Research gapAll disciplines

AI-text detectors in integrity cases: how reliable and how fair are they for real student populations?

Detectors that flag "AI-written" work are now part of many misconduct processes, yet benchmark studies disagree on whether they misjudge non-native English writers (Liang et al., 2023; Jiang et al., 2024). The evidence that would justify using a score in a decision about a student is still missing.

PhD Research Lab InsightsEditorial team, reviewed by a subject specialist
Posted
3 October 2026
Last checked
3 October 2026
Reading time
5 min
Sources
10
Research proposal support
The gap

We do not yet know whether commercial AI-text detectors, used at the thresholds institutions actually apply to authentic coursework, produce higher false-positive rates for writers of English as an additional language than for first-language writers, nor how integrity officers weigh those scores when deciding a case. Published estimates conflict (Liang et al., 2023; Jiang et al., 2024; Pratama, 2025), and almost none come from the integrity process itself.

Evidence gap (contradictory findings)Methodological gapPopulation gapPractical-knowledge gap

What we know

  • Seven public detectors misclassified an average of 61.3% of TOEFL essays by non-native writers as AI-generated, while classifying essays by US eighth-graders almost perfectly; low text perplexity was the likely cause (Liang et al., 2023).
  • A test of 14 tools, including Turnitin and PlagiarismCheck, found none fully accurate or reliable; accuracy fell sharply on manually edited (42%) and machine-paraphrased (26%) AI text (Weber-Wulff et al., 2023).
  • Simple evasion techniques cut the accuracy of six major detectors to 17.4% across 805 samples (Perkins et al., 2024).
  • On a benchmark of over six million generations, 12 detectors were easily fooled by adversarial attacks, decoding changes and unseen models (Dugan et al., 2024).
  • Detectors built on carefully sampled GRE essays reached near-perfect accuracy with no evidence of bias against non-native speakers, which contradicts earlier findings (Jiang et al., 2024).
  • With AI-assisted scholarly abstracts, GPTZero over-flagged non-native authors more often than native authors (Pratama, 2025), and human teachers performed no better, while being overconfident (Fleckenstein et al., 2024).

What we do not know yet

  • False-positive rates on authentic, institution-level coursework, stratified by language background, discipline and level of study (including doctoral writing).
  • Whether the contradictory bias findings reflect the detector, the sampling of essays, or the threshold chosen.
  • How legitimate assistive uses (grammar checkers, translation, editing support) shift scores for different groups.
  • How integrity officers and panels interpret a detector score alongside other evidence, and with what outcomes.
  • Whether institutional policies that restrict or disable detectors change case numbers or fairness of outcomes.

Research questions a PhD could ask

  1. At the thresholds used by a given institution, do commercial detectors yield different false-positive rates for first-language and additional-language English writers on authentic, verified human coursework?
  2. To what extent do sampling choices (test essays versus coursework, prompt type, text length) explain the conflicting bias estimates reported in benchmark studies?
  3. How do permitted assistive tools such as grammar checkers and machine translation alter detector scores, and does the effect differ by writer group?
  4. How do academic integrity officers interpret detector reports, and what weight do scores carry in case outcomes?
  5. What procedural safeguards (score disclosure, corroborating evidence, oral verification) reduce the risk of wrongful findings without removing scrutiny of genuine misconduct?
Methods that could close the gap
ApproachWhy it fits
Stratified benchmark studyTests detectors on verified human writing from diverse writer populations at real decision thresholds, extending designs used by Liang et al. (2023) and Jiang et al. (2024).
Controlled assistive-use experimentVaries grammar checking, translation and paraphrasing on the same texts to isolate their effect on scores, building on obfuscation tests (Weber-Wulff et al., 2023; Perkins et al., 2024).
Robustness evaluation on shared benchmarksUses open benchmarks such as RAID to compare detectors under attacks and unseen models before any field use (Dugan et al., 2024).
Field study of integrity casesAnalyses anonymised case records with ethics approval to see how scores enter decisions and outcomes across student groups.
Mixed methods with integrity officersCombines vignette experiments with interviews to show how officers weigh scores against other evidence, extending evidence on human judgement (Fleckenstein et al., 2024).

Why this gap matters now

An academic misconduct finding can delay a degree, end a scholarship or damage a career. When a detector score is one of the pieces of evidence behind that finding, its error rate stops being a technical detail and becomes a question of procedural fairness. The early evidence was stark: in one widely cited study, most essays written by non-native English writers were labelled as machine-generated (Liang et al., 2023). Some institutions responded by switching detectors off. Vanderbilt University, for example, disabled Turnitin's AI detector, noting that even a 1% false-positive rate could affect around 750 of the 75,000 papers it had submitted in 2022 (Vanderbilt University, 2023).

Yet the literature has not settled. Later work using large, carefully sampled assessment data found no bias against non-native speakers (Jiang et al., 2024), while a study of scholarly abstracts found that bias appeared mainly when humans used AI to polish their own writing (Pratama, 2025). These results are not necessarily in conflict; they test different tools, texts and thresholds. That is exactly what makes this an evidence gap in the sense described by Müller-Bloch and Kranz (2015) and Miles (2017).

Why the gap persists

Three features of the field keep the question open. First, most studies use convenience corpora: public essay sets, test responses or published abstracts. These are easier to obtain than coursework, but they differ in length, genre and the writing support students actually use. Second, detectors change quickly, and so do the models they try to catch, so findings date fast (Dugan et al., 2024). Third, the decisions that matter happen inside confidential integrity processes, which researchers rarely see.

  • Commercial vendors seldom publish method details or error rates by writer group.
  • Ground truth is hard to establish: a researcher must know, not guess, that a text is human-written.
  • Access to case records requires institutional trust, data-protection safeguards and ethics approval.

The result is a body of evidence strong on what detectors do in the laboratory and thin on what they do to students in practice (Weber-Wulff et al., 2023; Perkins et al., 2024).

Is this feasible for a doctorate?

Yes, if the scope is drawn tightly. A realistic thesis might combine one benchmark study on verified human coursework from two or three writer groups, one controlled experiment on assistive tools, and one qualitative or vignette study with integrity staff. Each strand can produce a publishable paper, and together they answer a single question about fairness in decision-making.

Practical planning for a doctoral project on detector fairness
ConsiderationWhat to plan for
Data accessConsent-based collection of drafts and final texts, or a partnership with an integrity office.
EthicsApproval covering student data, anonymisation and the risk of re-identification in case records.
Tool versionsRecord detector versions and dates; re-test before submission because tools change.
SkillsClassification metrics, fairness measures, and qualitative or mixed-methods analysis.
Disciplinary homeEducation, applied linguistics, information science or computer science, with co-supervision across them.

A focused study of this kind would be valuable whatever it finds. Showing that bias is small under realistic conditions matters as much as showing that it is large, because either result tells institutions how much weight a score can bear.

What to do now

  • Run a scoping search from 2023 onwards and map each study by tool, text type, writer group and threshold.
  • Decide which strand (benchmark, experiment or field study) is the core of your thesis and which are supporting.
  • Approach an integrity office or writing centre early about data access and ethics.
  • Type your gap explicitly, using an established typology (Müller-Bloch & Kranz, 2015), before drafting research questions.
Your next step

Could this gap become your thesis?

  1. 1Read the evidence and the open questions
  2. 2Talk to a mentor about fit, data and method
  3. 3Build a defensible proposal
See proposal support

Questions scholars ask

Has this not already been answered by the 2023 studies?

No. The early studies showed that detectors can fail, but later work reports different results under different conditions (Jiang et al., 2024; Pratama, 2025). The conditions that apply in real integrity cases remain largely untested.

Do I need to be a computer scientist to study this?

Not necessarily. Benchmark work needs some quantitative skill, but the questions about policy, judgement and procedure suit education, linguistics or law, ideally with an interdisciplinary supervisory team.

Will the topic date before I finish?

Specific tools will change, so frame the contribution around methods and decision processes, and report tool versions and test dates clearly (Dugan et al., 2024).

Sources

  1. Dugan, L., Hwang, A., Trhlík, F., Zhu, A., Ludan, J. M., Xu, H., Ippolito, D., & Callison-Burch, C. (2024). RAID: A shared benchmark for robust evaluation of machine-generated text detectors. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 12463–12492). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.acl-long.674
  2. Fleckenstein, J., Meyer, J., Jansen, T., Keller, S. D., Köller, O., & Möller, J. (2024). Do teachers spot AI? Evaluating the detectability of AI-generated texts among student essays. Computers and Education: Artificial Intelligence, 6, 100209. https://doi.org/10.1016/j.caeai.2024.100209
  3. 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, 105070. https://doi.org/10.1016/j.compedu.2024.105070
  4. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779
  5. Miles, D. A. (2017, August). A taxonomy of research gaps: Identifying and defining the seven research gaps [Paper presentation]. Doctoral Student Workshop: Finding Research Gaps, Research Methods and Strategies, Dallas, TX, United States. https://www.researchgate.net/publication/319244623
  6. Müller-Bloch, C., & Kranz, J. (2015). A framework for rigorously identifying research gaps in qualitative literature reviews. In Proceedings of the Thirty Sixth International Conference on Information Systems (ICIS 2015), Fort Worth, TX, United States. https://aisel.aisnet.org/icis2015/proceedings/ResearchMethods/2/
  7. Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., & Khuat, H. Q. (2024). Simple techniques to bypass GenAI text detectors: Implications for inclusive education. International Journal of Educational Technology in Higher Education, 21, 53. https://doi.org/10.1186/s41239-024-00487-w
  8. 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, e2953. https://doi.org/10.7717/peerj-cs.2953
  9. Vanderbilt University. (2023, August 16). Guidance on AI detection and why we're disabling Turnitin's AI detector. Brightspace at Vanderbilt. https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/
  10. Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, 26. https://doi.org/10.1007/s40979-023-00146-z

How to cite this page

Choose your referencing style, then copy.

PhD Research Lab Insights. (2026, October 3). AI-text detectors in integrity cases: how reliable and how fair are they for real student populations?. PhD Research Lab. https://www.phdassistance.com/insights/research-gap-ai-text-detectors/

PhD Research Lab Insights. “AI-text Detectors in Integrity Cases: How Reliable and How Fair Are They for Real Student Populations?.” PhD Research Lab, 3 Oct. 2026, www.phdassistance.com/insights/research-gap-ai-text-detectors. Accessed [access date].

PhD Research Lab Insights (2026) AI-text detectors in integrity cases: how reliable and how fair are they for real student populations?. PhD Research Lab. Available at: https://www.phdassistance.com/insights/research-gap-ai-text-detectors/ (Accessed: [access date]).

PhD Research Lab Insights. “AI-text Detectors in Integrity Cases: How Reliable and How Fair Are They for Real Student Populations?.” PhD Research Lab. October 3, 2026. https://www.phdassistance.com/insights/research-gap-ai-text-detectors.

PhD Research Lab Insights, ‘AI-text Detectors in Integrity Cases: How Reliable and How Fair Are They for Real Student Populations?’ (PhD Research Lab, 3 October 2026) <https://www.phdassistance.com/insights/research-gap-ai-text-detectors> accessed [access date].

PhD Research Lab Insights. AI-text detectors in integrity cases: how reliable and how fair are they for real student populations? [Internet]. PhD Research Lab; 2026 Oct 3 [cited [access date]]. Available from: https://www.phdassistance.com/insights/research-gap-ai-text-detectors/

PhD Research Lab Insights, “AI-text detectors in integrity cases: how reliable and how fair are they for real student populations?,” PhD Research Lab, Oct. 3, 2026. [Online]. Available: https://www.phdassistance.com/insights/research-gap-ai-text-detectors. [Accessed: [access date]].

@online{research_gap_ai_text_detectors_2026,
  author = {{PhD Research Lab Insights}},
  title = {{AI-text detectors in integrity cases: how reliable and how fair are they for real student populations?}},
  organization = {PhD Research Lab},
  date = {2026-10-03},
  url = {https://www.phdassistance.com/insights/research-gap-ai-text-detectors},
  urldate = {[access date]}
}

  • AI-text detection
  • Academic integrity
  • Algorithmic fairness
  • Non-native English writers
  • Higher education policy
AI-text detectors in integrity cases: how reliable and how fair are they for real student Research gap · All disciplines
Proposal support