Does ChatGPT Detect AI Writing?

ChatGPT is the model most people picture whenever "AI detection" comes up — it's the one named in nearly every detector's marketing and most schools' AI policies. So it's a reasonable guess that ChatGPT itself can tell you whether a piece of writing is AI-generated. It can't, and OpenAI's own attempt to give it that ability is a useful, concrete case study. Here's the honest picture, from both directions.

ChatGPT isn't a detection tool

ChatGPT's job is generating text and holding a conversation, not scanning a document and returning an AI-generated probability. There's no "detect AI" button anywhere in the product, and nothing about how the model works gives it a hidden, reliable ability to recognize its own output or anyone else's. If you've heard someone say "I asked ChatGPT and it said this was AI-written," that's not a result from a calibrated tool — more on exactly why below.

OpenAI actually built one — and shut it down for low accuracy

Unlike most of the chatbots covered on this blog, OpenAI didn't just skip building a detector — it tried, and the attempt is the most concrete data point in this entire comparison. On January 31, 2023, OpenAI launched a standalone tool called the AI Text Classifier, designed to estimate whether a passage of at least 1,000 characters was AI-generated. In OpenAI's own published evaluation, it correctly labeled just 26% of AI-written text as "likely AI-generated," while incorrectly flagging 9% of genuinely human-written text as AI. OpenAI kept it live for under six months: on July 20, 2023, it quietly pulled the classifier, adding a note to the original announcement rather than issuing a new one, and cited its low rate of accuracy as the reason. If the company that built ChatGPT couldn't make a reliable detector for its own model's writing, that's a strong signal about how hard this problem actually is — for every detector, not just OpenAI's.

Asking ChatGPT directly isn't detection either

With the dedicated classifier gone, some people just paste text into ChatGPT and ask whether it sounds AI-written. That isn't a workaround — it's the same company's model doing the same job, minus even the narrow calibration the retired classifier had. A chatbot answering "does this sound AI-written?" is pattern-matching on surface style and producing a plausible-sounding guess, not running the statistical analysis a purpose-built detector like Turnitin or GPTZero performs. Reword the question and you can get a different answer, with no accuracy benchmark behind any of it. Treat "I asked ChatGPT" the way you'd treat an unverified guess — not evidence.

So does ChatGPT's own output get flagged?

Yes, potentially — and for ChatGPT specifically, there's an extra wrinkle. Real detectors like Turnitin, GPTZero, Copyleaks and Originality.ai don't look for a signature tied to one chatbot; they score general statistical patterns — predictable word choice, uniform sentence rhythm — the way we cover in how AI detectors work. Because ChatGPT was the model that triggered this entire category of tool and remains the most-cited example in most detectors' own documentation and training data, its default writing style is, if anything, one of the easier patterns for an established detector to recognize — not a special case that slips through. Unedited ChatGPT text pasted straight into an assignment carries real detection risk; we cover the mechanics in detail in does Turnitin detect ChatGPT?, and the same general risk applies to Gemini, Claude and every other model on this blog.

What this means if you use ChatGPT for schoolwork

Read your institution's academic-integrity policy on AI first — the rules differ by school and assignment, and following them is your responsibility. Using ChatGPT to brainstorm, outline, or have a concept explained back to you carries no unusual detection risk, since the words you ultimately submit are your own. Accept a ChatGPT-drafted paragraph as-is, though, and you're submitting machine-written text with the same real risk of a flag as output from any other AI model covered here.

If ChatGPT-assisted writing gets flagged

A flag from any detector is a pattern-based estimate, not proof of misconduct. If you only used ChatGPT for outlining or explaining and wrote the actual draft yourself, your version history is stronger evidence than any single score. If your own honest writing tends to read as flat or overly uniform even without AI help, that's a writing-level issue worth addressing directly — see why writing gets flagged as AI.

The bottom line

ChatGPT can't reliably detect AI writing, and OpenAI's own attempt to build that feature failed clearly enough that the company shut it down within half a year. What ChatGPT can do is write text that, left unedited, reads as machine-generated to the detectors your school might actually run — the same risk as every other AI assistant on this blog, arguably more so, since it's the model most of those detectors were built around in the first place. Write the final version yourself, and where a draft needs smoothing into clearer, more natural English, Grade A Humanizer is built for exactly that.

Questions people ask