OpenAI just admitted it can’t identify AI-generated text. That’s bad for the internet and it could be really bad for AI models.::In January, OpenAI launched a system for identifying AI-generated text. This month, the company scrapped it.
OpenAI just admitted it can’t identify AI-generated text. That’s bad for the internet and it could be really bad for AI models.::In January, OpenAI launched a system for identifying AI-generated text. This month, the company scrapped it.
If the point of a language model is to emulate human language, being able to differentiate somewhat defeats the point. The same goes for image and video generation.
In fact, anyone interested in this goal should be using such adversarial detection techniques to better train their model, regardless.
If we can distinguish natural from artificial, the artificial still needs improvement. And at infinity, the two are indistinguishable, period.