Detect AI Generated Text

How AI text detection actually works

AI detectors do not read text the way a human reviewer does. They look at token-level statistics. The two terms that matter most are perplexity and burstiness. Perplexity measures how surprising the next word in a sentence is to a language model. Human writers reach for unusual phrasing, niche vocabulary, and the wrong-but-right word more often than a model trying to minimise loss, so human writing tends to score higher on perplexity. Burstiness measures sentence-to-sentence variation in length and structure. People write a long, winding sentence, then a short one. They throw in a fragment. Models, left to their defaults, produce sentences of remarkably similar length and cadence.

Detectors combine those signals with vocabulary frequency, punctuation patterns, and known stylistic fingerprints from specific models. The score that comes out is a probability, not a verdict. No detector is 100% accurate, and any vendor that claims otherwise is misleading you. The same sentence can flip from "AI" to "human" with a single rewrite, a translation pass, or a paraphrasing tool. Treat the output as a prompt for a second look, not as proof of anything.

When (and when not) to use an AI detector

Good use cases are editorial review (spot-checking a freelancer who claims human authorship), plagiarism investigation (paired with a plagiarism checker, not as a replacement), academic integrity reviews where a human educator makes the final call, and content audits where you want to see how much of your historical content was clearly drafted by a model. For SEO teams, this is the most common use: figuring out which pages on the site are thin AI output that needs a rewrite before the next core update.

The bad use cases are anything where the detector is the judge. AI detectors have well-documented false positive problems with non-native English speakers, with formal academic writing, and with anyone whose voice happens to land near a model's training distribution. Failing a student, firing a writer, or rejecting a submission on the basis of a detector score alone is a bad call. If you would not be comfortable explaining the decision in court or in a hiring tribunal, do not make the decision on a probability score. Use it as a signal, not a sentence.

Limitations of AI detection

Detectors struggle in four predictable places. First, paraphrased AI text: run a passage through a paraphrasing tool or a second model and most detectors lose the scent. Second, mixed human and AI content, where a writer used a model for the outline or a few paragraphs and edited everything else by hand. The signals get blurred and the score lands in the middle. Third, short snippets: anything under a few hundred words has too little statistical surface for a confident call. Fourth, fine-tuned model evasion. Anyone who deliberately fine-tunes or prompts a model to write in a more human, bursty, idiosyncratic style can dodge detection consistently. Detectors are useful, but they are not a lie detector and they are not future-proof.

How to read the source attribution scores

The tool returns independent likelihood scores for the major writing models, plus an Unknown / Mixed bucket. A high ChatGPT score means the text leans on phrasing, sentence shape, and punctuation patterns common to OpenAI's models (the famous love of em dashes, "in conclusion", and "it is important to note"). A high Claude score points to longer flowing sentences, hedged language, and the structural balance Anthropic's models tend to produce. A high Gemini score suggests the cleaner, more list-oriented Google style. A high Perplexity score is more about citation-heavy, summary-style structure. The Unknown / Mixed score going up means the text either does not match a single model cleanly, has been edited or paraphrased, or is a blend of multiple sources. Because each score is independent, they will not add up to 100. Use the strongest label as a starting hypothesis, not a confession.

What to do after running the detector

If a page scores high on AI likelihood and you want to keep it, the fix is not to run it through another model. The fix is to make it more useful. Edit for human voice (cut the throat-clearing intros, the "in today's fast-paced world" filler, the symmetrical three-bullet lists). Add first-hand experience: screenshots, numbers from your own work, anecdotes a model could not invent. Get a subject matter expert to review the page and add a quote or a correction. Tighten the title, the H2s, and the lead so the value is obvious in the first 50 words. Then recheck with the detector and read it out loud. If it still sounds like a model wrote it, the rewrite was not deep enough.