OpenAI has released a new detection tool designed to identify text generated by its own language models. The system, which the company is calling a classifier, is meant to help educators, journalists, and platform moderators distinguish between human and AI writing. This move comes as concerns about automated misinformation and academic dishonesty continue to grow.
How the detection system works
The classifier works by analyzing patterns in text that are common to AI generation. According to OpenAI, the tool looks for small statistical markers that distinguish synthetic content from organic human writing. These include choices in word frequency, sentence length variation, and the predictability of certain phrases. The system has been trained on a wide set of texts from OpenAI’s own models, including GPT-3 and GPT-3.5, and it attempts to flag content that likely came from those systems.
In testing, the classifier showed reasonable accuracy for longer passages but struggled with shorter texts. OpenAI admits the tool is not perfect. It mislabels human-written text as AI-generated about 9 percent of the time. For text written by an AI, it correctly identifies it less than 30 percent of the time in ideal conditions. Those numbers drop further for text under one thousand characters. The company is clear that this tool should not be used as the sole basis for any decision or accusation.
Pushing back against AI misuse
The launch of this tool is part of a broader push by OpenAI to address misuse of its technology. Over the past year, AI generated essays, fake reviews, and automated propaganda have sparked debates in schools and online platforms. Some universities have banned ChatGPT outright, while others are trying to adapt their policies. By releasing a detection classifier, OpenAI is offering a technical solution to what has become a social problem.
The company has also acknowledged the cat and mouse nature of this work. As detection tools improve, AI models can be tweaked to evade them. OpenAI says it is committed to continuing research in this area and will update the classifier as its models evolve. The tool is currently available as a web app for free, though users must create an account to access it.
Critics point out that the classifier is only trained on OpenAI’s own models, meaning it may not recognize text from other AI systems like Anthropic’s Claude or Google’s Bard. That limitation makes it less useful for detecting synthetic content at scale. Still, the company argues that visibility into its own models is a necessary first step toward broader accountability across the industry.
Educators have responded with cautious optimism. Several teachers told tech outlets that they see the tool as a helpful reference but not a definitive answer. Many are calling for clearer academic policies around AI usage rather than reliance on detection software alone. Others worry that the tool will push students to use more advanced or obscure AI models in order to avoid detection.
OpenAI has also released a separate watermarking proposal for AI generated text, though it has not been implemented yet. That system would embed invisible markers in the text at the time of creation, making it easier to verify later. The company says it is still testing the watermark method and is seeking public input before rolling it out broadly.
The launch order of these tools shows how quickly the conversation around AI writing is evolving. A year ago, the biggest question was whether AI could write a coherent paragraph. Today, the question is how to manage the flood of AI content that is already being published online. Detection tools and watermarking systems offer partial answers, but they do not solve the deeper issue of trust in digital text.
As more companies release their own AI writing models, the industry will need collaboration on standards for provenance and disclosure. OpenAI’s classifier is a starting point. It puts a tool in the hands of users who want more clarity, but it also exposes how difficult the problem really is. For now, the best defense against AI generated text may still be careful reading, context, and common sense.
For more on how AI tools are reshaping digital content and the latest from the artificial intelligence landscape, visit Mylistingo.
The Netherlands, for internationals
One email a week: the week's news for internationals in the Netherlands, new practical guides and what changed in the cities. No spam, unsubscribe any time.
