Re:Drowning in AI Generated Garbage: the silent war we are f...
by robocat on Monday December 05, 2022 @09:07PM
(#33874952)
Except at present we can also train classifiers that recognise GPT output, so ingressed text should be able to be filtered.
https://huggingface.co/openai-detector/ detects GPT2, but it still detects ChatGPT snippets most of the time, even though the algorithm has been improved and the training set is presumably quite different.
That said, I did just generate a complete lie https://news.ycombinator.com/item?id=33874757 using ChatGPT that scored 99.98% real on the detector, so adversaries can pre-filter too.
Also interesting is that historical corpora (every existing corpus) of data becomes more valuable, because they are untainted.
https://huggingface.co/openai-detector/ detects GPT2, but it still detects ChatGPT snippets most of the time, even though the algorithm has been improved and the training set is presumably quite different.
That said, I did just generate a complete lie https://news.ycombinator.com/item?id=33874757 using ChatGPT that scored 99.98% real on the detector, so adversaries can pre-filter too.
Also interesting is that historical corpora (every existing corpus) of data becomes more valuable, because they are untainted.