AI cheating in school: why detectors fail and what works instead

Short answer: AI detectors can flag honest writing as machine-made, so a detector score alone is weak evidence of cheating. Liang and colleagues (2023) found that seven detectors misclassified more than half of TOEFL essays by non-native English writers as AI-generated. Schools can respond with clear rules on permitted AI use and assessments that show a student's process.

EasWrk is a hosting platform, not a research institute. This essay summarises published sources, which are listed at the end. It connects to our other essays on AI and job displacement and AI journalism and trust, and to the research hub.

The Generative Turn in the Classroom

ChatGPT was released in November 2022. Within weeks, educators were asking whether students would use it to write essays and answer exam questions. Institutional responses varied widely, which showed deep uncertainty about what education is for and what credentials certify.

Stokel-Walker (2022), writing in Nature in December 2022, reported the first wave of concern about what ChatGPT actually was: not a better search engine or a more capable calculator, but a system that could produce, on demand, a competent essay on almost any topic in almost any academic register, without any of the cognitive struggle that the essay assignment was designed to provoke. The question that immediately arose, "should professors worry?", quickly revealed itself as the wrong question. The right question was: what are assignments for, what do they certify, and what happens to those purposes when the work can be trivially delegated to a machine?

Rudolph, Tan, and Tan (2023), in an early paper on ChatGPT and higher education assessment, asked whether traditional assessments could survive. The standard essay and written assignment formats that constitute the majority of undergraduate assessment were designed to measure a combination of knowledge, analytical skill, argumentation, and written communication. All of these could now be produced, at least superficially, by a sufficiently capable language model. The credential issued on the basis of such assessments would, if AI assistance was not controlled, certify the student's access to AI tools rather than their own competencies. Whether that credential still meant what institutions claimed it meant was a question that could not be deferred.

Cotton, Cotton, and Shipway (2024) examined the opportunities and challenges of ChatGPT in higher education, including the difficulty of detecting and preventing academic dishonesty. They suggested that universities develop policies and procedures, provide training and support, and use several methods to detect and prevent cheating. This essay's reading is that AI text differs from a purchased essay in three ways: it is effectively free, it is instant, and it can be hard to tell from competent student work by reading alone. Those differences change the scale of the problem more than its kind.

Perkins (2023) argued that it is not the use of an AI tool that defines a breach of academic integrity, but whether the student makes that use clear. He also demonstrated that text from large language models can avoid detection by existing tools and by trained academic staff, and concluded that institutional integrity policies need updating. In the author's view, the pressure falls first on disciplines that rely on written essays, such as the humanities, social sciences and business. STEM and professional disciplines face related questions about coding, problem sets and whether practitioners trained under older assessment are fit for purpose.

UNESCO published its guidance on generative AI in education and research in September 2023. It calls on governments to regulate generative AI in schools, to protect data privacy, to support a human-centred approach with teacher training, and to validate AI systems before students use them. It also raised a minimum age of 13 for classroom use as something to consider (UNESCO, 2023). The guidance is a framework. Applying it takes judgment at each institution about disciplines, students and assessment purposes.

Is Using AI Cheating in School? Academic Integrity in the Age of ChatGPT

Academic integrity frameworks, the policies, norms, and enforcement mechanisms through which universities define and respond to dishonesty, were built on assumptions that AI fundamentally challenged. The International Center for Academic Integrity (ICAI) defines academic integrity through six fundamental values: honesty, trust, fairness, respect, responsibility, and courage. The application of these values to AI use required resolving questions that the definition did not address: Was it dishonest to use an AI tool to improve a draft if the policy did not prohibit it? Was it a violation of trust to use AI for brainstorming if the final text was genuinely the student's own? Was it fair that some students had access to premium AI tools while others did not?

Institutional responses ranged from explicit prohibition to explicit endorsement, with a wide range of ambiguous middle positions. Some institutions updated their academic integrity policies to define unauthorised AI use as a form of academic misconduct equivalent to plagiarism. Others revised their policies to permit or even encourage AI use, with requirements for disclosure and reflection on how AI had been used. Many institutions operated in a state of policy ambiguity. A UNESCO survey of more than 450 schools and universities in May 2023 found that fewer than 10% had institutional policies or formal guidance on generative AI (UNESCO survey, 2023). Individual faculty were left to set their own course-level expectations, which creates inconsistency that disadvantages students who are unsure of the norms.

The credential problem is easy to picture. A medical student who uses AI to complete clinical case analyses may never develop diagnostic reasoning. A law student who uses AI to write briefs may never develop legal analysis skills. A teacher education student who uses AI to write lesson plans may never develop instructional design capability. Each could hold a credential that implies competence they lack. These are illustrations, not documented cases. The downstream risk for patients, clients and schoolchildren is the reason the question matters.

The question of what constituted "use" versus "assistance" versus "fraud" was genuinely difficult to resolve cleanly. Using a spell-checker was universally accepted; using a grammar checker was almost universally accepted; using a tool that improved sentence structure and clarity was widely accepted; using a tool that generated content from scratch was, in most institutional policies, a violation. But the line between these uses was not crisp. Grammar tools that suggest whole-sentence rewrites could already reshape prose while leaving the student as the nominal author. The novelty of ChatGPT was one of degree rather than kind, but the degree was sufficient to change the practical question from "did this student write this?" to "is there any meaningful sense in which this work is this student's?"

Floridi and colleagues (2018), in the AI4People framework, list autonomy among five principles for a good AI society. Applied to education, this is clarifying: AI use in assessment was educationally harmful to the extent that it substituted for, rather than augmented, the development of student autonomy: the capacity to identify problems, formulate questions, gather evidence, reason through uncertainty, and produce conclusions independently. Assessment formats that could be fully delegated to AI were, by this framing, revealing themselves to be poor instruments for evaluating the capacities education was supposed to develop.

The sociological dimension is not limited to individual student dishonesty. What follows is the author's opinion, not a finding. Institutional incentives may have shaped the response in ways that were not always educationally optimal. Universities that depend on tuition revenue may have reasons to avoid rigorous enforcement that could end in dismissals or degree revocations. Reputational concerns may also discourage public acknowledgment of integrity problems at scale. Some policy updates may signal concern without creating the means to enforce it.

AI Detector False Positives and Why Detection Fails

The market for AI detection tools, software designed to identify AI-generated text, expanded rapidly after the release of ChatGPT, driven by faculty demand for technical solutions to the integrity problem. Tools including Turnitin's AI writing detection feature, GPTZero, Originality.ai, and several others attracted institutional adoption. Independent tests have found limits that vendor benchmarks do not show.

The core technical challenge was that AI-generated text and high-quality human writing shared many statistical properties. Language models trained to produce fluent, coherent, contextually appropriate text necessarily produced text that was statistically similar to the fluent, coherent, contextually appropriate text in their training data, which was primarily written by humans. Detection tools attempting to identify AI-generated text based on statistical properties of the text itself were therefore operating on a signal that was inherently noisy, and whose noise level increased as AI models became more capable.

Independent evaluations have found false positives, cases where genuine student writing is flagged as AI-generated, that fall unevenly. Liang and colleagues (2023), a Stanford-led team, tested seven GPT detectors on 91 TOEFL essays written by non-native English speakers. The detectors misclassified more than half as AI-generated, an average false positive rate of 61.22%, while essays by native English writers were classified accurately. The authors point to text perplexity: simpler word choices read as machine-like to the detectors. A student who learned English as a second language and writes in a careful, accessible style may be at greater risk of wrongful accusation than a native speaker writing in a more varied register.

Turnitin's detector was not among the seven tested by Liang and colleagues (2023). In October 2023 Turnitin reported that for documents of at least 300 words the false positive gap between non-native and native English writers was small and not statistically significant, although the gap was larger on shorter documents (Turnitin, 2023). Turnitin said its target false positive rate was 1%. Vanderbilt University, which submitted 75,000 papers to Turnitin in 2022, noted that 1% of that volume is about 750 papers, and in August 2023 it disabled the AI detector (Vanderbilt University, 2023). A student accused on the strength of a detector score alone has little to answer, which is why the score is better treated as one signal among several.

Detection is also easy to evade. Perkins (2023) demonstrated that text from large language models can avoid existing detection tools. In the author's view, a student who is set on cheating faces little technical barrier, so detection alone cannot carry an integrity policy.

Cotton, Cotton, and Shipway (2024) recommend using various methods, not one tool, to detect and prevent cheating. This essay's reading of the evidence is that detection is an arms race in which the advantage lies with the generator, and that the cost of false positives falls on particular students. That points toward redesigning assessment rather than policing existing formats.

The Widening Digital Divide

The narrative around AI and education has often assumed, implicitly, a world in which all students have equivalent access to AI tools and the relevant policy question is how institutions should respond to a technology that is universally available. This assumption was inaccurate when ChatGPT was released and became more so as AI tools diversified into free and premium tiers with significantly different capabilities.

Noble (2018), studying how search engines reinforce racism, showed how technical systems embedded in unequal social structures can reproduce existing inequalities. A similar pattern is plausible for AI access. Students who can pay for premium AI subscriptions and have fast internet are likely to get more capable AI assistance than students who rely on free tiers or have no connection.

The scale of the connectivity gap is large. UNICEF and the International Telecommunication Union (2020) estimated that about 1.3 billion children aged 3 to 17, roughly two thirds of the world's school-age children, had no internet connection at home. About 60% of school-age children in urban areas lacked a home connection, against about 75% in rural households. COVID-19 school closures made these gaps visible.

Selwyn (2019) challenges the hype around AI in education and argues that the questions are matters of values, judgement and politics, not only technical capability. This essay applies that lens to access. Students with more social capital, more educated parents, access to tutoring and experience of academic expectations are probably better placed to use AI as a cognitive tool rather than a crutch. That last point is the author's inference, not a finding of the cited source.

The credential implications of differential AI access were serious. If assessments became implicitly competitions in AI-assisted performance, the credential would certify, in part, the quality of AI assistance available to the student rather than the student's own capabilities. This would represent a new mechanism for reproducing existing social inequality through educational credentialing, adding to rather than replacing existing mechanisms including differential access to tutoring, to preparatory schools, and to the social capital that helps students find their way through educational institutions.

Surveillance Technologies in Education

Institutional responses to academic integrity challenges increasingly involved not only assessment redesign and policy revision but surveillance, deploying technology to monitor students in ways designed to deter or detect dishonest behaviour. The COVID-19 pandemic pushed many institutions to adopt automated proctoring tools for remote examinations.

Proctorio, an automated proctoring platform, drew sustained controversy in 2020 and 2021, much of it in court. In 2020 the company sued Ian Linkletter, a University of British Columbia learning technology specialist, over tweets that linked to its training videos (The Verge, 2020). In April 2021 the Electronic Frontier Foundation sued Proctorio on behalf of Miami University student Erik Johnson, who said the company had misused a copyright takedown notice against his tweets criticising the software (TechCrunch, 2021).

In the author's view, these cases show a risk for institutions: companies selling surveillance tools have commercial reasons to resist criticism, and a school that has already signed a contract is poorly placed to test a vendor's claims independently.

Surveillance in education extends beyond proctoring. Learning management systems record data such as login times and assignment submission timing, and some institutions use that data to flag students as at risk.

Floridi and colleagues (2018) list autonomy among five principles for AI. In education, that principle sits in tension with institutional monitoring of student behaviour. Education has historically created conditions for cognitive risk-taking: the willingness to attempt difficult tasks, to make and learn from errors, and to develop positions that might be wrong. In the author's view, comprehensive surveillance is inconsistent with those conditions.

How to Prevent AI Cheating in School Essays: Assessment After ChatGPT

The disruption of standard assessment formats by AI was not an accident or an oversight. It was the predictable consequence of a disjunction between what assessments were formally designed to measure and what they were actually capable of measuring. Most written assessments, including essays, problem sets, case analyses, and literature reviews, were designed to measure knowledge and analytical capability by proxy: if a student could produce a sufficiently competent discussion of a topic, they probably understood it. The proxy worked when producing the discussion required genuine engagement with the material. When AI could produce the discussion without genuine engagement, the proxy broke.

The deeper question this revealed was whether the written product was ever an adequate proxy for the capability it purported to measure, or whether it had always been an imperfect proxy whose limitations had simply not been exposed because the technology needed to exploit those limitations did not exist. Research on deep and surface approaches to learning (Marton & Säljö, 1976) has long distinguished students who seek understanding from those who reproduce material, and the second group can often do adequately on standard assessments. AI made this gap visible rather than creating it.

A useful distinction in the AI era is between "performance on a task" and "evidence of capability". A performance on a task that could be delegated to AI was not evidence of the capability the task was designed to assess. The validity of the assessment, whether it actually measured what it claimed to measure, was undermined not by dishonest students but by the existence of a tool that could produce the required performance without the required capability.

Several alternative assessment formats are often suggested as more resistant to AI assistance: oral examinations, which required real-time synthesis and communication; practical demonstrations, which required the physical performance of skills; reflective portfolios documenting a learning process over time; in-class handwritten assessments without access to digital tools; collaborative projects assessed through contribution analysis and process documentation. Each of these formats had genuine advantages and genuine disadvantages in terms of equity, scalability, accessibility, and the specific capabilities they could assess. None was universally applicable across disciplines and student populations.

Assessment reform can create new inequities if it is not designed with care. In the author's view, in-person oral examinations may favour students with strong verbal skills and comfort with interpersonal evaluation, and portfolio assessment demands sustained self-organisation. Both are unevenly distributed across student populations. Substituting formats without attending to those effects would reproduce the underlying problem.

Reimagining Pedagogy for the AI Era

Beyond the immediate crisis of assessment integrity lay a more fundamental question: what should education do in a world where AI can perform many of the cognitive tasks that formal education has historically trained people for? The question was not new. Similar questions had been raised by calculators, by the internet, and by search engines. But AI's generality of capability gave it a scope that previous tools had not had. A calculator automated arithmetic; the internet automated information retrieval; AI was automating significant portions of reasoning, synthesis, and communication. The scope of what remained uniquely human in the cognitive tasks that education prepared people for had narrowed significantly.

UNESCO's (2023) guidance stresses a human-centred approach, with teacher training and validation of AI systems before students use them. In this essay's view, that means more than teaching students to use ChatGPT. It involves the critical judgment to evaluate AI outputs, awareness of AI limitations and biases, the ethical reasoning to decide when AI use is appropriate, and the metacognitive capacity to understand what a student is and is not learning when AI assists.

Selwyn (2019) argues that debates about AI in education are matters of values, judgement and politics. In this essay's reading, that directs attention away from the technical capabilities of AI tools and toward the conditions in which they are deployed: the power relationships between students and institutions, the economic interests of vendors, and the cultural assumptions embedded in the text the systems learn from.

Four of the five principles in Floridi and colleagues (2018) translate directly into educational contexts. Beneficence: AI should benefit students, not merely be efficient for institutions. Non-maleficence: AI should not harm students, including through privacy violations, false accusation, or the production of competency gaps. Autonomy: AI should support the development of student agency and independent judgment, not substitute for it. Justice: the benefits and burdens of AI in education should be equitably distributed.

In the author's view, the most productive pedagogical response to AI is likely to involve treating AI as a subject of study as much as a tool for study, developing in students the ability to work with AI critically, to understand its outputs as products of specific training processes with specific limitations, and to situate their own intellectual work in relation to what AI can and cannot do. This required faculty development at a scale that most institutions had not undertaken. Educators who had not themselves developed facility with AI tools could not effectively design AI-integrated pedagogy or guide students through the ethical and intellectual challenges AI posed.

The credential question remained unresolved. If educational credentials were to continue signifying genuine competency rather than access to AI tools, the competencies they certified would need to be either inherently difficult to delegate to AI or verified through assessment processes that could not be meaningfully assisted by AI. Both requirements pointed in the same direction: toward the development of deeper, more authentic, more transferable competencies rather than the performance of tasks on demand. Whether educational institutions had the political will, the economic resources, and the institutional flexibility to make that shift was uncertain. What was certain was that the question of what education was for, and what it owed the students and societies it served, had never been more urgently in need of a serious answer.