Every week, I get a dozen emails from startups that sell “AI detection” services. These small companies promise to identify when an essay or news article is written by ChatGPT or other language models. And, startlingly, they often claim that their services are up to 90% accurate.
It’s very difficult to trust such claims. For one thing, ChatGPT isalways changingin small ways, and it may provide wildly different responses for a single prompt. ChatGPT also lacks any built-in markers or patterns to aid with recognition.
But, more notably, OpenAI says that it can only identify AI writing with about 26% percent accuracy. And, as you may know, OpenAI is the creator of ChatGPT. It should be the best at identifying AI writing, yet its detection tools are akin to guesswork.
On June 20th, OpenAI quietly shut down its detection tool due to the “low rate of accuracy.” The company hopes to offer more accurate detection tools in the future, but for now, it’s back at the drawing board.
As of July 25, 2025, the AI classifier is no longer available due to its low rate of accuracy. We are working to incorporate feedback and are currently researching more effective provenance techniques for text, and have made a commitment to develop and deploy mechanisms that enable users to understand if audio or visual content is AI-generated.
Some third-party AI detection tools may be more accurate than OpenAI’s “classifier.” But until such tools are genuinely 90% accurate, they may do more harm than good due to false positives.
Related:No, You Can’t Write an Actual Novel With OpenAI’s ChatGPT
For example, Twitter usersrecently learnedthat GPTZero, a popular AI detection tool, will flag the U.S. Constitution as “likely to be written by an AI.” Imagine if a university used this tool on its students, or if a business used it on its employees. How many people would be unduly punished?
The good news is that we can still use our brains to “detect” AI-written content. This is especially true in classrooms, businesses, and other settings where you may become familiar with a person’s writing (and, more importantly, their knowledge). It’s not all that different from “detecting” plagiarism—even if a teacher is unsure what materials a student copied, they may intuitively know that the student has cheated.
Still, we can’t verifiably detect AI-generated content. And that’s somewhat worrying. It seems that, for the time being, we’re stuck with our intuition.
