In the midst of the boom of ChatGPT, the chatbot created by OpenAI that keeps growing at a dizzying pace with new uses, a controversy related to its use has started. The problem comes from the fact that many students are using this technology to do homework, cheat on exams or facilitate their homework. This is causing despair among teachers, as there is no real way to discern between real content and ChatGPT copied content.
But OpenAI is not willing to let the fame of its application go down because of this phenomenon, so it has put its foot down. According to the company’s announcement, it has created a tool with which to discern whether a text is created by Artificial Intelligence or is original. This will work mainly on ChatGPT texts, but will also be applicable to other chatbots.
Sam Altman, the CEO of OpenAI, believes that this tool can be a breakthrough for teachers. Instead of stepping aside or not engaging with the issue, he has decided to help teachers and launch this new application.
As they explain, the system is still in its infancy, but it will improve over the months. Currently, in the app’s native language (English), the new tool identifies 26% of all texts generated by AI. And, so far, it is wrong in 9% of the cases when it comes to original human text.
To test it, just go to this link, paste the text you want to check and click on the “Submit” button. As we said, it is only available in English, but it will be available in other languages (including Spanish) in the near future.
As the tool is not yet infallible, the results that appear are “very unlikely”, “unlikely”, “unclear” or “possible” when reporting whether the text has been generated by AI. In addition, it needs a minimum of 1,000 words to become operational, so for short texts it is not yet possible to test it.
How does the OpenAI text detector work?
To put this text detector to work, OpenAI has used the same processes they use for any Artificial Intelligence update. Basically, they have carried out a first manual training and then they have continued testing more and more texts for the tool to “train”.
The sources of the generated texts have been three different: Wikipedia, WebText and InstructGPT. Thus, it has several ways to analyze the texts. As we say, it is still early to have great results, but everything in AI is advancing very fast, so we are sure that it will soon be able to be used with certainty.
We have tested the detector and…
To see if it works correctly or, as it seems, still has a long way to go, we have tested this technology. We have used three texts to see if it detects them as original or not. These texts are as follows:
- A text generated in ChatGPT about the history of Apple.
The results have been quite enlightening. About the Softonic article, he said the following:
That is, it is very unlikely to have been generated by Artificial Intelligence. Ok, you’ve done well on this first test (even if you didn’t take much care in your answer).
About the Wikipedia article on Isaac Newton, the same answer has been repeated:
Perfect. At least it detects when an article has not been generated by AI, but what happens when it has?
On this occasion, it does detect that it has been generated with AI, although it only clarifies it as “possible”.
Thus, although it is still a very primitive tool, it is clear that it works. So it is only a matter of time before it is optimized and achieves better results.