Facing growing concerns about GenAI use, some instructors are turning to AI detection software such as GPTZero in an attempt to identify AI-generated student work. However, these tools have significant risks related to reliability, bias, and student privacy.
Given these limitations, the University of Minnesota does not centrally support or recommend the use of any GenAI detection tool. The most effective way to determine the role Generative AI played in a student’s work is to have a conversation about their process for completing the activity.
This resource outlines the key challenges associated with GenAI detectors and provides considerations for instructors navigating issues related to GenAI and academic integrity.
The Trouble with Trusting AI Detectors
While GenAI detection tools promise to offer a quick way to verify student authorship, they are far from transparent or definitive in their findings. Most detection tools operate by analyzing linguistic patterns and statistical anomalies in a textual sample and assigning a probability percentage score that indicates the extent to which the piece was AI-generated (Tiwari et al., 2024).
These scores are far from conclusive and can not be used to consistently detect AI-generated writing, which means that students can be falsely accused of misconduct.
Try It Yourself
Anecdotally, University faculty have found inaccurate results when submitting their own writing to GenAI detection tools. In many cases, a portion of their text has been flagged as AI-generated, even when entirely original or predates the emergence of GenAI!
Curious? Try submitting a piece of your writing and see how it is evaluated. Would you trust the results in your classroom?
Challenges with GenAI Detection Tools
Accuracy and Reliability
GenAI detection tools (CopyLeaks, Originality.ai, and GPTZero) cannot provide proof of AI authorship. These tools can merely assign an estimate of the likelihood that a text was created with GenAI based on patterns identified by its training and algorithmic model. Moreover, numerous scholarly studies have found that GenAI detectors are unable to reliably distinguish between AI and human-generated texts (Ardito, 2024; Chaka, 2024).
False Positives
Even tools that claim a low 1-2% false positive rate acknowledge the potential for error. Independent research, however, has found higher false positive rates, particularly for human-generated text (Elkhatat et al., 2023; Gegg-Harrison & Quarterman, 2024).
Any false positive result can have serious consequences for accused students, including:
- Undue stress and anxiety
- Disruption to financial aid, academic progress, or graduation
- Damage to instructor-student trust
Bias Against Marginalized Students
Studies have shown that GenAI detection tools are more likely to flag writing by:
- non-native English writers
- neurodivergent students
These groups may produce writing that deviates from the linguistic patterns GenAI detection tools associate with human-generated writing, resulting in disproportionate false accusations (Chaka, 2024; Liang et al., 2023, Gegg-Harrison & Quarterman, 2024).
Vulnerability to Adversarial Techniques
AI-generated text can easily be modified to avoid detection. Common “adversarial techniques” include prompting GenAI to:
- Add spelling or grammatical errors
- Vary sentence length
- Increase or decrease complexity
- Write as a non-native English speaker
- Change the tone, audience, mood, style, or skill level of a text
These methods deceive detectors, making detection easy to circumvent (Malik & Amjad, 2025; Perkins, 2024).
Data Privacy Concerns
As the University of Minnesota does not license a GenAI detection tool, use of such tools requires instructors to enter student coursework into unvetted 3rd-party tools. Faculty are not authorised to share student work with external companies without explicit written consent from the student.
Feeding your students’ work into these models also raises serious data privacy concerns:
- FERPA Violations: Student work is protected as part of their academic record, and entering it into these tools violates student privacy.
- Intellectual Property: Students retain rights to their original work; uploading it without permission infringes on those rights
- Data Use: Many tools do not disclose how uploaded data is stored or may be used to train future AI models.
Recommendations for Instructors
In light of the challenges described above, the use of AI detection tools is not recommended. Instead, consider alternative approaches:
- Foster a culture of ethical use. Include a strong statement about GenAI use in your syllabus, discuss GenAI and academic integrity with your students, and instruct students on how to acknowledge their use of GenAI appropriately.
- Use grading schemas that reward originality, accuracy, analysis, and voice. Focus on building skills that are harder to replicate with GenAI.
- Explain why it is important for students to develop their own ideas and be able to express them in their own words.
- Incorporate process-based assessments. Ask students to include drafts, outlines, reflections, and in-class writing in their final products.
See the guide GenAI-Resilient Assignments: Designing for Integrity, Curiosity, and Critical Thinking for additional ideas.
References
Ardito, C. G. (2024). Generative AI detection in higher education assessments. New Directions for Teaching and Learning, 1-18. doi.org/10.1002/tl.20624
Chaka, C. (2024). Accuracy pecking order – How 30 AI detectors stack up in detecting generative artificial intelligence content in university English L1 and English L2 student essays. Journal of Applied Learning & Teaching, 7(1). doi.org/10.37074/jalt.2024.7.1.33
Elkhatat, A. M., Elsaid, K., & Saeed, A. (2023). Evaluating the efficacy of AI content detection tools in differentiating between human and AI-generated text. International Journal for Educational Integrity, 19. doi.org/10.1007/s40979-023-00140-5
Gegg-Harrison, W., & Quarterman, C. (2024). AI detection’s high false positive rates and the psychological and material impacts on students. In S. Mahmud (Ed.), Advances in educational marketing, administration, and leadership, 199–219. IGI Global. doi.org/10.4018/979-8-3693-0240-8.ch011
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. doi.org/10.1016/j.patter.2023.100779
Malik, M. A., & Amjad, A. I. (2025). AI vs AI: How effective are Turnitin, ZeroGPT, GPTZero, and Writer AI in detecting text generated by ChatGPT, Perplexity, and Gemini? Journal of Applied Learning and Teaching, 8(1), 91-101. doi.org/10.37074/jalt.2025.8.1.9
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(1), 53. doi.org/10.1186/s41239-024-00487-w
Tiwari, S., Sharma, R., Sikarwar, R. S., Dubey, G. P., Bajpai, N., & Singhatiya, S. (2024). Detecting AI generated content: A study of methods and applications. Proceedings of International Conference on Communication and Computational Technologies, 161–176. doi.org/10.1007/978-981-97-7423-4_13