OpenAI LLC today introduced a cloud-based tool that can help users determine whether a piece of text was written by artificial intelligence.
The tool, which is available at no charge, is itself based on an AI model. It can detect text generated by not only OpenAI’s neural networks but also competing software from other companies. However, the startup cautioned that the tool currently has a number of significant accuracy limitations.
“In our evaluations on a ‘challenge set’ of English texts, our classifier correctly identifies 26% of AI-written text (true positives) as ‘likely AI-written,’ while incorrectly labeling human-written text as AI-written 9% of the time (false positives),” OpenAI researchers detailed in a blog post today.
OpenAI says that the tool is most adept at analyzing English-language texts comprising more than 1,000 characters. It can’t reliably evaluate text snippets that are shorter or written in other languages. Moreover, the tool is not yet capable of accurately detecting AI-generated software code.
There are also other limitations to the tool’s accuracy. According to OpenAI, AI-generated text can be modified to make detection more difficult. The startup’s researchers added that “for inputs that are very different from text in our training set, the classifier is sometimes extremely confident in a wrong prediction.”
The training dataset that OpenAI used to develop the tool included a collection of text prompts submitted by users to its cloud-based AI models. According to the startup, its researchers entered each prompt into multiple neural networks to produce a series of machine-generated responses. The startup also used text from other sources during the development process.
OpenAI’s efforts to develop software for detecting AI-generated text date back several years. In 2020, the startup released GPT-2 Output Detector, a tool for spotting text generated by its GPT-2 neural network. The neural network is a predecessor of ChatGPT that was released in 2019 and offers a more limited feature set.
According to OpenAI, the new tool it released today is significantly more adept at detecting text generated by recently developed AI models.
“While it is impossible to reliably detect all AI-written text, we believe good classifiers can inform mitigations for false claims that AI-generated text was written by a human: for example, running automated misinformation campaigns, using AI tools for academic dishonesty, and positioning an AI chatbot as a human,” OpenAI’s researchers explained.
The release of the tool comes a week after the startup raised a multibillion-dollar investment from Microsoft Corp. to advance its AI research. The investment includes not only capital but also cloud computing infrastructure. Previously, OpenAI raised $1 billion from Microsoft in 2019 and partnered with the company to build a cloud-based supercomputer for its researchers.