How to Tell if Text Is AI Generated

How can you tell if text is AI generated?

You tell if text is AI generated by combining three kinds of evidence: what the writing looks like, what a detector measures, and what you know about how it was produced. Learning how to tell if text is AI generated is mostly learning which of these to trust, and how much.

Reading alone is weak evidence, because good editors, templates and grammar tools produce the same polish. Process evidence, such as drafts and a writer who can explain their choices, is strong. A detector sits in between: useful for spotting which passages deserve a closer look.

What signals suggest something was written by AI?

The common signals of AI writing are generic content, uniform rhythm and confident but unverifiable claims. Each has an innocent explanation, which is why you should treat them as a reason to look closer, never as a conclusion.

How does an AI writing detector decide?

An AI writing detector mostly asks one question: how predictable is each word, given the words before it? Language models write by choosing likely next words, so their text tends to sit in the most probable path. People take more detours.

Researchers call this measure perplexity: low perplexity means the text is easy for a model to predict. Some detectors also look at how much predictability varies from sentence to sentence, since human writing tends to be burstier.

The AnyFact AI text detector uses the Fast-DetectGPT method with an open-source language model (Qwen2.5). In plain words, it compares each word the writer chose with the alternatives the model expected. Text that keeps landing on the model's favorite choices scores as more likely generated.

This also explains the weaknesses. Simple vocabulary, formulaic genres such as abstracts or cover letters, and very short passages are all predictable for human reasons.

Other tools work differently. Some train a classifier on large sets of human and generated text; some look for watermarks that a model provider deliberately hides in its word choices. Whatever the method, the output is a statistical estimate, and the same text can score differently in two tools. If you specifically suspect a chatbot, our ChatGPT detector page explains what that kind of check can and cannot tell you.

How do you detect AI writing in essays, emails and reviews?

You detect AI writing differently in each genre, because each one has its own normal style. Compare the text with what a real person would write in that situation, not with an abstract idea of “human” prose.

In every genre, the strongest clue is a mismatch between the text and the situation: the writer, the task, the audience or the timing. Style alone rarely carries a case.

What about texts that are part human, part AI?

Mixed texts are now the norm, so a good check looks at passages, not just the whole document. A student may write the argument and ask a chatbot for the introduction; a marketer may generate a draft and rewrite half of it.

A single score for the whole text averages these parts away. AnyFact scores each passage of 2–3 sentences, so a partly generated text comes back as mixed with the suspicious passages highlighted. That tells you where to focus your questions.

Read the pattern of highlights, not just their number. A flagged introduction and conclusion around a hand-written middle tells a different story from scattered single passages, which may simply be the most formulaic sentences in the piece. Ask about the highlighted parts specifically: where did this paragraph come from, and what did you change?

Heavy human editing of AI output, or AI polishing of human writing, both blur the signal. Expect more “uncertain” results on these texts, and read that as an honest answer rather than a failure.

Why do AI text checkers flag human writers?

AI text checkers flag human writers when their writing is predictable for innocent reasons. The groups most exposed are non-native writers, people writing in a rigid format, and anyone submitting a very short text.

• Non-native writers often use a smaller, safer vocabulary and textbook transitions, which lowers perplexity.
• Formulaic genres such as lab reports, legal summaries and product descriptions follow strict patterns.
• Grammar and rewriting tools smooth out the irregularities that make writing look human.
• Short texts give the detector too little evidence, so random variation dominates.

Published research has documented the first problem in detail; our guide on whether AI detectors are accurate covers it. If you work with students, the page for teachers and schools explains how to build these limits into your policy.

How to tell if text is AI generated, step by step

You check text in a fixed order so that your own first impression does not drive the conclusion. Write down what you see at each step before moving on.

1. 01Compare the text with the writer's earlier work: vocabulary, typical errors, sentence length, how they open and close.
2. 02Verify every checkable specific: quotes, references, dates, numbers and names.
3. 03Look for personal or local detail that only the writer would know, such as a class discussion or a project constraint.
4. 04Run the text through an AI text checker and note which passages it flags, not only the overall label.
5. 05Ask for the process evidence: drafts, notes, version history.
6. 06Talk to the writer before you draw any conclusion.

For essays, the AI essay checker applies the same method. For a full report or slide deck, the AI document detector checks text block by block and embedded images too, page by page.

What should you do with an AI detection result?

Use the result to start a conversation, never to end one. A detector result is an indication, not proof, and nobody should be sanctioned on a detector result alone.

Checklist: detect AI writing fairly

• Did I compare with known samples from the same writer?
• Did I verify the quotes, sources and figures?
• Is the text long enough, and in a language the checker supports?
• Did the checker flag specific passages, and do they match my own concerns?
• Could the writer be a non-native speaker, or have used a grammar tool?
• Have I seen the drafts or version history?
• Have I spoken to the writer before deciding anything?

If you have worked through every question and still have doubts, record the case as unresolved. An honest “we could not tell” is better than a wrong accusation, and it keeps the door open if new evidence appears.

Code raises the same questions with even weaker signals; see how to detect AI-generated code.

Frequently asked questions