ChatGPT maker OpenAI has released a “classifier” tool that can detect text written by AI, but notes that it should not be relied upon.
The breadth of prompts that ChatGPT could create answers startled educators as some students began turning in AI-generated essays and homework assignments as their own.
Also: What is ChatGPT and why is it important?here’s everything you need to know
Education is one field that could benefit from OpenAI’s new classifier, which automatically detects AI-generated essays and homework assignments. It might allow a different response than say, limiting its use, as the New York school did recently. School departments across Australia have also banned the use of ChatGPT.
The classifier could also help companies like developer Q&A site Stack Overflow, which banned ChatGPT answers after its moderators were inundated with sometimes correct solutions.
While classifiers 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 affect its effectiveness.
The classifier correctly identified 26 percent of the AI-written text as “probably AI-written”—its true positive rate. It incorrectly identified human-written text as AI-written 9 percent of the time—false positives. In other words, there’s a good chance it won’t detect text submitted by someone who didn’t reveal it was written by AI, and it’s likely that it will mislabel the text in both cases. But it claims that the classifier is “more reliable on text from state-of-the-art AI systems” than previous GPT-2-based detectors.
“We are making the classifier publicly available to get feedback on whether such imperfect tools are useful,” OpenAI said in a press release, adding that it hopes the tool will “ignite conversations about AI literacy.”
A free web classifier tool is available here. To use it, the user needs to copy and paste a portion of text into the text box and press submit. It will rate the text as “very unlikely”, “unlikely”, “unclear”, “likely”, or “likely” to be generated by AI.
The classifier needs at least 1,000 characters (approximately 150 to 250 words) to analyze text. OpenAI warns that editing AI-generated text can easily evade detection programs. Also, the tool is more likely to mislabel text written by children and non-English text, since it was trained primarily on English-language content written by adults.
Additionally, OpenAI said it “did not thoroughly evaluate the effectiveness of classifiers at detecting content written in collaboration with human authors.”
It said the tool “should not be used as a primary decision-making tool” but rather as a complement to other methods of determining the origin of a piece of text.
