Education is one field that could benefit from OpenAI's new classifier by automating the detection of AI-generated essays and homework. It might allow a different response than say, restricting its use as New York Schools did recently. School departments across Australia have banned ChatGPT too.
The classifier could also help companies like developer Q&A site, Stack Overflow, which banned ChatGPT answers after its moderators were flooded with sometimes-correct solutions.
While the classifier can eliminate some detection work, OpenAI says it's impossible to reliably detect all AI-written text and there are a number of limitations that impact its effectiveness.
The classifier correctly identifies 26% of AI-written text as "likely AI-written" -- its true positive rate. It incorrectly identifies human-written text as AI-written 9% of the time -- its false positive. In other words, there is a reasonably high chance it will fail to detect text submitted by a human who doesn't disclose it was written by AI and some chance it will mislabel text both ways. But it claims the classifier is "significantly more reliable on text from more recent AI systems" than its previous detector based on GPT-2.
"We're making this classifier publicly available to get feedback on whether imperfect tools like this one are useful," OpenAI said in a press release, adding that it hopes the tool will "spark discussions on AI literacy."
The free web classifier tool is available here. To use it, the users needs to copy and paste a portion of text into the text box and then press submit. It will rate the text as 'very unlikely', 'unlikely', 'unclear if it is', 'possibly', or 'likely' AI-generated.
The classifier needs at least 1,000 characters -- about 150 to 250 words -- to analyze a text. OpenAI warns it's easy to edit AI-generated text in order to avoid detection programs. Also, the tool is more likely to mislabel text written by children and non-English text, because it was primarily trained on English content written by adults.
Also, OpenAI says it has "not thoroughly assessed the effectiveness of the classifier in detecting content written in collaboration with human authors."
It says the tool should "not used as a primary decision-making tool", but as complement to other methods of determining the source of a piece of text.