Science

Is it possible to detect text written by ChatGPT?

An American student recently tried to create a program capable of recognizing texts written by ChatGPT. Unfortunately, as we explained to you in a previous article, many texts written by AI have been identified as written by humans. However, this does not mean that there is no way to be suspicious. Thus, a university professor was able to determine that half of his students wrote their papers using this artificial intelligence. How ? Realizing that the same ideas are coming back and always in the same order.

To understand how to dump texts written by ChatGPT, you must first understand how it works. Artificial intelligence scans the entire network (until 2021) to answer the question asked by the user. Like a student who will draw on multiple sources to write a paper, or a worker who will draw inspiration from existing cover letters to write their own. Except that ChatGPT doesn’t copy sentences, it paraphrases them… leaving some clues along the way.

Some Tips for Detecting Texts Written by ChatGPT

– General formulas. ChatGPT does indeed calculate the most likely continuation of text. As BFM TV explains, “his answer will consist of the words most likely to appear in his database. Be the password everywhere.” Artificial intelligence can be complex, proof that it has already been able to pass the medical test offered by American institutions, but the text will tend to be uniform, use simple words, and dare not use stylistic effects.

the same structures. If several people ask him the same question, he will answer it in the same structure, using the same arguments. The shape will differ from one to the other, but the structure will most often be the same.

– No reference to the backstory of the person writing. When he writes a text, a person inevitably puts a piece of himself into it. He appeals to his own knowledge of the world, his personal experience … AI will never be able to reproduce this. Similarly, AI cannot copy a person’s writing style. Thus, teachers who know their students well can quickly catch on to something if their writing style suddenly changes.

– Lack of recent events. ChatGPT scanned the entire network until 2021 to create its database. Therefore, he is stunned if we talk to him about recent events. Therefore, in the texts written with ChatGPT, there is (yet) no mention of the events of 2022, as he himself explains to us.

Example of a question asked in ChatGPT ©chatGPT

– Spelling mistakes. Will spelling errors become the prerogative of human-written texts? No, of course not, but if someone who used to make a lot of mistakes suddenly doesn’t make any anymore, it might be interesting to ask questions.

AI, the only real way to detect other AI?

The above advice is largely based on the reader’s feelings… which are not infallible. But everything that artificial intelligence creates, it can technically detect. Therefore, expert groups are working on technologies capable of detecting computer-generated texts. In 2019, people at the MIT-IBM Watson AI Lab and the Harvard Natural Language Processing Group developed an algorithm that was supposed to determine if text was written by a robot. This is called the Giant Language Model Test Room (GLTR).

Why Microsoft is ready to put a billion dollars on OpenAI startup and its ChatGPT chatbot

OpenAI, which developed ChatGPT, also announced the development of software capable of detecting texts written by artificial intelligence. In a lecture at the University of Austin, Professor Scott Aaronson, an OpenAI researcher, revealed that this would be in the form of a watermark, in other words a watermark that would be present in every text written by ChatGPT. However, it is unclear exactly how this will be implemented. “We have always called for transparency regarding the use of AI-generated text,” said an OpenAI spokesperson, however. “Our policy requires users to be honest with their audience when using our tools.” But according to some independent experts, this watermark can be easily hacked and therefore become useless…

Not to mention that if a person goes back to text and improves the AI ​​generated database, it will make any detection impossible… Therefore, the emergence of such a tool will inevitably lead to changes in society. Rather than oppose it, AI proponents advise thinking about how to get the most out of it.

Back to top button

Adblock Detected

Please consider supporting us by disabling your ad blocker.