OpenAI Released an AI Text Classifier yesterday

It was just put up yesterday and it appears to be usable with just a free account if you have toyed with ChatCPT: https://platform.openai.com/ai-text-classifier

Anyone else check it out and what kind of results are you seeing and how does it compare to the Huggingface one (which appears somewhat recently to detect a little better with some changes).

I'll kick things off. With some of my earlier AI I'm getting that it's unlikely AI, and with my more recent stuff that I alter via custom code I'm getting very likely AI.

I'm going to try to generate a couple of ChatGPT things to run through the detector which logic would dictate it would pick up being it's the same company.

Thank you, I just tried it.

Trick: (If your text have less 1000 characters you can repeat the text until you have more than 1000 characters and it will work fine)

i tested some AI articles that i've generated yesterday and today and it was never detected as AI content, without any changes to the content.
"The classifier considers the text to be very unlikely AI-generated." lol

Most of the ones I just generated using ChatGPT it's picking up.

Using some I just generated with a stock GPT3 tool it was somewhat picking it up most of the time.

I made a typo in my first post regarding my latest AI. With my own custom coding for AI articles I'm getting very unlikely AI-generated. I think it's a given at this point that you have to alter any AI beyond what it outputs to do well.

These all AI detectors are shit I think. As Google use own algo and detector.
Just checked AI article on:

contentatscale.ai/ai-content-detector - 1% human.
openai-openai-detector.hf.space - 99.77% fake.
writer.com/ai-content-detector - 0% human.
platform.openai.com/ai-text-classifier - very unlikely AI-generated.

Passing on any detector does not mean it'll pass Google as they have their own detector. However, it does give an indicator as they all operate on the same principle.

AI writes content by using most likely next words, etc. The more it conforms to an algorithm(such as lower temperature), the more predictable it becomes.

AI detectors work in a similar way, in that they check what the most likely next word is, and the more matches the more likely it's AI. Remember, it's trying to detect how AI works. There is no magic way of knowing other than trying to reverse engineer current AI writers.

That's how they work. Adding a human touch means adding imperfection (or personality) and that non-predictability is what makes it more human. One thing to note is that the AI detector needs to be just as well trained if not more than the AI generator to be effective. It's all about data here. So Google does have an advantage if they choose to pursue it. Obviously with recent updates Google is still having a hard time when it's not obvious.

For an AI detector to be effective it needs to work like that or else it's useless. The writer.com one is a bit optimistic and the huggingface(hf.space) one is a bit pessimistic. However, I'd say it's worth checking them all to get overall feedback on tests as it's always a best guess.

Shitty repetitive article from articleforge, nichesss it struggle to detect.
While open ai GPT-2 detector can detect them easily.