The maker of ChatGPT, OpenAI, has just released a “classification” tool capable of detecting texts written by AI. But he makes it clear that you shouldn’t rely on it too much.
The breadth of questions for which ChatGPT can create answers alarms the teaching community, as some students are beginning to turn in AI-generated assignments passing it off as their own work.
The education sector is therefore an area that could benefit from OpenAI’s new classifier, by automating the detection of AI-generated essays and assignments. It could allow a different response from that which consists in restricting or prohibiting its use, as recently done in France by Science Po, or even the schools of New York and Australia.
Unable to reliably detect all texts written by AI
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.
Although the classifier can eliminate some of the detection work, OpenAI says it is impossible to reliably detect all text written by the AI and a number of limitations impact its effectiveness.
The classifier correctly identifies 26% of texts written by the AI as “probably written by the AI” – its true positive rate. It also incorrectly identifies texts written by humans as being written by AI in 9% of cases (false positive rate). In other words, there is a high probability that it will not detect a text submitted by a human that did not reveal that it was written by the AI. And a certain probability that he is wrong in both cases. But the company says the classifier is “significantly more reliable on texts from newer AI systems” than its previous GPT-2-based detector.
250 words to do an analysis
“We are making this classifier available to the public to get feedback on the usefulness of flawed tools like this,” OpenAI said in a statement, adding that it hopes the tool “will spur knowledge discussions. of AI”.
The free web classification tool is available here. To use it, you have to copy and paste some of the text into the text box, then click on “Send”. The tool will classify the text as “very unlikely”, “unlikely”, “uncertainly”, “possibly” or “probably” generated by the AI.
The classifier needs at least 1,000 characters – or about 150-250 words – to analyze a text. OpenAI warns that it’s easy to edit AI-generated text to evade detection programs. Also, the tool is more likely to mislabel text written by children and non-English speaking text, as it was trained primarily on English content written by adults.
Furthermore, OpenAI claims that it has not “thoroughly evaluated the effectiveness of the classifier in detecting content written in collaboration with human authors.” The company clarifies that the tool should “not be used as a primary decision-making tool”, but as a complement to other methods for determining the source of a text.
Source: “ZDNet.com”