ChatGPT detectors such as GPTZero, Turnitin’s AI indicator, Copyleaks and Originality.ai can flag text that looks machine-generated, but they cannot prove who wrote something. They make mistakes in both directions: they miss edited AI text and sometimes flag human writing, especially from non-native English speakers. Use their score as a reason to ask questions, not as evidence.
How AI detectors work
Most detectors look at statistical features of the text. Two common signals are perplexity, how predictable each word is to a language model, and burstiness, how much sentence length and structure vary. AI text tends to be very predictable and even; human writing is usually less so. Some tools also use classifiers trained on examples of human and AI text. None of them can see how the text was actually written.
How accurate are they?
Accuracy depends heavily on the text. Detectors do best on long, unedited passages straight from a chatbot. They do worse on short texts, on AI text that has been paraphrased or edited by a person, and on writing in simple, formulaic styles. Vendors publish high accuracy figures from their own tests, but independent studies have repeatedly found much more variation.
Two data points are worth knowing. OpenAI withdrew its own AI classifier in 2023 because of its low rate of accuracy. And a 2023 Stanford study found that several popular detectors flagged a large share of essays written by non-native English speakers as AI-generated, because their simpler vocabulary looked predictable.
Why human writing gets flagged
Clear, simple and well-structured writing, the kind schools teach, can look “low perplexity” to a detector. So can technical writing, writing that follows a template, and text run through grammar tools. That is why a high AI score on its own is weak evidence.
What detectors miss
Editing breaks the statistical signals detectors rely on. A person who rewrites AI output, mixes it with their own sentences, or asks the model to vary its style can push scores down. Newer models also write less uniformly than older ones.
Advice for students
Know your school’s rules on AI use, which vary from course to course. Keep drafts, notes and version history, since those show your process if a score is questioned. If AI use is allowed, use it for brainstorming, explanation and feedback rather than submitting its text as yours. Our guide to AI for students covers ways to use AI that most schools accept.
Advice for teachers
Treat a detector score as a prompt for conversation, not a verdict. Compare with the student’s previous work, ask them to explain their argument, and check sources, since invented references are a stronger sign of unedited AI use than any score. Clear rules about which AI uses are allowed reduce disputes before they start.
Can AI text be watermarked instead?
Some providers are researching watermarks, hidden statistical patterns added during generation, and content credentials attached to images. Text watermarks are not yet standard across chatbots and can be weakened by rewriting, so they do not solve detection today.
For the background on how these models generate text, read what a large language model is.
Frequently asked questions
Can teachers tell if you used ChatGPT?
Detectors give a probability, not proof. Teachers are more likely to notice AI use from writing that does not match a student's usual work, invented sources, or answers that cannot be explained in conversation.
Is there a 100% accurate AI detector?
No. Every detector produces false positives and false negatives, and accuracy drops further on short texts, edited AI text and writing by non-native speakers.
Did OpenAI make a ChatGPT detector?
OpenAI released an AI text classifier in January 2023 and withdrew it in July 2023 because of its low accuracy.