Can ChatGPT Detect AI-Written Text? (OpenAI Couldn't)
A lot of people assume ChatGPT can simply tell whether a piece of text was written by an AI, after all, it is the AI. That assumption is wrong, and it is worth correcting because it leads students, teachers, and employers to trust a tool that does not do the job. Here is what OpenAI itself says, what happened to its own detector, and which third-party tools actually check AI text today.
ChatGPT has no built-in 'was this AI?' button
ChatGPT does not know whether text was written by a model. In its own educator FAQ, OpenAI states plainly that ChatGPT has no knowledge of what content could be AI-generated and that when you ask it 'did you write this?' it will sometimes make up responses that are random and have no basis in fact. In other words, a 'yes' from ChatGPT is a coin flip dressed up as an answer, not evidence.
That confuses a lot of people because the model sounds confident. Confidence is not accuracy. If you need to know whether a draft is machine-written, you need a purpose-built detector, not a chat window.
OpenAI's own detector was retired after six months
OpenAI did build a dedicated AI Text Classifier, launched in January 2023. By July 20, 2023 it was shut down, with the company's own words: no longer available due to its low rate of accuracy. In evaluation it caught only 26% of AI-written text (true positives) and falsely flagged 9% of human writing (false positives).
The striking part is that the company which trained the model could not reliably detect its own output. If the creator misses three out of four of its own documents, no one should expect a generic chat tool to do better. This is exactly why independent detectors such as Originality.ai exist and why they focus their accuracy work on the hard cases, like paraphrased text, where naive classifiers fall apart.
How real detectors actually work
Third-party detectors do not ask the model. They measure statistical properties of the text itself, mainly perplexity (how predictable each word is) and burstiness (variation in sentence length and structure). Human writing is higher-perplexity and more bursty; model output trends toward uniform predictability.
Copyleaks is a good example of a detector built to hold up across languages, keeping quality over 30-plus languages while layering plagiarism checking on top. The method matters more than the brand: any tool that only returns a single percentage without showing you why is hiding the weakness of its own signal.
The most-searched free option, with a catch
ZeroGPT is the casual first stop: paste text, get a percentage, no account needed, and it has grown to tens of millions of users. For a quick ballpark read it is fine.
The honest tradeoff is accuracy. Independent testing puts ZeroGPT's real-world performance well below its marketing, with false-positive rates measured between 15% and 33%, and it misflags a large share of non-native English writing. For any consequential decision about a student or a client's work, the more accurate paid tools are the safer call.
For schools already inside a system
Turnitin did not set out to detect AI, but because schools already pay for its plagiarism platform, its AI indicator arrived inside the workflow they use. No new tool to adopt, no separate login.
The serious concern is fairness. Independent testing shows false-positive rates as high as 50% on ESL writers, a real equity problem for campuses with international students. It remains institutional-only, with no individual pricing, so individuals still need a separate checker for a private self-test.
The workflow that actually works
If you want a trustworthy self-check before submitting or publishing, start with a free student-focused detector. Scribbr's detector is free, requires no account, explains its flags in plain language, and sits around 90% accurate, which is strong for a no-cost tool.
Then, if the stakes are high, confirm with a second, independent detector rather than trusting any single score. No detector is definitive, and the one built into a chat box is the least definitive of all.
Bottom line
No, ChatGPT cannot detect AI-written text, and OpenAI says so itself: its answers to 'did you write this?' are random and baseless. OpenAI's own dedicated classifier was shut down in July 2023 after catching only 26% of AI text and falsely flagging 9% of human writing, so even the model's creator could not reliably detect its own output. To actually check a draft, use a purpose-built detector such as GPTZero, Originality.ai, Copyleaks, ZeroGPT, Turnitin, or Scribbr, and confirm any high-stakes call with a second independent tool rather than trusting one score.