How Do AI Detectors Work? A Plain-English Breakdown

How do AI detectors check for AI?

They read the words you paste. They do not see the prompt, the account, or the time of night. A checker splits the draft into sentences or short windows, scores each window, and turns those scores into a share of the page.

That share is the percentage. It is not a probability that someone cheated. Two tools can disagree because they were trained on different examples and they draw the line in different places. OpenAI shut down its own AI classifier in 2023 after it proved too weak to use as a verdict.

DetectGPT does the public version of this job. Paste text or a public URL, up to 5,000 words, no account. You get a percentage and highlighted lines. The AI checker is that box. It is not Turnitin, and it is not a watermark reader.

Perplexity and burstiness explained

Perplexity is a fancy word for surprise. A language model assigns odds to the next word. If the next word is always one of the obvious choices, perplexity stays low. Machine prose often looks unsurprising to another model, because both were trained to pick safe continuations.

Burstiness is the jumpiness of the writing. People change speed. A short sentence lands after a long one. A specific noun shows up after a string of abstractions. Model prose hums. The sentences match each other in length and in caution.

Those two pictures are still useful. They are not the number on a 2026 dashboard. Products moved to classifiers and then described the result as a percentage. How DetectGPT scores a draft is that report on this site. This section is the older vocabulary, which still explains why a flat page looks machine-made.

Classifier models: trained on human vs AI text

A classifier is a model that has seen many human documents and many generated documents and has learned a boundary. You give it a new passage. It says which side the passage is closer to, often as a score rather than a yes or no.

The boundary is only as good as the examples. If the human examples are news articles and the new text is a methods section, the boundary gets confused. If the generated examples are raw ChatGPT and the new text was edited for an hour, the boundary gets confused the other way.

There is no shared boundary. GPTZero, ZeroGPT, campus tools, and DetectGPT each train their own. A 20% on one is not a 20% on another. How to read an AI writing score is how to treat the number you actually received.

Watermarking, SynthID, and OpenAI's approach

A watermark is a signal added when the text is generated. The reader cannot see it. A detector that knows the scheme can sometimes still see it. That is a different test from perplexity. Style detectors work on any prose. Watermark detectors work only on text from a lab that planted the mark, and only if the mark survived editing.

Google DeepMind's SynthID is the public name for that family of marks across images, audio, text, and video. The text and video note describes watermarking the generation process itself. Developer docs are for people integrating the tool, not for a student pasting an essay into a website.

OpenAI has described its own text watermark and a restricted detector. What is the ChatGPT text watermark covers what they published, including how synonym edits weaken the mark. DetectGPT does not read SynthID or OpenAI's mark. If you need the style question instead, the signs are listed in signs of AI writing.

Can AI detectors detect Claude, Gemini, and GPT-5?

They can score prose that came from those models. They do not receive a label that says Claude or GPT-5. The habits overlap: even sentences, safe wording, a close that could fit any topic. The defaults differ a little, which is why a careful reader sometimes guesses the model and is still wrong.

A newer model does not reset the problem. If the prose is still the likely next word, a classifier trained on that family can still notice. If the prose was heavily revised, the notice gets weaker. DetectGPT is aimed at the finished draft from ChatGPT, Claude, Grok, Gemini, and other models. It is not a test that names the product.

What is the 30% rule of AI?

There isn't one. People repeat a 30% rule as if a course, a company, or a detector agreed on it. They did not. Some instructors treat a low band as noise. Some treat any highlight as a conversation. DetectGPT's own bands are rough: under about 20% is a weak signal, the middle is mixed, and over 50% means a real share of the draft still looks generated.

Do not bargain a policy down to a number you found in a thread. Ask what the assignment allows. Then read the lines. A 28% with three red paragraphs is a different paper from a 28% that is spread as a faint tint.

Why AI detectors produce false positives

False positives are human drafts that look generated. The usual reasons are tidy grammar, a second language that sticks to safe words, a template the course required, and technical prose that has no room to be surprising. Are AI detectors accurate stays on that failure.

False negatives are the other miss. A generated draft that was rewritten in a specific voice, with uneven sentences and local details, can fall under the line. Paraphrase tools are a weak version of that edit. They change words and keep the skeleton.

OpenAI's retired classifier is the clean public example. On their challenge set it caught only a minority of AI-written text and still mislabeled some human text. That is why a percentage is a map, not a finding.

How DetectGPT analyzes text

DetectGPT is a classifier-style check on the page you give it. It does not claim a private watermark, a campus archive, or a single number that other vendors must match. You paste 50 to 5,000 words, or you point it at a public URL. The report is a percentage plus yellow, orange, and red on the lines that carried the score.

Use the AI checker when you want that map. Use the plagiarism tab when you want overlap with the public web. Those are different questions. How DetectGPT scores a draft is the click-by-click version.

FAQs

How do AI detectors work in one sentence?

They score how much the finished prose resembles generated writing, using predictability, rhythm, or a classifier, and they do not open your chat.

What is perplexity?

A measure of how surprising the next word is to a language model. Low surprise is common in model prose. It is a picture, not the percentage on every modern dashboard.

Do detectors read SynthID or ChatGPT watermarks?

Only a detector built for that mark can try. DetectGPT does not. It scores style. A watermark check and a style check answer different questions.

Is 30% the cutoff?

No. There is no shared 30% rule. Read the lines and the policy you actually have.

Why do two detectors disagree?

They were trained separately and they threshold separately. Disagreement is normal. Edit the stiff lines instead of shopping for a kinder badge.