Since you posted about GPT-2's AI-identification software, I thought I'd ask you about something that's been haunting me for a while. You may have seen a viral twitter post with pictures of a script, labeled something like "I forced an AI to watch 1000 hours of lawyer commercials and this is what it wrote." The person has since done others and even compiled a book. The thing is - the text, to me, absolutely does not in any way read like it was written by an ML program. With jokes like "it's full of lemons, the justice fruit that only lawyers may touch" and "the phone digits appear. It's your social security number," it reads a lot more like a human pretending to be an AI than it does like actual AI writing. Do you have any input? I'd be curious to hear.

You're correct - as far as I know, the author has never used actual machine learning to generate any of those jokes. The first ones came out when AI wasn't even remotely capable of writing that coherently. I wrote a thread about it that ended up in a Know Your Meme article: https://knowyourmeme.com/memes/i-forced-a-bot