Giant Language Model Test Room (GLTR) is a tool that looks for that level of randomness and was developed by researchers at Harvard and the MIT-IBM Watson AI Lab in 2019. Its functionality can still be tested online, which colors words according to the probability of your prediction according to the text just before it, on the left. If you see a lot of green in a text, which means a word is in the top 10 most predictable, it was most likely typed by a machine.
But the AI moves quickly, and if you pass some ChatGPT-generated text through the above tool, you may have trouble detecting it. GPT-3 (Generative Pre-trained Transformer 3), the language model on which ChatGPT is based, is the improved version of GPT-2, the one used by this tool, GLTR.
The larger and more powerful the language model, the more difficult it is to distinguish between text generated by AI and humans, as warned by experts cited by MIT Technology Review. In a cat-and-mouse game, most AI generated text detector is based on GPT-2 or earlier models, and are not efficient against the latest generation of language models.
How are the texts generated by AI?
What characteristics of the texts generated by AI can give us a sign that they have not been produced by humans, or rather, that there is an AI behind their creation?
Anyone with a bit of experience will quickly realize something: chatbots are generators, as David Karpf, a researcher at George Washington University, called ChatGPT after using it. This feature is typical of machine-generated language: they imitate and therefore repeat.
Artificial intelligence is very good at recognizing and also reproducing patterns. ChatGPT tries to predict the next word in a sentence, and in the process compares multiple options. A method proposed by a University of Maryland computer science professor, Tom Goldstein, for detecting AI-generated text relies on those patterns.
Goldstein proposes using watermarks, which are not visual as their name might lead us to think, but rather a “secret” function in plain sight, dumped into computer code. The function of these watermarks would be to contain certain word patterns. When doing a text crawl, if the rules contained in those watermarks have been broken many times, we can suspect that there is a human being behind it.
AI-generated texts are usually perfect in grammar because it is based on rules, and also in spelling. A typo is a good indicator of human trace. Our texts are also highly variable, with a mixture of styles and their own jargon.
However, if you are looking for the best AI detection tool free, then zerogpt.com is the best tool.
