AI article detector: catching AI-generated content in essays and long-form text

An AI article detector is a tool designed to analyse long-form text - essays, articles, blog posts, reports - and flag content that appears to have been generated by an AI writing system rather than a human author. The technology has grown rapidly since ChatGPT made AI-generated text ubiquitous in early 2023, and the tools vary considerably in what they can realistically deliver.

How article-level detection differs from sentence-level

Some detection tools work at the sentence level, flagging individual sentences as AI-generated. Others work at the document level, producing an overall probability score for the full article or essay. The distinction matters because AI-generated content rarely appears in clean isolation - most cases of AI use in student essays involve a mixture of generated text, lightly edited generated text, and genuinely original writing.

Document-level scores are generally more reliable than sentence-level classifications. A tool that highlights individual sentences as 'likely AI-generated' is making very fine-grained probabilistic claims that carry significant uncertainty at that resolution. An overall document score is still uncertain, but the averaging effect makes it more robust than any single sentence attribution.

What the detection is actually doing

The underlying approach in most AI article detectors involves measuring text predictability. Language models generate text by predicting the most likely next token; the output tends to be uniformly smooth, with word choices that are statistically expected. Human writing - especially analytical, argumentative writing - tends to be less predictable: writers use idiosyncratic phrasing, make unexpected connections, and vary their stylistic register in ways that diverge from statistical expectation.

Perplexity and burstiness are the two most commonly cited metrics. Low perplexity and low burstiness together produce a high AI-detection score. The problem is that formal, careful writing also trends towards lower perplexity and lower burstiness - which is why the tools have documented false positive problems with ESL writers and formal academic prose.

Paragraph-level versus full-article detection

An AI paragraph detector works by applying the same analysis to individual paragraphs rather than the full document. This has appeal for articles where AI use was selective - perhaps an introduction generated by AI and a body written by the author - because a full-article score would average out to something that looks mostly human. Paragraph-level analysis surfaces the unevenness.

In practice, paragraph-level detection is even more uncertain than document-level. A single paragraph is too short for reliable statistical analysis; the confidence intervals are wide enough that the classification is close to noise. Where you see clearly AI-generated introductory paragraphs followed by clearly human-written analysis, the pattern may be visible to an experienced reader before it's visible to a detector.

The mixed-origin problem

The hardest case for any AI article detector is mixed-origin content: an article that started as AI-generated text but was substantially edited by a human, or one where specific sections were generated while others were written independently. Most tools are trained and evaluated on cleanly generated or cleanly human-written content; the mixed case is less well-covered, and detection rates fall accordingly.

For teachers, this means the most concerning scenario - a student who used AI as a starting point and edited the output - is also the scenario where detection tools are least reliable. A student who pasted in a generated essay and changed a few phrases might still produce a high detection score; one who generated a structural outline and then wrote to it independently might score very low.

What to do with a detection result

A high AI article detection score is a starting point for further investigation, not a finding in itself. It means: look more carefully at this submission and have a conversation with the student about their process. Ask them to walk you through how they approached the essay, which sources they drew on, what changed between their first and second draft.

If you have process data alongside the detection result - from a tool that captured the writing session - the combination is considerably more useful. A high AI detection score plus a submission that arrived as a single large paste event provides two independent signals pointing in the same direction. That combination is meaningfully stronger than either signal alone.