Do AI Detectors Work, and What Happens When They Get It Wrong?

The question "do AI detectors work" sounds simple, but it sits at the centre of a genuinely consequential problem for students: a tool that gets the answer wrong can damage an honest writer's academic record. This guide explains the statistical logic behind detection, the conditions under which it fails, and the practical steps any student can take to document their original work and protect themselves if a false accusation arises.

How AI detectors actually measure text

Every mainstream AI detector is, at its core, a statistical model. It does not read for meaning; it reads for probability. Two measurements dominate the field.

The first is perplexity, a technical term borrowed from information theory that measures how surprised a language model is by a given sequence of words. When a large language model generates text, it selects high-probability continuations at each step, so the resulting output tends to be low-perplexity: the words follow each other in ways a language model would have predicted. Human writing, by contrast, is often surprising. A novelist reaches for an unexpected verb; a student makes an idiosyncratic structural choice. These decisions register as high perplexity and push a piece toward the "human" end of the scale.

The second is burstiness, which describes the variation in sentence length across a passage. Human writers naturally alternate between long, complex sentences and short ones. Generated text tends toward uniform length and syntactic complexity throughout a paragraph. A detector that sees low burstiness alongside low perplexity has stronger statistical grounds for flagging a piece as machine-generated.

Most detectors also train on corpora of known human and AI text, learning the distributional fingerprints of each category. The output is usually a percentage score or a categorical verdict. Neither is a certainty; both are probabilistic estimates, and that distinction matters enormously in an academic-integrity context.

Where detection breaks down: the false-positive problem

The core weakness of the perplexity-and-burstiness approach is that it measures statistical style, not authorial intent. Any writing style that happens to produce low perplexity and low burstiness will score as suspicious, regardless of whether a human or a machine produced it. Several common writing situations produce exactly those patterns.

Formal academic prose. The conventions of academic writing, hedged argumentation, consistent register, disciplinary vocabulary repeated in disciplined ways, systematically flatten the stylistic variation that detectors read as human. A student who has learned to write well within genre conventions may write text that looks, statistically, more like generated output than a casual email would.

Non-native English. Writers working in a second or third language often rely on high-frequency vocabulary and syntactically safer constructions. The resulting text can score as low-perplexity for the same reason generated text does: it favours predictable paths through the probability space of English. Several independent audits of detector performance have found substantially elevated false-positive rates for non-native speakers, a finding with serious equity implications.

Technical and scientific writing. A chemistry lab report or a close reading of a legal text necessarily repeats specialist terminology. Detectors trained primarily on general prose can misread that repetition as the kind of lexical conservatism that characterises generated output.

Highly edited or polished drafts. Revision tends to smooth out eccentricity. A student who has carefully revised their prose for clarity and concision may inadvertently remove the burstiness that signals human authorship.

For a fuller treatment of the accuracy data behind these failure modes, see our companion piece on how accurate AI detectors are.

Do AI detectors work for students in practice?

The honest answer is: inconsistently, and in ways that skew toward false accusation of certain student populations. Independent benchmarks have found that detectors correctly identify AI-generated text at rates that sound impressive in marketing material but carry significant error rates when applied to real student work. A tool that is 90 percent accurate on balanced test sets can still produce a false positive for one in ten human-written submissions under realistic conditions, which is not a tolerable threshold when the consequence is an academic misconduct charge.

Instructors and institutions are increasingly aware of this limitation. Most educational guidance now frames detector output as one data point among many, not as definitive evidence. A flagged score is a reason to ask questions; it is not a finding of guilt. Students should understand that distinction, because it changes how they need to respond if their work is questioned.

You can run your own work through our AI detector tool to see how a piece scores before you submit it, not to game the result, but to understand what a reader of that score will see and to prepare your documentation accordingly.

Academic integrity and the right framing

Detector accuracy is a technical question, but the underlying issue is an ethical one. Academic integrity means submitting work that honestly represents what you understand and what you can do. The concern about AI-generated submissions is legitimate: an essay written by a language model tells an instructor nothing about whether the student has read the text, developed an argument, or learned the discipline's methods of reasoning.

That concern, however, should not translate into treating detector scores as proof. The two questions, did this student submit work that is genuinely theirs, and does this text score as AI-generated on a statistical model, are related but not identical. A student can submit genuinely original work that a detector flags. A student can also submit AI-generated work that a detector misses. Neither error is acceptable as the basis for an academic judgment, which is why process documentation has become more important than ever.

The broader landscape of AI in academic writing is still developing. Our AI and writing guide covers the range of questions students are navigating, from citation practices to permitted uses of AI assistance at different institutions.

How to protect your original writing

The most reliable protection against a false-positive accusation is a paper trail that predates the submission. The following practices do not change your writing; they document it.

Keep every draft. Version history in a word processor, timestamped saves, or a series of named files (draft-1, draft-2, and so on) all establish that a piece developed over time. AI generation does not produce drafts; human writing does.

Save your research trail. Browser histories, annotated PDFs, library search records, and notes on sources show the intellectual work that preceded the writing. An essay that grew out of visible research is harder to mistake for generated output.

Preserve feedback exchanges. If a tutor, writing centre consultant, or peer reviewer commented on your work, save those exchanges. Instructor feedback embedded in a draft is particularly strong evidence of a genuine writing process.

Write with specificity. Generated text tends toward the general because it optimises for plausibility across a wide range of readers. Analytical writing that makes precise claims, cites specific passages, and builds arguments from textual detail is both stronger academically and harder to mistake for output that was generated without genuine engagement with a source.

Vary your sentence architecture. This is a craft recommendation independent of detection concerns: mixing sentence lengths and structures produces more readable prose. It also raises burstiness scores and reduces the statistical resemblance to generated text. The two goals align.

If your work is flagged despite these precautions, present your documentation calmly and completely. A writing process that is visible and dateable is the strongest counter-argument to a probabilistic score. Detectors measure statistics; they cannot see your notes, your drafts, or the two hours you spent rereading a chapter before you wrote a word.

What detectors cannot measure

It is worth being precise about what falls outside the scope of any statistical detector. Detectors cannot identify whether a student understood what they wrote. They cannot assess whether an argument is original or derivative. They cannot detect paraphrasing from a source, selective quotation used misleadingly, or unacknowledged ideas, all of which are integrity concerns that predate large language models entirely. The problem of academic dishonesty is wider than AI generation, and detectors address only a narrow slice of it.

Conversely, detectors cannot confirm that a piece is genuinely the student's own work simply because it scores as human. A low AI-probability score is reassuring to no one; it is the absence of one kind of signal, not the presence of evidence for authentic authorship. Authentic authorship is demonstrated through what a student knows and can discuss, through the specificity and consistency of their written voice across a course, and through the documented process behind a finished piece.

Understanding these limits helps students engage with the technology honestly: not as a hurdle to clear, but as an imperfect instrument operating in a context where honest work, carefully documented, remains the soundest position.

Frequently Asked Questions

Do AI detectors work accurately enough to be used as proof of cheating?

No published detector achieves accuracy high enough to serve as standalone proof. False-positive rates in independent studies regularly reach 10 percent or higher on human-written text, which means a significant share of flagged submissions are entirely original. Most educational guidance treats a detector result as a prompt for conversation, not a verdict.

Why does my own writing get flagged as AI-generated?

Detectors flag text that scores low on perplexity, meaning it uses predictable word choices, and low on burstiness, meaning sentence lengths stay uniform. Academic writing styles, non-native English, technical vocabulary, and formal registers all tend to produce those same statistical patterns, which is why original student work is misidentified more often than most people expect.

What can I do if my original work is flagged?

Keep your drafts, notes, search histories, and any feedback exchanges that document the writing process. A portfolio of evidence showing how your essay developed over time is far stronger than any counter-argument made after the fact. Submit that documentation to your instructor alongside a clear explanation of your process.

Does writing style affect how detectors score a piece?

Yes. Short declarative sentences, technical terminology repeated across a paragraph, and restrained hedging language all resemble patterns that detectors associate with generated text. Varying sentence structure, grounding claims in specific textual evidence, and writing with genuine analytical unpredictability all reduce the statistical similarity to generated output.

Sources

No external sources were cited in this article.