Open AI’s chatbot, ChatGPT, went viral shortly after it was launched by Open AI on November 30th. Many students have adopted it and rely on it for their essays, book summaries… to the great despair of teachers who cannot prove fraud. The start-up Draft & Goal comes to their rescue: it offers them a ChatGPT – GPT3 content detector that is 93% reliable.
DnG.AI or Draft & Goal, a Montreal startup co-founded by Nabil Tayeb and Vincent Terrasi, is developing an automated AI content workflow solution for its clients. She defends “an ethical AI that is not used in a deceptive way” and has trained a detection algorithm that helps teachers spot fraud, but it can also be useful for fighting misinformation and fake news, among other things.
How does the content detector work?
Operating the ChatGPT-GPT3 Content Detector is very simple. Just enter a text of at least 400 characters and click on analyze. The detector quickly determines whether the text was written in whole or in part by an AI and what percentage.
At the moment it only works with texts in English, but a French version is planned very soon.
Draft & Goal analyzed its performance on 2,000 texts and the reliability was 93%, the start-up keeps training its algorithm to improve the results.
Detect AI footprints
Teachers typically use software to detect plagiarism in texts published on the Internet. Models like chatGPT synthesize the information they contain to generate new texts, teachers, even when in doubt, cannot confirm that students have used them.
According to Vincent Terrasi, ChatGPT tends to do a lot of repetition and exaggeration, inventing, often using “hollow” words like the verb faire, so many clues for the detector. However, he acknowledges that the model’s task is much more difficult if the student has added a few personal details or slightly modified the generated text. This one is completely free at the moment and accessible to everyone.
Display AI-generated content
OpenAI is aware of the problem and is considering how to set up an imperceptible signature in the texts generated by GPT in order to be able to identify them if necessary.
Other AI detectors have been developed, such as Neuralext’s GPT Radar or GPTZero, developed by Edward Tian, a computer science and journalism student at Princeton University (USA), during the last Christmas holidays to detect student fraud.