The AI Classifier for Indicating AI-Written Text is a tool capable of distinguishing between human-written and AI-generated textual content. OpenAI developed this classifier, which has been trained to recognize text generated from a variety of AI providers. While it's not foolproof in detecting all AI-written content, it can help curb the misuse of AI-produced texts, such as for automated misinformation campaigns, academic dishonesty, or presenting AI chatbots as humans.Despite its utility, the tool is not entirely reliable. For instance, it has limitations in identifying short texts and may incorrectly label human-written text as AI-generated. Moreover, the tool's effectiveness mitigates for texts in languages other than English, and it performs poorly on code-based texts. Additionally, it does not accurately identify texts that are very predictable. Efforts to deceive the classifier by editing AI-written text are also possible. And lastly, neural network-based classifiers like this one tend to be poorly calibrated outside of their training data, leading to inaccurate predictions.The classifier's training involves fine-tuning a language model on a dataset of AI-written and human-written text pairs on the same topics. The responses were generated from numerous different language models from various organizations, which were then divided into prompts and responses. For the web application, the tool's confidence threshold is adjusted to maintain a low false positive rate.The AI Classifier is open for public use, with OpenAI interested in feedback regarding the tool's usefulness and effectiveness. The team anticipates the tool's impact to extend to sectors like journalism, research, and education.
