports 98% accuracy for Turkish and Arabic and 99% for English (Bostancioglu et al., 2026).
The other seven do not name Türkiye in their abstracts: a systematic review on detectors in education (Caner, 2025); a study by an Italy-Türkiye author team on texts from four neurosurgery journals, in which detectors reached AUCs of 0.75 to 1.00 but none achieved 100% reliability (Erol et al., 2025); three detector development studies (Hazim & Ata, 2025, 2026; Utku, 2026); a study of AI-generated e-commerce reviews (Balci et al., 2026); and a policy analysis of the EU AI Act and a US Executive Order (Knott et al., 2024).
What is missing is substantial: no study set in a Turkish university or journal, no test of Turnitin, GPTZero or other commercial detectors on Turkish-language texts or on English texts by Turkish students or researchers, no evidence on false positives among Turkish writers, and no analysis of Turkish institutional or journal policies. Findings on L2 writers and other languages should be transferred with caution.
Debates and contradictions in the literature
Score as evidence or as a prompt for conversation. Some studies treat detectors as integrity tools: used together, they become a powerful tool for instructors when intuition gets it wrong (Hyatt et al., 2025), and selected tools may support safeguards because human judgement alone is insufficient (AL-Rawas et al., 2026). One position paper rejects evidential use, arguing that validating scores through linguistic markers, multiple tools or past work introduces confirmation bias rather than independent verification (Bassett et al., 2026). In between, flags open a dialogue (Giray, Sevnarayan, & Maphoto, 2026; Maphoto et al., 2026).
Vendor claims versus independent tests. Turnitin and similar detectors claim high accuracy, yet a critical article drawing on recent studies argues that they show systemic biases against multilingual writers (Maphoto et al., 2026). Independent results for Turnitin range from zero false positives (Gosling et al., 2024) to moderate discrimination (Villa et al., 2026) and as low as 45.7% in some obfuscation conditions (Makhmutova et al., 2026). High benchmark accuracy meets real-world complexity and counter-detection (X. Liu et al., 2025). Even rising detection rates are read as an early, modest effect of generative AI in six orthopaedic journals (Porto et al., 2025) but as likely software behaviour in a surgical journal (Bohler et al., 2026).
Banning versus integrating AI. Critics argue that detectors are ineffective, unreliable and harmful, with a chilling effect that paradoxically encourages AI use to avoid false accusations (Giray, Roe, & Diesta Espiritu, 2026), and that policing undermines learning (Ardana & Christiana, 2026).
Equity for non-native writers. Reviews and critiques report bias (Caner, 2025; Das Deep et al., 2025; Maphoto et al., 2026), whereas carefully built detectors on Graduate Record Examinations essays showed none (Jiang et al., 2024). One reading (the author’s) is that bias depends on tools and training data; a fairness-enhanced Indonesian model that narrowed the false positive gap fits it (Guspian et al., 2026). Teachers are not neutral either: linguistic deficit cues risk pathologising proficient learner writing (Barrot, 2026).
The detection arms race. An arms race between cheating and detection technology was anticipated (Gosling et al., 2024). One review still judges detection feasible if multiple approaches, domain customisation and ethics are combined (Kehkashan et al., 2025). The corpus contains both new attacks, such as fluent sentence insertions that evade detectors (Tu et al., 2026), and new defences reported to resist adversarial rewriting (Yin & Wang, 2026).
Research gaps and agenda
- Hybrid and edited texts. Earlier detectors mostly failed on human-machine collaboration such as polished texts (Yang et al., 2024); hybrid authorship remains neglected (Barrot & Aranda, 2025).
- Generalisation. Out-of-distribution data, attacks and weak evaluation frameworks are open problems (Wu et al., 2025); deployment needs cross-domain, mixed-content testing and calibration (Alluri et al., 2026).
- Languages beyond English. Cross-linguistic studies are needed (Sciandra et al., 2026); work on slightly polished text exists in English but not Arabic (Almohaimeed et al., 2026).
- Scientific style. Methods adapted to scientific writing are needed (Savelev et al., 2026), including span-level localisation and calibration across disciplines (Yin & Wang, 2026).
- Pedagogical effects. The impact of flawed detectors on learning is under-examined (Ardana & Christiana, 2026); longitudinal research on student perceptions is needed (Caner, 2025).
- Policy and training. Institutional policies, consistent enforcement and faculty training are lacking (Das Deep et al., 2025).
- Trustworthy detectors. Detection systems must themselves demonstrate fairness, robustness and accountability (Crothers et al., 2023).
Implications for practice: what to do with an “AI detected” result
The steps below translate the corpus recommendations into a procedure for academics, supervisors and editors; the sequence and the actions in the decision table are the author’s synthesis, and the last column names the evidence behind each step.
- Do not treat the score as a finding. It is a probabilistic estimate; one position paper argues that it cannot be verified against the true origin of the text (Bassett et al., 2026), and an evaluation study advises using it only to prompt inquiry (Hadra et al., 2026).
- Check error-prone conditions. Is the text short or technical (Sun et al., 2026), a literature review section (Dalalah & Dalalah, 2023), written by an L2 writer (Maphoto et al., 2026), or AI-edited for grammar or polish (Almohaimeed et al., 2026; Chemaya & Martin, 2024)? Low shares of AI words can produce inaccurately high Turnitin scores (Atamhenwan, 2026).
- Know the tool and threshold. Use a pre-specified threshold (Villa et al., 2026) and separate screening from diagnostic use (Rosenthal et al., 2026); tools give different scores for the same text (Savelev et al., 2026).
- Do not stack tools as “confirmation”. Agreement can lower false positives (Hyatt et al., 2025), but one position paper argues that it is confirmation bias rather than independent verification (Bassett et al., 2026).
- Read the work before the report. Warning numbers and red highlighting anchor judgement (Du et al., 2026).
- Collect process evidence. Drafts, reflective commentaries and an oral explanation (Maphoto et al., 2026); a short viva is one option (Hutson et al., 2026).
- Talk before you accuse. Use the flag to start a dialogue (Giray, Sevnarayan, & Maphoto, 2026) and record the evidence used.
- Apply a clear policy. Unclear rules leave grading perceived as inconsistent or unfair (Mirzayev et al., 2026).
Why base rates matter (author’s illustration with fictional numbers). Suppose a course receives 1,000 essays, 50 of them written with AI. A detector that catches 90% of AI essays and wrongly flags 2% of human essays produces 45 true and 19 false flags, so about three in ten flagged essays are human-written. If AI use were rarer, the share of false flags would rise. These numbers come from no specific tool; the interactive module lets you vary them.
| Situation after a flag | Suggested action (author’s synthesis) | Evidence behind the step |
|---|---|---|
| Score only; text short or technical | No action beyond your own reading | Systematic detector failures on short coursework and code Sun et al., 2026 |
| Score with plausible innocent explanation (L2 writer, grammar editing, permitted AI use) | Conversation; explain the tool’s limits | Chemaya & Martin, 2024; Maphoto et al., 2026 |
| Score plus concrete concerns (unrealistic data, content the writer cannot explain) | Request drafts, notes or a short oral explanation | Hutson et al., 2026; Maphoto et al., 2026; Stadler et al., 2025 |
| Score plus independent evidence of undeclared AI use against stated rules | Formal procedure based on the independent evidence; score not used as evidence | Balance of probabilities standard, which detection scores are argued not to meet Bassett et al., 2026 |
| Journal manuscript flagged at submission | Preliminary screening only; ask authors about AI use under the reporting policy | Pan & Florian-Rodriguez, 2024; Villa et al., 2026 |
Summary
- AI text detection classifies a text as human- or machine-written and reports a probability, unlike plagiarism detection, which matches verifiable sources.
- Classifiers, statistical signals, watermarks, stylometry and human judgement each fail under some conditions, especially domain shift and paraphrasing.
- Independent tests of Turnitin, GPTZero and other tools give high, low and mixed results that depend on text type, language, model version and editing.
- False positives are rare in some tests and frequent in others; edited, hybrid, short, technical and non-English texts and L2 writers face more risk.
- Human reviewers range from near chance to high accuracy across studies, and detection reports can anchor their judgement.
- The corpus advises against using a score alone for high-stakes decisions and favours dialogue, process evidence, clear policy and assessment redesign.
- Eight Türkiye-affiliated records exist; one mentions Turkish (as a language), and none studies a Turkish university, journal or writer group.
Discussion questions
- Why is an AI score different in kind from a similarity score?
- Turnitin produced zero false positives in one study and detected only 30% of rephrased AI text in another. How can both be true?
- Is aggregating detectors a safeguard against false positives or a form of confirmation bias? Use two sources.
- How can a policy protect L2 writers without ignoring undeclared AI use?
- Applied: draft a one-page departmental procedure for an “AI detected” result.
- Applied: design a study testing a commercial detector on Turkish-language and EFL texts by Turkish students, naming the error measures you would report.
Further reading
- Weber-Wulff et al. (2023): a broad early test of 14 tools, including Turnitin, with translation and obfuscation.
- Caner (2025): a PRISMA 2020 systematic review of 25 empirical studies on detection tools in education.
- Wu et al. (2025): a survey of watermarking, statistical, neural and human-assisted detection.
- Rosenthal et al. (2026): five tools on human, AI, humanised and hybrid texts, with a threshold-based framework.
- Bassett et al. (2026): a position paper arguing against detection scores in integrity investigations.
- Maphoto et al. (2026): a critique of Turnitin’s AI detector in multilingual writing with a dialogic alternative.
Glossary
| Term | Turkish | Short definition |
|---|---|---|
| AI-generated text | Yapay zekâ ile üretilmiş metin | Text produced by a large language model |
| AI detection | Yapay zekâ tespiti | Classifying a text as human- or machine-written, usually with a score |
| False positive | Yanlış pozitif | A human-written text classified as AI-generated |
| Academic integrity | Akademik dürüstlük | Norms and procedures that protect honest scholarship |
| Plagiarism | İntihal | Presenting others’ words or ideas as one’s own |
| Watermark | Filigran | A signal embedded in generated text so that it can be identified |
| Authorship | Yazarlık | Who produced a text and is responsible for it |
| Large language model | Büyük dil modeli | A model trained on large text collections that generates fluent text |
| Humanizer | İnsanlaştırıcı araç | A rewriting tool used to make AI text evade detectors |
| Hybrid text | Karma (hibrit) metin | A text combining human-written and AI-generated content |
Frequently asked questions
Can Turnitin or GPTZero prove that a text was written by AI? No. One position paper argues that their probabilistic estimates cannot be verified against the true origin of a text and do not meet the balance of probabilities standard (Bassett et al., 2026); an evaluation study advises using them only as a prompt for further inquiry (Hadra et al., 2026).
How accurate are AI detectors? It depends on the tool, text and model. Some tests report near-perfect separation of fully human and fully AI text (Rosenthal et al., 2026), others accuracies from 14.3% to 71.4% (Wakjira et al., 2025), and accuracy falls for hybrid and humanised texts (Rosenthal et al., 2026).
Are non-native English writers more likely to be flagged? Several reviews and critiques say so (Caner, 2025; Maphoto et al., 2026), but carefully built detectors on Graduate Record Examinations essays showed no evidence of such bias (Jiang et al., 2024); the risk depends on the tool.
Can people spot AI writing better than software? Usually not reliably: in one study 63 lecturers and AI detectors were both only slightly better than chance, with no significant difference between them (Fiedler & Döpke, 2025).
What should I do if a student’s essay is flagged? Treat the flag as a cue, not a verdict: triangulate it with drafts, reflective commentary and an oral explanation (Maphoto et al., 2026), and follow your institution’s policy.
Related on this site: When “AI-detected” does not mean “AI-written”: a reading of a new Turnitin study
References
- Ahmad, Z., Torres-Ruiz, M., Mahmood, A., Quintero, R., Ameer, I., & Bölücü, N. (2026). Human or machine? A survey on machine-generated text detection. IEEE Access, 14, 34113-34136. https://doi.org/10.1109/ACCESS.2026.3666781
- Alluri, N., Gottumukkala, P. S. V., & Indukuri, H. (2026). Detecting AI-generated text and code: An empirical study of cross-generator and cross-domain generalization. AI, 7(8), Article 319. https://doi.org/10.3390/ai7080319
- Almohaimeed, S., Almohaimeed, S., Jari, M., Alobaid, K. A., & Alotaibi, F. (2026). AI text detectors and the misclassification of slightly polished Arabic text. Journal of Big Data, 13(1), Article 146. https://doi.org/10.1186/s40537-026-01492-8
- AL-Rawas, M., Abdul Qader, O. A. J., Lin, G. S. S., Beh, Y. H., Sabarudin, M. A., Ang, Y., Low, J. F., Abdullah, J. Y., & Noorani, T. Y. (2026). Human versus artificial intelligence: Investigating ability of young academics from research and non-research institutions to identify ChatGPT-generated dental research abstracts. Scientific Reports, 16(1), Article 12287. https://doi.org/10.1038/s41598-026-42555-3
- Alshammari, H., El-Sayed, A., & Elleithy, K. (2024). AI-generated text detector for Arabic language using encoder-based transformer architecture. Big Data and Cognitive Computing, 8(3), Article 32. https://doi.org/10.3390/bdcc8030032
- Ardana, S. G., & Christiana, M. (2026). AI detectors, student anxiety, and authorial alienation: A qualitative study of affective control. Teaching in Higher Education, 31(6), 1081-1100. https://doi.org/10.1080/13562517.2026.2643825
- Atamhenwan, L. E. (2026). How are combinations of human-written words and LLM-generated words by ChatGPT, Copilot, Gemini and Grammarly detected by Turnitin? Education and Information Technologies, 31(14), 6483-6517. https://doi.org/10.1007/s10639-026-14049-2
- Balci, M. A., Erdogan, D. R., Akgüller, Ö., & Gaban, L. (2026). Detecting AI-generated e-commerce reviews with sheaf-hypergraph consensus residuals: Interpretable signals and cross-domain asymmetry. Applied Sciences, 16(14), Article 7071. https://doi.org/10.3390/app16147071
- Barrot, J. (2026). Reconsidering AI-detection in EFL writing: A response to De Wilde. ELT Journal, 80(1), 160-163. https://doi.org/10.1093/elt/ccaf071
- Barrot, J. S., & Aranda, M. R. R. (2025). Efficacy of AI-text detection tools in distinguishing student-produced, AI-edited, and AI-generated essays. Technology, Knowledge and Learning. Advance online publication. https://doi.org/10.1007/s10758-025-09884-0
- Bashir, S. (2025). Using pseudo-AI submissions for detecting AI-generated code. Frontiers in Computer Science, 7, Article 1549761. https://doi.org/10.3389/fcomp.2025.1549761
- Bassett, M. A., Bradshaw, W., Bornsztejn, H., Hogg, A., Murdoch, K., Pearce, B., & Webber, C. (2026). Heads we win, tails you lose: AI detectors in education. Journal of Higher Education Policy and Management. Advance online publication. https://doi.org/10.1080/1360080X.2026.2622146
- Bohler, F., Tran, T., Burmeister, J. R., Hadid, K., Joshi, S., Gao, S., & Chaiyasate, K. (2026). Detectable AI-generated text in the Journal of Craniofacial Surgery: A comparison of 2014 and 2024 publications. Journal of Craniofacial Surgery, 37(3/4), 841-844. https://doi.org/10.1097/SCS.0000000000012366
- Bostancioglu, A., Das, B., & Bulut Ozek, M. (2026). Multilingual AI-generated text detection in Arabic, English, and Turkish using a hybrid transformer-graph convolutional network. Applied Sciences, 16(14), Article 7249. https://doi.org/10.3390/app16147249
- Burden, M. (2026). Utilizing and detecting AI in higher education: A reflection from an online, asynchronous, criminal justice course. Journal of Criminal Justice Education, 37(2), 313-323. https://doi.org/10.1080/10511253.2026.2646601
- Caner, H. N. (2025). AI detection tools: A systematic review of empirical evidence and their implications for education. Khazar Journal of Humanities and Social Sciences, 28(4), 234-246. https://doi.org/10.5782/2223-2621.1580
- Chemaya, N., & Martin, D. (2024). Perceptions and detection of AI use in manuscript preparation for academic journals. PLOS ONE, 19(7), Article e0304807. https://doi.org/10.1371/journal.pone.0304807
- Cheng, A., Lin, Y., Reedy, G., Joseph, C., Wirkowski, S., Mallette, V., Nagesh, V., Krieser, D., & Calhoun, A. (2025). Ability of AI detection tools and humans to accurately identify different forms of AI-generated written content. Advances in Simulation, 10(1), Article 66. https://doi.org/10.1186/s41077-025-00396-6
- Crothers, E. N., Japkowicz, N., & Viktor, H. L. (2023). Machine-generated text: A comprehensive survey of threat models and detection methods. IEEE Access, 11, 70977-71002. https://doi.org/10.1109/ACCESS.2023.3294090
- Dalalah, D., & Dalalah, O. M. A. (2023). The false positives and false negatives of generative AI detection tools in education and academic research: The case of ChatGPT. International Journal of Management Education, 21(2), Article 100822. https://doi.org/10.1016/j.ijme.2023.100822
- Das Deep, P., Edgington, W. D., Ghosh, N., & Rahaman, M. S. (2025). Evaluating the effectiveness and ethical implications of AI detection tools in higher education. Information, 16(10), Article 905. https://doi.org/10.3390/info16100905
- Desaire, H., Chua, A. E., Kim, M.-G., & Hua, D. (2023). Accurately detecting AI text when ChatGPT is told to write like a chemist. Cell Reports Physical Science, 4(11), Article 101672. https://doi.org/10.1016/j.xcrp.2023.101672
- De Wilde, V. (2024). Can novice teachers detect AI-generated texts in EFL writing? ELT Journal, 78(4), 414-422. https://doi.org/10.1093/elt/ccae031
- Du, P., Liu, T., & Xian, X. (2026). Automation bias in teachers’ evaluation of student writing: Effects of algorithmic warnings and visual risk cues in AI detection reports. Frontiers in Psychology, 17, Article 1889402. https://doi.org/10.3389/fpsyg.2026.1889402
- Elkhatat, A. M., Elsaid, K., & Almeer, S. (2023). Evaluating the efficacy of AI content detection tools in differentiating between human and AI-generated text. International Journal for Educational Integrity, 19(1), Article 17. https://doi.org/10.1007/s40979-023-00140-5
- Erol, G., Ergen, A., Erol, B. G., Kaya Ergen, S., Bora, T. S., Çölgeçen, A. D., Araz, B., Sahin, C., Bostanci, G., Kiliç, I., Macit, Z. B., Sevgi, U. T., & Güngör, A. (2025). Can we trust academic AI detective? Accuracy and limitations of AI-output detectors. Acta Neurochirurgica, 167(1), Article 214. https://doi.org/10.1007/s00701-025-06622-4
- Fiedler, A., & Döpke, J. (2025). Do humans identify AI-generated text better than machines? Evidence based on excerpts from German theses. International Review of Economics Education, 49, Article 100321. https://doi.org/10.1016/j.iree.2025.100321
- Fine, K. S., Zona, E. E., O’Shea, A. W., Via, E. C., Attaluri, P. K., Wirth, P. J., Dingle, A. M., & Poore, S. O. (2026). Evaluating the detection accuracy of AI-generated content in plastic surgery: A comparative study of medical professionals and AI tools. Plastic and Reconstructive Surgery, 157(1), 129e-136e. https://doi.org/10.1097/PRS.0000000000012242
- Foltynek, T., & Newton, P. M. (2025). What does YouTube advise students about bypassing AI-text detection tools? A pragmatic analysis. Journal of Academic Ethics, 24(1), Article 8. https://doi.org/10.1007/s10805-025-09675-3
- Gao, C. A., Howard, F. M., Markov, N. S., Dyer, E. C., Ramesh, S., Luo, Y., & Pearson, A. T. (2023). Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors and blinded human reviewers. npj Digital Medicine, 6(1), Article 75. https://doi.org/10.1038/s41746-023-00819-6
- Giray, L., Roe, J., & Diesta Espiritu, J. (2026). AI writing detectors are ineffective, unreliable and harmful. English Teaching: Practice and Critique. Advance online publication. https://doi.org/10.1108/ETPC-07-2025-0155
- Giray, L., Sevnarayan, K., & Maphoto, K. B. (2026). GPTZero and the challenges of AI detection in assessing writing. Assessing Writing, 69, Article 101078. https://doi.org/10.1016/j.asw.2026.101078
- Gosling, S. D., Ybarra, K., & Angulo, S. K. (2024). A widely used generative-AI detector yields zero false positives. Aloma: Revista de Psicologia, Ciències de l’Educació i de l’Esport, 42(2), 31-43. https://doi.org/10.51698/aloma.2024.42.2.31-43
- Guspian, I., Lin, C.-C., & Yang, S. J. H. (2026). A fairness-enhanced hybrid model for Indonesian AI text detection with component-based explainability. KSII Transactions on Internet and Information Systems, 20(9), 4541-4563. https://doi.org/10.3837/tiis.2026.09.008
- Habibzadeh, F. (2023). GPTZero performance in identifying artificial intelligence-generated medical texts: A preliminary study. Journal of Korean Medical Science, 38(38), Article e319. https://doi.org/10.3346/jkms.2023.38.e319
- Hadra, M., Cambridge, K., & Mesbah, M. (2026). Evaluating the accuracy and reliability of AI content detectors in academic contexts. International Journal for Educational Integrity, 22(1), Article 4. https://doi.org/10.1007/s40979-026-00213-1
- Hazim, L., & Ata, O. (2025). Bridging the gap: Ensemble learning-based NLP framework for AI-generated text identification in academia. International Arab Journal of Information Technology, 22(6), 1054-1068. https://doi.org/10.34028/iajit/22/6/2
- Hazim, L. R., & Ata, O. (2026). HQML-NLP: A hybrid quantum machine learning framework for scholarly AI-text detection. Applied Soft Computing, 191, Article 114634. https://doi.org/10.1016/j.asoc.2026.114634
- Hua, H., & Yao, C.-J. (2024). Investigating generative AI models and detection techniques: Impacts of tokenization and dataset size on identification of AI-generated text. Frontiers in Artificial Intelligence, 7, Article 1469197. https://doi.org/10.3389/frai.2024.1469197
- Hutson, J., Poyer, K., Ogoe, E., & Atologun, K. A. (2026). Beyond AI detection: A pilot study of Integrevise™ and viva-based verification of student understanding in AI-mediated assessment. Trends in Higher Education, 5(3), Article 59. https://doi.org/10.3390/higheredu5030059
- Hyatt, J.-P. K., Bienenstock, E. J., Firetto, C. M., Woods, E. R., & Comus, R. C. (2025). Using aggregated AI detector outcomes to eliminate false positives in STEM-student writing. Advances in Physiology Education, 49(2), 486-495. https://doi.org/10.1152/advan.00235.2024
- Ibrahim, K. H. S., Al Otaibi, D., & Sibai, F. N. (2025). The robustness of AI-classifiers in the face of AI-assisted plagiarism: The case of Turnitin AI content detector. International Journal of Computer-Assisted Language Learning and Teaching, 15(1), Article 375428. https://doi.org/10.4018/IJCALLT.375428
- Jiang, Y., Hao, J., Fauss, M., & Li, C. (2024). Detecting ChatGPT-generated essays in a large-scale writing assessment: Is there a bias against non-native English speakers? Computers & Education, 217, Article 105070. https://doi.org/10.1016/j.compedu.2024.105070
- 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, Article 100793. https://doi.org/10.1016/j.cosrev.2025.100793
- Knott, A., Pedreschi, D., Jitsuzumi, T., Leavy, S., Eyers, D., Chakraborti, T., Trotman, A., Sundareswaran, S., Baeza-Yates, R., Biecek, P., Weller, A., Teal, P. D., Basu, S., Haklidir, M., Morini, V., Russell, S., & Bengio, Y. (2024). AI content detection in the emerging information ecosystem: New obligations for media and tech companies. Ethics and Information Technology, 26(4), Article 63. https://doi.org/10.1007/s10676-024-09795-1
- Liu, J. Q. J., Hui, K. T. K., Al Zoubi, F., Zhou, Z. Z. X., Samartzis, D., Yu, C. C. H., Chang, J. R., & Wong, A. Y. L. (2024). The great detectives: Humans versus AI detectors in catching large language model-generated medical writing. International Journal for Educational Integrity, 20(1), Article 8. https://doi.org/10.1007/s40979-024-00155-6
- Liu, X., Li, Y., & Li, K. (2025). Enhancing the robustness of AI-generated text detectors: A survey. Mathematics, 13(13), Article 2145. https://doi.org/10.3390/math13132145
- Lv, Z., & Sun, J. (2026). Zero-shot detection of LLM-generated text via multilevel information compression and multidimensional feature fusion. Neurocomputing, 671, Article 132731. https://doi.org/10.1016/j.neucom.2026.132731
- Makhmutova, A., Sharimbayev, B., Amirzhanov, A., & Shalkarbay-uly, A. (2026). Testing the limits: Evaluating AI detectors’ accuracy and the impact of obfuscation techniques on AI-generated text. Journal of Advances in Information Technology, 17(3), 438-449. https://doi.org/10.12720/jait.17.3.438-449
- Makiev, K. G., Asimakidou, M., Vasios, I. S., Keskinis, A., Petkidis, G., Tilkeridis, K., Ververidis, A., & Iliopoulos, E. (2023). A study on distinguishing ChatGPT-generated and human-written orthopaedic abstracts by reviewers: Decoding the discrepancies. Cureus, 15(11), Article e49166. https://doi.org/10.7759/cureus.49166
- Maphoto, K. B., Giray, L., & Sevnarayan, K. (2026). Automated integrity or automated injustice? Turnitin’s AI detection in multilingual writing contexts. Assessing Writing, 69, Article 101086. https://doi.org/10.1016/j.asw.2026.101086
- Milicka, J., Marklová, A., Drobil, O., & Pospísilová, E. (2025). Learning to detect AI texts and learning the limits. PLOS ONE, 20(10), Article e0333007. https://doi.org/10.1371/journal.pone.0333007
- Miralles-González, P., Huertas-Tato, J., Martín, A., & Camacho, D. (2026). Not all tokens are created equal: Perplexity Attention Weighted Networks for AI-generated text detection. Information Fusion, 125, Article 103465. https://doi.org/10.1016/j.inffus.2025.103465
- Mirzayev, A., Kamal, S. S. L. A., Turdiyeva, G., Tojiboyeva, M., & Irgasheva, N. (2026). AI detection and grading fairness in EFL assessment: A qualitative study of student and teacher perceptions at a university in Uzbekistan. Cogent Education, 13(1), Article 2742806. https://doi.org/10.1080/2331186X.2026.2742806
- Murch, H., Worley, M., & Volk, F. (2025). A comparative analysis of the prevalence of artificial intelligence and plagiarism in doctoral dissertations using Turnitin. Journal of Academic Ethics, 23(4), 1915-1933. https://doi.org/10.1007/s10805-025-09634-y
- Murgia, M., Reforgiato Recupero, D., & Spathoulas, G. (2026). A comprehensive evaluation of fine-tuned LLMs as AI-generated text detectors under adversarial and multi-domain conditions. IEEE Access, 14, 66819-66837. https://doi.org/10.1109/ACCESS.2026.3689209
- Nitu, M., & Dascalu, M. (2024). Beyond lexical boundaries: LLM-generated text detection for Romanian digital libraries. Future Internet, 16(2), Article 41. https://doi.org/10.3390/fi16020041
- Oghaz, M. M., Saheer, L. B., Dhame, K., & Singaram, G. (2025). Detection and classification of ChatGPT-generated content using deep transformer models. Frontiers in Artificial Intelligence, 8, Article 1458707. https://doi.org/10.3389/frai.2025.1458707
- Pan, E. T., & Florian-Rodriguez, M. (2024). Human vs machine: Identifying ChatGPT-generated abstracts in gynecology and urogynecology. American Journal of Obstetrics and Gynecology, 231(2), 276e1-276e10. https://doi.org/10.1016/j.ajog.2024.04.045
- Pantic, I. V., & Mugosa, S. (2025). Artificial intelligence strategies based on random forests for detection of AI-generated content in public health. Public Health, 242, 382-387. https://doi.org/10.1016/j.puhe.2025.03.029
- Paustian, T., & Slinger, B. (2024). Students are using large language models and AI detectors can often detect their use. Frontiers in Education, 9, Article 1374889. https://doi.org/10.3389/feduc.2024.1374889
- Popescu-Apreutesei, L.-E., Iosupescu, M.-S., Necula, S. C., & Pavaloaia, V.-D. (2025). Upholding academic integrity amidst advanced language models: Evaluating BiLSTM networks with GloVe embeddings for detecting AI-generated scientific abstracts. CMC: Computers, Materials & Continua, 84(2), 2605-2644. https://doi.org/10.32604/cmc.2025.064747
- Popkov, A. A., & Barrett, T. S. (2025). AI vs academia: Experimental study on AI text detectors’ accuracy in behavioral health academic writing. Accountability in Research: Ethics, Integrity and Policy, 32(7), 1072-1088. https://doi.org/10.1080/08989621.2024.2331757
- Porto, J. R., Morgan, K. A., Hecht, C. J., II, Burkhart, R. J., & Liu, R. W. (2025). Quantifying the scope of artificial intelligence-assisted writing in orthopaedic medical literature: An analysis of prevalence and validation of AI-detection software. Journal of the American Academy of Orthopaedic Surgeons, 33(1), 42-50. https://doi.org/10.5435/JAAOS-D-24-00084
- Pratama, A. R. (2025). The accuracy-bias trade-offs in AI text detection tools and their impact on fairness in scholarly publication. PeerJ Computer Science, 11, Article e2953. https://doi.org/10.7717/peerj-cs.2953
- Rosenthal, S. R., Jacob, S., Dalling, S. C., Villa, M. E. L., Kasprzak, M., Bruno, S. R., & Kulp, A. E. (2026). Interpreting AI text detection outputs: Empirical evaluation and a threshold-based framework. Studies in Educational Evaluation, 90, Article 101665. https://doi.org/10.1016/j.stueduc.2026.101665
- Rydell, F., Hansen, P., Kanestad, L., Ljungar, A., & Öberg, S. (2026). Low-resource single-pass paraphrasing attacks on AI-generated text detection. Intelligent Systems with Applications, 31, Article 200710. https://doi.org/10.1016/j.iswa.2026.200710
- Salnikov, E., & Bonch-Osmolovskaya, A. (2025). Detecting LLM-generated text with trigram-cosine stylometric delta: An unsupervised and interpretable approach. Journal of Language and Education, 11(3), 138-151. https://doi.org/10.17323/jle.2025.22211
- Sarvazyan, A. M., González, J. A., Franco-Salvador, M., Rangel, F., Chulvi, B., & Rosso, P. (2023). Overview of AuTexTification at IberLEF 2023: Detection and attribution of machine-generated text in multiple domains. Procesamiento del Lenguaje Natural, (71), 275-288. https://doi.org/10.26342/2023-71-21
- Sarvazyan, A. M., González, J. A., Rangel, F., Rosso, P., & Franco-Salvador, M. (2024). Overview of IberAuTexTification at IberLEF 2024: Detection and attribution of machine-generated text on languages of the Iberian Peninsula. Procesamiento del Lenguaje Natural, (73), 421-434. https://doi.org/10.26342/2024-73-32
- Savelev, I. L., Ivannikov, A. L., Solodov, S. V., Maryushin, L. A., Valova, A. A., & Korenkova, T. V. (2026). Experimental evaluation of the reliability of AI-generated text identification in scientific literature. Physics of Particles and Nuclei, 57(4), 549-552. https://doi.org/10.1134/S1063779626700085
- Sciandra, A., Dal Cero, F., Cortelazzo, M. A., & Tuzzi, A. (2026). AI detection in Italian essays through different text representations and adversarial robustness evaluation. Scientific Reports, 16(1), Article 21499. https://doi.org/10.1038/s41598-026-51613-9
- Shamsi, A., Wang, T., Amraei, M., & Raju, N. V. (2026). Evaluating AI text detection tools for distinguishing human-written from AI-generated abstracts in Persian-language journals of library and information science. Acta Informatica Pragensia, 15(1), 126-134. https://doi.org/10.18267/j.aip.293
- Sharimbayev, B., & Kadyrov, S. (2026). Advancing machine-generated text detection: A comprehensive evaluation of transformer-based models. Journal of Advances in Information Technology, 17(7), 1321-1330. https://doi.org/10.12720/jait.17.7.1321-1330
- Stadler, R. D., Sudah, S. Y., Moverman, M. A., Denard, P. J., Duralde, X. A., Garrigues, G. E., Klifto, C. S., Levy, J. C., Namdari, S., Sanchez-Sotelo, J., & Menendez, M. E. (2025). Identification of ChatGPT-generated abstracts within shoulder and elbow surgery poses a challenge for reviewers. Arthroscopy: The Journal of Arthroscopic and Related Surgery, 41(4), 916-924. https://doi.org/10.1016/j.arthro.2024.06.045
- Sun, Y., Liao, Y., & Ma, X. (2026). Trusting AI to detect AI? A systematic evaluation of the reliability and robustness of current AIGC detection tools for student academic work. Computers & Education, 249, Article 105616. https://doi.org/10.1016/j.compedu.2026.105616
- Tamuka, N. (2026). Performance evaluation and interpretability of a Qwen LangGraph agent for AI written and AI rewritten text detection. Discover Computing, 29(1), Article 745. https://doi.org/10.1007/s10791-026-10710-z
- Tayebipour, F., Siddique, M., & Hamid, A. (2026). Omani college students’ and lecturers’ perceptions of Turnitin and artificial intelligence in relation to academic integrity. Frontiers in Education, 11, Article 1864395. https://doi.org/10.3389/feduc.2026.1864395
- Teraif, Z. (2026). Who wrote this? An EAP think-aloud study on AI detection. Language Learning & Technology, 30(1), 1-22. https://doi.org/10.64152/10125/73709
- Tu, R., Kang, X., Tan, C. W., Chi, C.-H., & Lam, K.-Y. (2026). Sentences based adversarial attack on AI-generated text detectors. IEEE Transactions on Big Data, 12(1), 80-91. https://doi.org/10.1109/TBDATA.2025.3600034
- Utku, A. (2026). MILA: An innovative approach to identifying AI-generated content using BERT, CNN, and BiLSTM. Arabian Journal for Science and Engineering, 51(5), 5653-5669. https://doi.org/10.1007/s13369-025-10466-6
- Van Vlasselaer, M., Van Droogenbroeck, F., & Spruyt, B. (2026). Who wrote this? Evaluating the reliability of AI detection tools in higher education. International Journal for Educational Integrity, 22(1), Article 16. https://doi.org/10.1007/s40979-026-00226-w
- Villa, J., Garcovich, D., Lombardo, L., Siciliani, G., & Adobes Martin, M. (2026). AI text detection in dentistry: A comparative analysis across generative models. Research Integrity and Peer Review, 11(1), Article 46. https://doi.org/10.1186/s41073-026-00228-9
- Wakjira, T. G., Tijani, I. A., Alam, M. S., Mashal, M., & Hasan, M. K. (2025). Can we trust AI content detection tools for critical decision-making? Information, 16(10), Article 904. https://doi.org/10.3390/info16100904
- Waltzer, T., Pilegard, C., & Heyman, G. D. (2024). Can you spot the bot? Identifying AI-generated writing in college essays. International Journal for Educational Integrity, 20(1), Article 11. https://doi.org/10.1007/s40979-024-00158-3
- Wang, W., & Feng, A. (2021). Self-information loss compensation learning for machine-generated text detection. Mathematical Problems in Engineering, 2021, Article 6669468. https://doi.org/10.1155/2021/6669468
- Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltynek, T., Guerrero-Dib, J., Popoola, O., Sigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19(1), Article 26. https://doi.org/10.1007/s40979-023-00146-z
- Wu, J., Yang, S., Zhan, R., Yuan, Y., Chao, L. S., & Wong, D. F. (2025). A survey on LLM-generated text detection: Necessity, methods, and future directions. Computational Linguistics, 51(1), 275-338. https://doi.org/10.1162/coli_a_00549
- Xiang, L., Li, N., Liu, Y., & Hu, J. (2026). AI-generated text detection: A comprehensive review of active and passive approaches. CMC: Computers, Materials & Continua, 86(3), Article 073347. https://doi.org/10.32604/cmc.2025.073347
- Yang, L., Jiang, F., & Li, H. (2024). Is ChatGPT involved in texts? Measure the Polish Ratio to detect ChatGPT-generated text. APSIPA Transactions on Signal and Information Processing, 13(2). https://doi.org/10.1561/116.00000250
- Yin, Z., & Wang, S. (2026). Span-level detection of AI-generated scientific text via contrastive learning and structural calibration. Knowledge-Based Systems, 334, Article 115123. https://doi.org/10.1016/j.knosys.2025.115123
- Zhang, Y., Jiang, X., Sun, H., Zhang, Y., & Tong, D. (2025). CurveMark: Detecting AI-generated text via probabilistic curvature and dynamic semantic watermarking. Entropy, 27(8), Article 784. https://doi.org/10.3390/e27080784
- Zhao, Y., Borelli, A., Martinez, F., Xue, H., & Weiss, G. M. (2024). Admissions in the age of AI: Detecting AI-generated application materials in higher education. Scientific Reports, 14(1), Article 26411. https://doi.org/10.1038/s41598-024-77847-z
Source verification and use of AI. Author names, publication years, titles, journals, volumes, issues, pages or article numbers and DOIs in the reference list were taken from the Web of Science records of the included studies, with author names taken from the full-name (AF) field. Initials were derived from the full given names in AF, keeping hyphens; names were corrected only for obvious export errors and diacritics, and two title export errors were fixed (“AIto”; “these” for “theses”, checked against the DOI’s Crossref record); journal names were normalised from the capitalised Web of Science form, and titles were set in sentence case. Claims attributed to sources were checked against the abstracts in an independent verification at abstract level; full texts were not reviewed, so numbers are reported as the abstracts state them. AI tools supported screening, evidence extraction from abstracts, drafting and figure generation. Responsibility for the content rests with the author, Prof. Dr. Mehmet Nurullah Kurutkan.
Data and supplementary material. The full list of the 185 included records, the screening log and the verification report are provided in the supplementary file “TU_101_supplementary.xlsx”, available from the author on request.
Cite this page (APA 7)
Kurutkan, M. N. (2026, November 18). AI-Generated Text Detection and Turnitin: An Introduction (101). The Health Topics. https://healthtopic.org/topics/ai-generated-text-detection-and-turnitin-an-introduction-101/
