Was This Text Written by AI?

How to use the AI text detector, step by step

Paste the text, run the analysis and read the highlighted passages before the overall verdict. The passages are where the useful information is.

1. 01Paste at least 300 characters of English, French or Arabic text. Longer extracts give more passages to score.
2. 02Leave out material that is not the author’s prose, such as reference lists, tables or code, when you can.
3. 03Keep passages whole. Trimming the text down to the part you suspect removes the context around it.
4. 04Read the verdict: probably human, uncertain, probably AI or mixed.
5. 05Open the highlighted passages and read the explanation, then download the PDF report if you need a record.

A text check uses 1 credit. To check a whole file with its images and metadata, use the AI document detector instead.

How does an AI writing detector measure a text?

It measures how predictable each word choice is to a language model. Generated text tends to follow the choices a model would make itself; human writing wanders more.

AnyFact uses the Fast-DetectGPT method with an open-source language model (Qwen2.5). For each passage of 2–3 sentences, it compares the words that were actually written with the alternatives the model considers likely. Passages that match the model’s own preferences very closely score toward AI.

When a score sits in the grey zone, Claude reviews the passage as a second opinion. It can move the score only a little and never decides alone.

How to read an AI text checker result

Look at where the highlighted passages fall, not just at the overall level. The pattern often tells you more than the verdict.

A mixed verdict is common for texts written with some assistance. It describes the text, not the author’s intent, so it calls for questions rather than conclusions.

How to check if text is AI-generated without a tool

Read for content, not just style. These checks work on any text, and they are worth doing whatever a detector says.

• Sources: look up the references. Citations that cannot be found are a strong reason to ask questions.
• Specifics: generated text often stays general where an informed author would give names, figures or examples.
• Consistency with the author: compare with their earlier writing, emails or work done in person.
• Leftovers: phrases addressed to the reader as a chatbot would, or a sentence that answers a question nobody asked.
• Confident errors: plausible-sounding facts that turn out to be wrong.

Every one of these can also appear in human writing. Our guide on how to tell if text is AI-generated explains how to combine them without jumping to conclusions.

Which texts are hardest to detect as AI?

Texts that are predictable by nature are the hardest, because human and generated versions look alike. Read results on these genres with extra care:

• Formulaic writing: legal boilerplate, abstracts, product descriptions, standard letters.
• Machine-translated text, which is also produced by a model.
• Lists, templates and texts close to the 300-character minimum.
• Texts full of quotations or set phrases that many authors share.

Cover letters are a common example: the genre itself is formulaic. Our page on hiring and recruiting suggests questions that reveal more than a score.