If desired, the detector can still be bypassed.
To the developers at OpenAI managed create a detector that identifies texts generated by the ChatGPT chatbot. To do this, the company modifies the text generation in a special way – as a result, certain patterns appear in them that can be detected.
According to OpenAI, the detector defines generated text with 99.9% accuracy. There is a nuance: the detector can be bypassed if you rephrase the text (for example, using another neural network) or run it through a translator.
And here’s another story
The company claims that embedding watermarks into the text does not degrade the quality of ChatGPT generation. However, OpenAI has not yet decided whether to release the detector to the public domain.
It is expected that ChatGPT users will be unhappy with the appearance of such a detector. According to The Wall Street Journal, according to a survey conducted by OpenAI, about 30% of ChatGPT users intend to use the neural network less often if such a detector appears.
