AI detectors do not compare a student’s essay against a database of known AI output the way plagiarism checkers compare text against existing sources. They cannot. Every response a generative AI produces is, technically, unique. Instead, detection tools analyze statistical properties of the text itself.
The two most common signals they rely on are perplexity and burstiness.
Perplexity measures how predictable a piece of text is. Language models are trained to generate the most statistically likely next word given what came before it. As a result, AI-generated text tends to be very predictable: it flows smoothly, uses common word choices, and rarely surprises. Human writing tends to have higher perplexity because people make idiosyncratic choices, use unusual phrasings, and sometimes write in ways that are stylistically distinctive but statistically unlikely.
Burstiness measures variation in sentence structure and length. Human writers naturally alternate between short punchy sentences and longer, more complex ones. AI-generated text tends to maintain a more consistent sentence rhythm throughout a passage, because the model is optimizing for coherent output rather than stylistic variety.
Detection tools train classifiers on large samples of known human writing and known AI-generated writing. They learn which patterns in those two signals, and others like them, tend to predict AI origin. When you submit a new text, the tool scores it against those learned patterns and returns a probability estimate.1
That is the key word: probability. The tool is not detecting AI authorship. It is estimating the likelihood that the text’s statistical properties resemble AI output more than human output. That is a meaningful distinction, and it is where the problems begin.
A false positive occurs when a detector flags human-written text as AI-generated. This is not a rare edge case. It is a systematic and documented problem.
In 2023, researchers at Stanford University tested several leading AI detection tools against writing samples from non-native English speakers.
2 They found that essays written by international students were flagged as AI-generated at dramatically higher rates than essays written by native English speakers. The reason is straightforward: non-native speakers often write in ways that are syntactically consistent, favor common vocabulary, and produce lower perplexity scores. The same statistical fingerprint the tools associate with AI.
The implications are serious. A detection tool trained primarily on native English writing will systematically disadvantage students who learned English as a second or third language. It will penalize clarity, careful sentence construction, and conservative word choice. Simultaneously, these are qualities many teachers and writing instructors actively encourage.
False positives also appear in other categories of legitimate human writing:
- Highly structured writing, such as lab reports, legal briefs, or technical summaries, tends toward low perplexity because the genre demands precision and consistency
- Students who have received extensive writing instruction and practice clear, direct prose may score higher for AI likelihood than students whose writing is less polished
- Edited drafts, where a student has revised for clarity and removed informal constructions, can register as more “AI-like” than the rougher first draft
- Some subjects, particularly STEM fields, require factual, declarative writing that naturally exhibits the low perplexity profile detectors treat as suspicious
Education Week documented multiple cases in which students with strong academic records received failing grades or academic misconduct charges based on detection reports, only for the accusations to be withdrawn after further review.3
The other side of the problem is equally important. Detection tools miss a substantial amount of AI-generated content. This is called a false negative: text that was written by AI but scored as human.
As generative AI models have become more sophisticated, the statistical signatures detectors rely on have become harder to identify.4 More recent models produce text with greater variety in sentence structure, more idiosyncratic word choices, and higher perplexity scores than earlier versions. In other words, the models have gotten better at writing in ways that look human, which means the detectors trained on older AI output are increasingly out of date.
Students have also learned that detection can be circumvented. Common workarounds include:
- Asking the AI to write in a more casual or personal voice, which increases burstiness and perplexity
- Running AI output through a paraphrasing tool before submission
- Using the AI to generate an outline or argument structure, then writing the prose themselves
- Editing AI output sentence by sentence to introduce stylistic variation
None of these approaches require technical sophistication. They are being discovered and shared among students through social media and peer networks. A student motivated to use AI without detection and willing to invest ten additional minutes can routinely beat the tools that schools are paying to catch them.
Accuracy claims are usually derived from controlled test conditions: a balanced dataset of known AI and known human samples, often drawn from similar domains and writing styles. Real-world classroom use is messier. Students write across subjects, skill levels, languages, and genres. The test conditions that produce a 98% accuracy headline may bear little resemblance to the conditions under which the tool is actually used.
Turnitin, one of the most widely adopted tools in schools, has noted in its own documentation that its AI detection feature is intended to support educator judgment, not replace it, and that no detector should be used as the sole basis for an academic integrity finding.5 That is a meaningful disclaimer. It is also one that does not always make it into the conversation when a teacher receives a detection report and has to decide what to do with it.
GPTZero, another widely used tool, similarly acknowledges in its FAQ that false positives occur and recommends that results be treated as one signal among many rather than as definitive evidence.6
When vendors themselves counsel against relying solely on their products, schools that do rely solely on those products are operating outside the intended use of the technology.
Academic integrity accusations carry real consequences. A student found responsible for academic misconduct may receive a failing grade, be required to repeat coursework, be placed on academic probation, or have a notation added to their record. In competitive academic environments, those consequences can affect college admissions, scholarships, and future opportunities.
The Washington Post documented cases in which students, including honor students with no prior disciplinary history, were accused of AI use based entirely on detection tool output.7 In several of those cases, the students could demonstrate, through saved drafts, writing process documentation, and teacher testimony, that the work was their own. The accusations were ultimately dropped. But the process itself was damaging: weeks of uncertainty, strained relationships with teachers and administrators, and the experience of having one’s integrity questioned without sufficient evidence.
For students who cannot easily document their writing process, or who lack the confidence to challenge an accusation from a position of authority, the outcomes can be worse.
There is also a chilling effect that extends beyond individual cases. When students know their work may be flagged regardless of how they wrote it, some respond by deliberately writing below their ability level, choosing simpler vocabulary and shorter sentences to avoid triggering a detector. The tool meant to protect academic integrity ends up discouraging the kind of clear, precise writing that academic training is supposed to develop.
None of this means schools should stop caring about academic integrity or ignore the real challenge that generative AI presents to writing-based assessment. It means the tools currently available cannot bear the evidentiary weight that many schools are placing on them.
A more reliable approach focuses on the writing process rather than the final product. Practices that support genuine assessment without depending on flawed detection technology include:
- Requiring students to submit drafts, outlines, or process notes alongside final work, which documents the development of ideas over time
- Incorporating in-class writing components that establish a baseline for each student’s voice and ability
- Designing assignments that require personal reflection, local context, or specific class discussion content that an AI cannot access
- Having conversations with students about their work: asking them to explain their argument, describe their research process, or expand on a specific paragraph
- Treating detection tool output as a prompt for that conversation rather than as a finding in itself
These approaches require more effort than uploading a document to a detection service. They also actually work. A student who can fluently discuss the argument in their own essay has demonstrated something a detection score cannot: that they engaged with the material.
