OpenAI Launches Tool to Detect AI-generated Text
- By DSAITrends editors
- February 08, 2023
OpenAI has released an AI Text Classifier tool to identify text generated by its AI, as part of belated efforts to curb potential abuses of ChatGPT. In a blog post, it also issued a call for input from educators and others impacted by ChatGPT.
“We’ve trained a classifier to distinguish between text written by a human and text written by AIs from a variety of providers… 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.”
Not fully reliable
OpenAI warns that its free classifier is work-in-progress and “not fully reliable”. In its evaluations using a set of English texts, it says the classifier correctly identified just one quarter (26%) of AI-written text. In one in 10 cases (9%), the tool also incorrectly identified human-written text as AI-written.
It is also unreliable on short texts of less than 1,000 characters and is not recommended on non-English text or programming code. Crucially, the tool can be extremely confident yet wrong on non-ChatGPT models.
However, authors Jan Hendrik Kirchner, Lama Ahmad, Scott Aaronson, and Jan Leike noted that the AI Text Classifier is “significantly more reliable” on the text generated by ChatGPT than the earlier GPT-2 Output Detector.
While many have lauded its ability to help generate ideas, write marketing copy, and write software, ChatGPT poses a unique headache for educators. On that front, the race to detect AI-assisted plagiarism is on, with solutions such as GPTZero cropping up. GPTZero relies on two different metrics to assess if a piece of content was written by AI, evaluating the randomness in a sentence (perplexity), and randomness for all the sentences in a text (burstiness).
Popular plagiarism detection service Turnitin is also developing a system to help educators determine if a student assignment was written by AI models such as ChatGPT. Currently in closed testing, the AI detection system works by differentiating the idiosyncratic style of human writing from how AI models are more likely to select the most commonly used words in a particular context.
OpenAI is also working on technology that relies on a cryptographic function to output text that incorporates a “watermark” to identify AI-generated content. A prototype is available, and it is understood that deleting random words or rearranging sentences will not thwart the detection system.
The release of the AI Text Classifier comes at a time when the European Union (EU) is developing regulations for AI, which is increasingly incorporated into smartphones, autonomous vehicles, and manufacturing systems. While AI could offer opportunities, there are also risks involved, according to EU Commissioner for Internal Market Thierry Breton.
“This is why we need a solid regulatory framework, to ensure trustworthy AI based on high-quality data,” said Breton.
AI Text Classifier can be accessed here.
Image credit: iStockphoto/ThitareeSarmkasat