This seems like a waste of energy. There is not enough text generated at once to be able to tell confidently if something is AI generated or not. Additionally, I'd imagine in adversarial applications, adding an extra step to reword certain things, using dictionary synonyms and phrases, unifying punctuation, and using a secondary grammar check/modifier like Grammarly is trivial.
Open alternatives, like GPT2, have already been released under an MIT license. Community language model trained on NSFW text already exist as well. In my opinion, a move like this would only push away most people in the long term. Given the choice between a non-censored good enough model, and a censored great model, I would choose the good enough model for 99% of my long term projects.
9 comments
[ 1.6 ms ] story [ 34.6 ms ] threadOthers will emerge.
Open alternatives, like GPT2, have already been released under an MIT license. Community language model trained on NSFW text already exist as well. In my opinion, a move like this would only push away most people in the long term. Given the choice between a non-censored good enough model, and a censored great model, I would choose the good enough model for 99% of my long term projects.
If I know you put steganography in there I can make damn sure it doesn't stay in, for small enough chunks of text.