Good call, that was only suggested as a could-be option. There have been better models such as GPT4-Turbo and Claude 3 since that blog post which could hint at better performance in theory
This wasn’t a swipe at the overall approach here.
Using an LLM to look for features in the text that might suggest it is spam is a really great goal and I welcome the addition wholeheartedly.
My only issue was with the claim that you can automatically identify AI generated text, which is a very hard problem.
Note the distinction:
- Is identifying general spam with LLM = great and I can see how this might be tractable and very worthwhile
- Is identifying that the spam was generated by AI - with or without the use of an LLM this is hard. I don’t think current models are able to do this well specifically and positives are probably just hallucinations.
I hear you! yeah it’s a much more nuanced issue. In fact in my original testing I tried to have it setup for both but focusing on just spam probably yields better results here
So we are in agreement.
I do think there are some tell-tale signs of AI though - like post length and overuse of exceptionally flowery language … so maybe those aspects can be analysed … post length is probably hard for an LLM though “flowery-ness” might be less difficult. You can use deterministic functions for post length flagging though, of course.
Then again, if someone has used ChatGPT to refine e.g. their English (as second language) then flagging their post is probably not helpful to overall Community inclusion.
Going one step further here, general machine learning techniques might be useful but more complex to implement than just sending text to an LLM and getting the result.