#API to detect AI work?
9 messages · Page 1 of 1 (latest)
An openAI/ gpt detector already exists. The source code is on github. https://huggingface.co/openai-detector/
That only detects GPT-2 output. Detecting GPT-3 output is far more difficult, and may be impossible to do reliably in an adversarial environment where the output may be generated from prompts specifically designed to produce a style that doesn’t match what the detector is expecting.
It detects the output of chatGPT at 99%
This thread has a good overview of the state of play: https://twitter.com/emollick/status/1604872727973830657?s=46&t=_GvBcQLAAiOpR8-8NCwd5g
There are a lot of techniques being shared online by teachers on how to automatically detect ChatGPT output.
It is very doubtful that any of them work well, or if they do, that they do it consistently. Certainly not enough to accuse someone of cheating without other evidence.
Where do you see that? I just pasted my last several ChatGPT completions in, and the ones that it didn’t error on due to <> characters were about 50/50 Real vs. Fake, and almost always highly overconfident (99% one way or the other).
Sounds like a bad loss function or training data (probably the latter) to me
How could I integrate this into a chrome extension?
I think you need a better “this” first, which is an unsolved research problem.