AI Detectors: Evidence and Recommendations for Use in Education

AI detectors are an increasingly heterogeneous category of software and tools that claim to provide information regarding the likelihood that a written text has in some way been generated, altered, or otherwise shaped, in part or in whole, by generative AI tools that use large language models (LLMs). The first AI detector was launched in January 2023 and the market has since expanded to include dozens of freestanding tools, usually accessed via web browser, as well as added features within existing software platforms.

To use an AI detector, the user pastes text or uploads a file into the browser-based tool and after a brief analysis the tool provides a dashboard with an overall summary and, often, line by line breakdowns similar to what we’re accustomed to seeing with “originality” or “similarity” checkers such as TurnItIn. (Importantly, what AI detectors are doing is very different from what similarity checkers are doing.)

Based on the available evidence, AI detectors exhibit a wide range in efficacy for positive and negative detection as well as false positives and false negatives. In addition, the tools’ efficacy is not fixed; their effectiveness changes as LLMs and generative AI tools are released, altered, updated, and retired, often on a time scale much shorter than an academic semester.

AI detectors remain a controversial strategy for enforcing academic integrity in assessments. (A recent career feature in Nature covers the issues as they stand during the summer of 2026.) Researchers continue to assess the efficacy of AI detector tools in connection with a range of variables and contextual factors (see the references at the end of this document for a partial list of those studies). Processing tracking has been proposed as a less harmful alternative to AI detectors and comprises both dedicated software as well as informal methods such as adding versioning review into the assessment workflow. This approach, however, introduces its own limitations and risks.

CELT does not recommend the use of AI detectors for educational assessment purposes. (Note that UK does not have an institutional license for an AI detector.)

AI detectors look for statistical patterns in a text to determine the probability that it was produced exclusively or in part by an LLM-based AI tool. AI detectors use a range of methods, often more than one, with the most prevalent being (1) analyzing the style and composition of text against known text-output patterns of LLM-based tools and (2) drawing from machine learning on datasets with both AI-generated and human-authored texts, during which the detector “learns” patterns unique to the two categories. Some AI detectors provide an overall estimate of the likelihood that a text has been partially or completely generated by AI, while others provide a segment-by-segment breakdown that indicates which parts of a text may be AI-generated or altered in some way. Some detectors have begun to add a feature that estimates whether text has been “humanized” after AI-based generation, owing to the rise of “humanizer” tools that purport to use AI to render an AI-generated text as “human-looking” to AI detectors.

Despite advancements (both verified and merely claimed) in AI detectors, they operate within several fundamental limitations:

• They cannot verify their findings outside of their statistical analyses; all of it remains an estimation or probability. (Compare this to similarity checkers, which can point to other texts or documents that contain similar or the same language.) Ultimately, the number that an AI detector provides is the end result of sophisticated guesswork, not fact.
• They remain prone to false positives. These can occur for a number of reasons. For instance, some writers, and some writing, may be more prone to being flagged by AI detectors simply because they use more common, predictable, or unvaried words and phrasing.
• LLMs were developed specifically to emulate human writing and they continue to be optimized for this purpose. The majority of the data on which LLMs were trained is human-authored text. Thus, human writing may be flagged as LLM-authored simply because it looks statistically similar to text that, in turn, attempts to be statistically similar to human writing.
• AI detectors need to keep up with the developments in LLM-based generative AI. New models or updates mean new statistical “fingerprints” that AI detectors must respond to. Moreover, between the model releases and updates, AI tools frequently undergo tweaks in their system prompts, model weights, etc. In other words, AI detectors are leveraging a complex statistical analysis against a constantly moving target.
• Many genres and writing situations discourage stylistic and compositional choices that diverge from established norms. Boilerplate, narratives of specific methods, definitions of well-established concepts, and discursive conventions around phrasing and word choice can render some writing situations more at risk for false positives from AI detectors.
• Similarly, unusual or unique genres and writing situations are not as well represented in the training datasets for AI detectors. Their accuracy may be less dependable when they deal with texts and writing approaches for which they are less prepared or optimized.
• LLMs may be prompted to output text in ways that strategically diverge from their output conventions. Additional human and AI-assisted modification of AI-generated text often renders it invisible to AI detectors. Studies have shown that altering syntax, language, style, and organization while also introducing human “tics” such as typos, inaccuracies, and surprising choices can bypass AI detectors’ thresholds.
• AI detectors analyze for the presence of AI-generated text but do not infer how the AI- generated text was used. For example, an AI detector score lacks the context to discern whether a paragraph was generated entirely by AI or if a student drafted the paragraph themselves and then asked an AI tool to clean it up or copyedit it. (Both cases may still run afoul of a course or assignment policy.)

In addition to the limitations above, there are several risks in using AI detectors to assess the provenance of student-submitted work in educational settings.

• Privacy and legal compliance (e.g., FERPA) are implicated for any education record. Submitting student information included in copied text or uploaded files effectively shares that information with an external organization. Even if a text does not contain specific student records such as name, course, etc., it still may have information that could be used to infer the student’s identity, e.g., unique personal experiences or ideas.
• In addition, some institutions of higher education have developed policy or guidance around the relationship of student intellectual property and third-party services such as AI detectors. Even if the text in question does not risk disclosing FERPA-protected information, it remains the student’s intellectual property. Sharing student work with organizations or services that do not have a contract or agreement with the University requires the intellectual property owner’s consent. (For instance, AI detectors may store and use uploaded text for product improvement and other purposes. Their terms of use may require that users have the rights or at least the permissions to upload content.)
• Reliance on AI detectors risks creating an atmosphere of suspicion and surveillance in a course, damaging the rapport between instructor and students and compromising the conditions under which students learn effectively. Even if no accusation is made, the omnipresence of AI detectors communicates particular assumptions about students that detract from student belonging, learning, and academic performance.
• In addition to the effect on overall course climate, there is a growing record of students and families alleging that schools, colleges, and universities have wrongfully accused or taken punitive action against them as a result of an AI detector score (or undisclosed methods). Cases such as these may place the burden of proof on students who are not familiar with navigating the academic misconduct process, while a wrongful accusation can have an outsized effect on a student’s experience of their educational journey, their relationship with the school, and their self-conception as a rising scholar or professional.
• Automation bias (the tendency to treat machine output as more accurate, objective, or authoritative than it really is) may lead some users to interpret AI detector results as definitive, even if those users are also aware of the limitations and risks of AI detectors. Similarly, quantification bias leads to similar effects in that numerical expressions may be received as inherently objective, rigorous, or trustworthy. Broadly, the McNamara fallacy also potentially applies to this scenario; the fallacy refers to excessive valuation of what can be measured—in this case, the methods that AI detectors use to analyze text—at the expense of establishing whether those measurements represent valid constructs.
• AI detectors will likely not flag all students’ work, leading to a “scrutiny asymmetry effect” by which some students’ work is heavily scrutinized for provenance while other students’ work is not, even if the evidence suggests that AI detectors produce false negatives at a rate that usually exceeds false positives.

This question is fiercely debated by researchers, educators, and other stakeholders. Some research suggests that some AI detectors are mostly effective at distinguishing text that has been generated entirely by LLM-based AI tools and submitted in unmodified form. CELT does not recommend any particular AI detector tools, though some initial research may identify tools that perform better than others under controlled or experimental conditions. Some instructors use AI detectors carefully and transparently as part of a larger trust strategy with students. It is important to note that the threshold for effective and ethical use of AI detectors is much higher for teaching and learning settings because of educational institutions’ obligation to learners, particularly at a land-grant, public flagship university whose strategic plan emphasizes putting students first, ensuring trust, transparency, and accountability, and serving the state.

CELT does not recommend the use of AI detectors for educational assessment purposes. (Note that UK does not have an institutional license for an AI detector.)

Additionally, CELT strongly discourages the use of prompt injections, trojan horses, and other deceptive strategies by which hidden instructions are embedded in assignment descriptions and other course material so that, if they are used in a prompt, AI output exhibits unique markers. This places instructors and students in an adversarial relationship that undermines learning and trust while conveying implicit messages about how students are seen.

Many instructors have explored how assessments may be adapted to reduce the likelihood of inappropriate generative AI use. First, however, a few caveats are worth reviewing.

• Assessments best serve learning and students when they triangulate authenticity, validity, and proportionality. Authenticity refers to the involvement of tasks, criteria, or scenarios that engage students in meaningful practices that are situated in disciplinary, professional, or realistic contexts. Validity refers to the degree that an assessment allows the instructor to interpret the outcomes and/or skills for which it purports to generate evidence of learning. Proportionality refers to alignment with levels of learning (such as first-year undergrad versus capstone masters), levels of outcomes (such as lower versus higher complexity), and scope of time and effort (such as one week versus one month in a class). If assessment modification moves beyond this triangulation, its primary function for learning has been deprioritized.
• Many instructors have moved towards more process-oriented assessments that either segment larger tasks into smaller ones with checkpoints or require that students submit additional deliverables attesting to that process. While this may allow the instructor to observe student learning better, it also may place a documentation burden on students that distracts from the main goals of the assessment. Balancing the tension between surfacing process and focusing effort is key for assessment redesign. Similarly, adding more touchpoints of observation, interaction, or feedback may prohibitively increase the demand on instructors’ time and focus.
• The allure of so-called “AI-proof” assessment strategies is a powerful one when the grounds of academic integrity seem unstable. However, given the versatility of LLMs and other media-based AI tools, as well as prompting strategies and other AI fluency skills, it is unlikely that any deliverable is immune from simulated academic performance via the output of AI tools. Direct observation is the only concrete method for verifying the use or non-use of AI tools during an assessment, but this may also not triangulate well with authenticity, validity, and proportionality. Limitations such as paper-only activities (with accessibility exceptions) may align with some circumstances while generating barriers, challenges, or misalignments in other circumstances.
• Overall, assessments are powerful tools for communicating an instructional stance and even a philosophy of teaching. The design of assessments might follow primarily from a desire to prevent and/or detect transgressions, or it might follow primarily from a desire to create meaningful learning opportunities. Those poles are not necessarily mutually exclusive, but the latter will take priority with learning-centered design and pedagogy.

Given those caveats, instructors may adapt assessments strategically in the following ways.

• Communicate specific AI policies for assessments using tools such as the CELT GAI Use Scale.
• Use frameworks such as transparent assignment design to communicate expectations clearly to students.
• Segment more complex and long-term assessments into parts or iterations with checkpoints and feedback.
• Use specific course materials, local contexts, recent class discussions, datasets, cases, or constraints when they meaningfully support the learning goals of the assignment.
• Include time during class meetings for students to work intentionally on specific aspects of larger assignments with instructor interaction or engagement.
• Frontload low-stakes activities with formative feedback that allow students to gain momentum, purpose, and self-efficacy in their work.
• Include moments or steps when students speak to their work or process in some way.
• Incorporate revision or retries if appropriate.
• Indicate in syllabi or assignment instructions that students may be asked to meet to discuss questions or concerns if they arise, and potentially to revise work accordingly.
• Allow students to use AI tools in particular ways that do not compromise the value of the assessment and to document their AI use clearly.
• Lead students in activities involving the critique or analysis of AI-generated information, text, artifacts, arguments, etc., with specific reference to disciplinary expertise.
• Shift homework or assignment activities to class meetings and move content coverage into homework (i.e., flipping), but be cognizant of support that students may need if they struggle with learning the content.
• Consider the best genre, form, and modality of expression for the assessment given the learning goals of the assessment and course.
• Align grading criteria, weights, and feedback with the outcomes and overall purpose of the assessment. Evaluate student work for learning and growth rather than polish unless the latter is important for the assessment.

CELT regularly consults with individuals and groups on issues and practices related to AI in teaching, learning, and higher education. In addition, we facilitate workshops on assignment design with and/or without AI by request for groups that are interested. Each semester, we offer a series of AI-focused events for the university community.

Alshammari, H., & Rao, P. (2025). Evaluating the performance of AI text detectors, few-shot and chain-of-thought prompting using DeepSeek generated text (arXiv:2507.17944). arXiv. https://doi.org/10.48550/arXiv.2507.17944 

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

Chaka, C. (2024). Reviewing the performance of AI detection tools in differentiating between AI-generated and human-written texts: A literature and integrative hybrid review. Journal of Applied Learning & Teaching, 7(1), Article 14. https://doi.org/10.37074/jalt.2024.7.1.14 

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, Article 66. https://doi.org/10.1186/s41077-025-00396-6 

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. The International Journal of Management Education, 21(2), Article 100822. https://doi.org/10.1016/j.ijme.2023.100822

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, Article 17. https://doi.org/10.1007/s40979-023-00140-5

Erol, G., Ergen, A., Gülşen Erol, B., Kaya Ergen, Ş., Bora, T. S., Çölgeçen, A. D., Araz, B., Şahin, C., Bostancı, G., Kılıç, İ., 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, 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

Garland, N. A. (2026). AI detectors fail diverse student populations: A mathematical framing of structural detection limits (arXiv:2603.20254). arXiv. https://doi.org/10.48550/arXiv.2603.20254 

Gehring, L., & Paaßen, B. (2025). Assessing LLM text detection in educational contexts: Does human contribution affect detection? (arXiv:2508.08096). arXiv. https://doi.org/10.48550/arXiv.2508.08096 

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, Article 4. https://doi.org/10.1007/s40979-026-00213-1 

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 

Jabarian, B., & Imas, A. (2025). Artificial writing and automated detection (NBER Working Paper No. 34223). National Bureau of Economic Research. https://doi.org/10.3386/w34223 

Kar, S. K., Bansal, T., Modi, S., & Singh, A. (2025). How sensitive are the free AI-detector tools in detecting AI-generated texts? A comparison of popular AI-detector tools. Indian Journal of Psychological Medicine, 47(3), 275–278. https://doi.org/10.1177/02537176241247934

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. https://doi.org/10.1016/j.patter.2023.100779

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, Article 8. https://doi.org/10.1007/s40979-024-00155-6

Liu, Z., Yao, Z., Li, F., & Luo, B. (2024). On the detectability of ChatGPT content: Benchmarking, methodology, and evaluation through the lens of academic writing. In Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (pp. 2236–2250). Association for Computing Machinery. https://doi.org/10.1145/3658644.3670392

Orenstrakh, M. S., Karnalim, O., Suárez, C. A., & Liut, M. (2024). Detecting LLM-generated text in computing education: A comparative study for ChatGPT cases. In 2024 IEEE 48th Annual Computers, Software, and Applications Conference (COMPSAC) (pp. 121–126). IEEE. https://doi.org/10.1109/COMPSAC61105.2024.00027

Pak, T. K., Han, E. C., Montelongo Hernandez, C., Collins, K., Carrasco, A., Nekovei, A., King, D., Robinson, D. M., & Brenner, A. M. (2025). Accuracy of artificial intelligence detection software for residency personal statements. Journal of Graduate Medical Education, 17(6), 722–726. https://doi.org/10.4300/JGME-D-25-00420.1 

Paullet, K., Pinchot, J., Kinney, E., & Stewart, T. (2025). Utilizing GPTZero to detect AI-generated writing. Information Systems Education Journal, 23(6), 44–52. https://doi.org/10.62273/PZWW7741 

Perkins, M., Roe, J., Postma, D., McGaughran, J., & Hickerson, D. (2024). Detection of GPT-4 generated text in higher education: Combining academic judgement and software to identify generative AI tool misuse. Journal of Academic Ethics, 22(1), 89–113. https://doi.org/10.1007/s10805-023-09492-6

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, Article 53. https://doi.org/10.1186/s41239-024-00487-w

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 

Pudasaini, S., Miralles-Pechuán, L., Lillis, D., & Llorens Salvador, M. (2024). Survey on plagiarism detection in large language models: The impact of ChatGPT and Gemini on academic integrity. Journal of Academic Ethics, 23, 1137–1170. https://doi.org/10.1007/s10805-024-09576-x 

Pudasaini, S., Miralles-Pechuán, L., Lillis, D., & Llorens Salvador, M. (2025). Benchmarking AI text detection: Assessing detectors against new datasets, evasion tactics, and enhanced LLMs. In F. Alam, P. Nakov, N. Habash, I. Gurevych, S. Chowdhury, A. Shelmanov, Y. Wang, E. Artemova, M. Kutlu, & G. Mikros (Eds.), Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect) (pp. 68–77). International Conference on Computational Linguistics. https://aclanthology.org/2025.genaidetect-1.4/

Pudasaini, S., Miralles-Pechuán, L., Lillis, D., & Llorens Salvador, M. (2026). Why AI-generated text detection fails: Evidence from explainable AI beyond benchmark accuracy (arXiv:2603.23146). arXiv. https://doi.org/10.48550/arXiv.2603.23146 

Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., & Feizi, S. (2024). Can AI-generated text be reliably detected? (arXiv:2303.11156). arXiv. https://doi.org/10.48550/arXiv.2303.11156

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 

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, Article 16. https://doi.org/10.1007/s40979-026-00226-w 

Walters, W. H. (2023). The effectiveness of software designed to detect AI-generated writing: A comparison of 16 AI text detectors. Open Information Science, 7(1), Article 20220158. https://doi.org/10.1515/opis-2022-0158

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, Article 26. https://doi.org/10.1007/s40979-023-00146-z