I think some people would be surprised how much hidden information already exists in what seem to be pretty standard AI responses.

A game I like to play with Claude is:

1. Send it a photo of a celebrity
2. Get the standard rejection "Sorry, I can't name people in images, but I can describe what I see..." which tends to be followed by a pretty general description
3. Then asking a different instance of Claude to guess the celebrity, based on the standard rejection response

Since Sonnet 3.5-7, Claude has been pretty ok at this game.

I used to play a different version with famous YouTube links:

1. Send YT link to Sonnet 3.5
2. "Sorry I can't open YouTube links..."
3. Ask a new instance to guess what the link I sent to Claude was

This only worked with extremely famous video links (e.g. https://www.youtube.com/watch?v=dQw4w9WgXcQ), and was much less reliable than the celebrity version.

THIS LINK VERSION ONLY EVER WORKED WITH NEVER GONNA GIVE YOU UP. I have not tested with never models.

Claude was never perfect at these types of games. For example, I found there were clusters of celebrities who shared similar 'vibes' that confuse Claude {Emma Watson, Emma Stone, Anna Kendrick}. But given the space of all possible celebrities/YT links, I was impressed that Claude succeeded at this task as often as it did.

With prompting, Claude got way better at these games (e.g. "Fill out this template of characteristics the person in this photo has: Eye Colour, Hair Colour, Hair Length, Sex, etc")