How to check whether text is AI-generated

There are four ways to check text for AI generation, and only one of them holds up on its own. Provider watermarks are the most reliable signal but you cannot run them; detector access is limited to approved organisations. Classifier detectors are free and instant but fallible, disagree with each other, and are wrong more often for non-native and plain-styled writers. Provenance metadata is excellent for images and rare for text. Process evidence, meaning version history, drafts, typing replay and the research trail, survives rewriting and does not depend on anyone's algorithm. Use classifiers to decide where to look, and process evidence to decide.

Method one: provider watermarks

If the text came from Gemini, a future Claude model, or ChatGPT in the EU, it probably carries a statistical mark. The mark is strong evidence when present and long enough, and the providers' detectors are built to report only whether their own model likely produced part of the text. The catch is access. Google, Anthropic and OpenAI each limit detection to regulators, researchers, media, educators and similar organisations, by application. A school district, a newsroom or a compliance team can apply; an individual cannot. If you qualify, this is the first check to make, and the marks section links to each programme.

Method two: classifier detectors

Classifiers such as Pangram, GPTZero, Originality.ai, Copyleaks and Turnitin work on any text from any model, which is their value. Their limits are well documented: accuracy varies by model, domain and length; paraphrasing lowers it; and the one peer-reviewed fairness study found a 61.3 percent average false positive rate on non-native English essays. Use them this way: run at least two, expect disagreement, ignore anything under a few hundred words, and treat a high score as a reason to ask for process evidence rather than as the evidence. For a reproducible second opinion that no vendor controls, Binoculars is open source. For a sense of how detectors compare on equal terms, the RAID leaderboard is the only neutral scoreboard.

Method three: provenance metadata

If the content is a file, especially an image, audio or video file, check it for Content Credentials with the free verify tool. A signed C2PA manifest records the originating tool and the edit chain. For text this rarely helps, because text is pasted rather than sent as a file, and the manifest stays behind. A PDF or a document file can carry one, so it is worth a look when a file is what you have.

Method four: process evidence

How was the text made? A document with version history shows the writing happening: the outline, the false starts, the paragraph that was deleted. Grammarly's authorship feature and GPTZero's writing replay record keystrokes. Drafts, notes, sources and browser history show the research. None of this changes when text is paraphrased, and an absence of history on a polished text is itself informative. In any setting where a decision about a person follows, this is the evidence to ask for first and weigh most. The response guide explains how to assemble it.

Signals that do not work

Typographic tells such as em dashes, certain vocabulary, or tidy structure are folklore. Humans use them and models can be told not to. Hidden characters and zero-width spaces are not how production watermarks work and are trivially stripped. A single detector score, a reader's hunch, and the fact that a text is fluent are not evidence of anything. The honest state of the art in late 2026 is layered: watermark if you have access, classifiers for triage, metadata for files, and process evidence for the decision.

Questions people ask

Is there a free tool to check if text is AI-generated?

Free classifiers exist (GPTZero, Pangram, Scribbr, Sapling and others). They are signals, not proof, and they disagree with each other.

Can I check a text for a ChatGPT or Claude watermark?

Not with a public tool. Detector access is limited to approved organisations at Google, Anthropic and OpenAI.

What is the most reliable way to tell if someone used AI?

Process evidence: document version history, drafts, typing replay and research trail. It is unaffected by rewriting and does not depend on any detector.