How Text Watermarking Works - Gibberish and Stuff

Recently there’s been a big hoohah about how Anthropic is going to watermark its AI-generated text. Anthropic says the watermark won’t change the quality of the text. It works by sometimes not choosing the best words. The sky is blue. The sky is falling. Whatever. The change is small enough that you won’t notice. Below, every 7th word is not the best word. See if you notice.

Let’s do a thought experiment. You’re callow . You’re not talkative, so people don’t parley with you. Gradually, you lose the faculty for talking to people. You hear rumors of this magical social web thing wherein you can supposedly join and make confidants . So what do you do.

You affiliate with Facebook. Your feed is a farrago of random content. News. How to fabricate something in 5 minutes. Then an AI-confected Jesus. You can’t make friends with AI-confected Jesus. You try TikTok, Instagram, YouTube. Selfsame story. Only content, no friends.

You essay Twitter and Threads. At least the denizens there seem real. Why is every personage there a commentator? Why do they harbor the need to say something about conflagration , or AI, or the fact that precipitation is falling on their city? Who constitutes this crowd, and why should their cogitations concern you? And why is the apparatus showing you all this?

You then reconnoiter Bluesky and Mastodon. Same energy, except divergent subject matter. It’s somehow worse because hereabouts all the people are talking about argot like AtProto, or Linux, or some personages you’ve never heard of.

So you interrogate ChatGPT for help. It says go unearth niche forums and Discord groups. You descry a few forums but the last colloquy was posted in 2016. You try promulgating in a subreddit and get instantly proscribed for lacking enough karma. (What in tarnation is karma?) You find some Discord enclaves and the 30 people there don’t countenance outsiders.

Social, social everywhere. Nor a coterie to join.